Khongor Dissertation

i
Ph.D. Program of Agriculture Science

National Chiayi University
Ph. D. Dissertation

Lands at TM

Modeling forests stand structures, and aboveground biomass
estimation using Landsat TM imagery in Mongolia

Advisor: Chinsu Lin, Ph. D.

Graduate Student: Khongor Tsogt
:

January, 2013
102 1

ii

iii

Acknowledgements
This thesis would not have been possible unless with support of Taiwanese government and
National Chiayi University scholarships and financial support of my supervisor, Chinsu Lin. I am
heartily thankful to my supervisor whose encouragement, guidance and support from the initial to
the final level enabled me to develop an understanding of the subject. Besides I would like to thank
my father, Tsogt Zandraabal, and my colleagues for providing me great information resources. It is
an honor for me to study in Taiwan with support of my mother Dagiisuren Tserennadmid and
family members. Lastly, I am indebted to my many of my friends, Liang-Ting and Marco Lee to
morally support me and lab mates, J ia Wei and A-Zhi boosted me in any respect during the
completion of the study.

iv

Table of Contents
Table of Contents ........................................................................................................................... iv
List of Figures ................................................................................................................................ ix
List of Tables ................................................................................................................................. xi
Part I. Modeling forest stand structures for volume estimation ........................................................ 1
Part I. Background .......................................................................................................................... 1
Chapter 1 ........................................................................................................................................ 4
1 Diameter structure analysis of a larch forest stand and selection of suitable model................... 4
1.1 Abstract ........................................................................................................................... 4
1.2 Introduction ..................................................................................................................... 4
1.3 Materials and Methods..................................................................................................... 5
1.3.1 Field measurement ....................................................................................................... 5
1.3.2 Data analysis................................................................................................................ 5
1.4 Results ............................................................................................................................. 6
Table 1.1 Parameter estimates of the three distribution models for the study plot ............................ 6
Fig. 1.1 Model comparison for the study plot .................................................................................. 7
1.5 Discussion and Conclusion .............................................................................................. 7
Chapter 2 ........................................................................................................................................ 9
2 Diameter and height distributions of natural even-aged Pine forests......................................... 9
2.1 Abstract ........................................................................................................................... 9
2.2 Introduction ..................................................................................................................... 9
2.3 Materials and Methods................................................................................................... 10
2.3.1 Field measurement ..................................................................................................... 10
Table 2.1 Inventory information of the study site .......................................................................... 11
Table 2.2 Descriptive diameter and height statistics of the study site ............................................. 11
2.3.2 Data analysis.............................................................................................................. 11
2.4 Results ........................................................................................................................... 12
Table 2.3 Parameter estimates of the dbh and h distribution models for pine forests, P1-P3 ........... 12
2.5 Discussion ..................................................................................................................... 13
2.6 Conclusion .................................................................................................................... 16
Chapter 3 ...................................................................................................................................... 17
3 Modeling diameter and height distributions of a spruce-larch mixed forest ............................ 17

v

3.1 Abstract ......................................................................................................................... 17
3.2 Introduction ................................................................................................................... 17
3.3 Material and Methods .................................................................................................... 19
Table 3.1 D and H characteristics of the spruce-larch forest in the study site ................................. 19
3.3.2 Data analysis.............................................................................................................. 19
3.3.3 A flexible mixture model for dbh structure modeling ................................................. 20
3.3.4 Model comparison ..................................................................................................... 21
3.4 Results ........................................................................................................................... 21
3.5 Discussion ..................................................................................................................... 26
3.6 Conclusion .................................................................................................................... 28
Chapter 4 ...................................................................................................................................... 29
4 A flexible modeling of irregular diameter structure for the volume estimation of larch forest
stands ............................................................................................................................................ 29
4.1 Abstract ......................................................................................................................... 29
4.2 Introduction ................................................................................................................... 29
4.3.2 Data analysis.............................................................................................................. 32
4.3.3 dbh-h relationship modeling and assumption tests ...................................................... 32
4.3.4 Volume estimation and accuracy assessment .............................................................. 33
4.4 Results ........................................................................................................................... 34
4.4.1 Simple dbh distribution models of larch forest ........................................................... 34
4.4.2 Mixture dbh distribution models of larch forest .......................................................... 35
4.4.3 dbh-h relationship of larch forest................................................................................ 37
4.4.4 Accuracy assessment of the volume estimation for the models derived by unimodal
approach and multimodal approach ....................................................................................... 40
4.5 Discussion and Conclusion ............................................................................................ 40
Chapter 5 ...................................................................................................................................... 42
5 Forests aboveground biomass and carbon storage estimation in Eastern Khentii, Mongolia ... 42
5.1 Abstract ......................................................................................................................... 42
5.2 Introduction ................................................................................................................... 42

vi

5.3.1 Study area .................................................................................................................. 43
Fig. 5.1 Tree species of study area, Mongon morit soum, Tov aimag, Mongolia. ........................... 44
Table 5.1 General characteristics of studied plots .......................................................................... 45
5.3.3 Biomass calculation ................................................................................................... 46
5.3.4 Carbon and other elemental composition estimation ................................................... 47
Table 5.2 Four main elemental compositions of larch and birch tree species in Mongolia .............. 47
5.4 Results and Discussion .................................................................................................. 47
5.5 Conclusion .................................................................................................................... 49
Part I. Summary ............................................................................................................................ 50
Part II. Aboveground biomass estimation using Landsat TM imagery in Mongolia ....................... 52
Part II. Background ....................................................................................................................... 52
6 Regression models for forest aboveground biomass and leaf area index in North-Eastern
Mongolia ...................................................................................................................................... 53
6.1 Abstract ......................................................................................................................... 53
6.2 Introduction ................................................................................................................... 53
6.2.1 Background ............................................................................................................... 53
6.2.2 Spectral characteristics of vegetation ......................................................................... 55
6.2.3 Objective of the research ............................................................................................ 57
6.3.1 Study area and field measurement .............................................................................. 58
6.3.2 Abovegound biomass estimation ................................................................................ 58
6.3.3 LAI-2000 Plant canopy analyzer (Li-COR) ................................................................ 59
6.3.4 Landsat TM data ........................................................................................................ 59
6.3.5 Vegetation indices ..................................................................................................... 60
Table 6.1 Vegetation indices (VIs) used in this study .................................................................... 60
6.3.6 Data analysis.............................................................................................................. 61
6.4 Results ........................................................................................................................... 63
6.4.1 Correlations between biomass of tree components and/or LAI and Landsat TM bands63
6.4.2 Correlation between AGB and/or LAI, and band ratios .............................................. 64
6.4.3 Correlation between AGB and/or LAI, and VIs .......................................................... 65
6.4.4 Models for AGB and LAI .......................................................................................... 66
6.4.5 Index validation ......................................................................................................... 69

vii

6.5 Discussion ..................................................................................................................... 69
6.5.1 Relationship between AGB, LAI, Landsat TM bands and band ratios ........................ 69
6.5.2 Relationship between AGB, LAI and VIs................................................................... 71
6.5.3 AGB and LAI estimation using regression equation ................................................... 72
6.5.4 Effect of the undergrowth vegetation and background reflectance .............................. 72
6.5.5 Model verification ..................................................................................................... 73
6.5.6 Suggestions................................................................................................................ 73
6.6 Conclusion .................................................................................................................... 74
Chapter 7 ...................................................................................................................................... 76
7 Mapping land cover and forest types using multi-temporal Landsat imageries with biophysical
information of dominant forest types in mountainous area ............................................................. 76
7.1 Abstract ......................................................................................................................... 76
7.2 Introduction ................................................................................................................... 76
7.3 Materials and Methods................................................................................................... 78
7.3.1 Study area .................................................................................................................. 78
7.3.2 Landsat TM data ........................................................................................................ 78
Fig. 7.1 Framework of image processing and forest type map........................................................ 79
7.3.3 Image preprocessing .................................................................................................. 80
7.3.4 Classifiers .................................................................................................................. 80
7.3.4.1 Support vector machines (SVM) ........................................................................ 80
7.3.4.2 Maximum Likelihood ......................................................................................... 81
7.3.5 Region of interest (ROI) for classification .................................................................. 81
Table 7.1 Training pixels and reference pixels used for image classification and validation ........... 82
7.3.6 Classification accuracy assessment ............................................................................ 83
7.3.7 Ancillary biophysical information for post-classification ........................................... 83
Table 7.2 Forests altitudinal belt zones in Eastern Khentey province, Mongolia ............................ 84
Table 7.3 Conditional distributions of forest types by elevation ..................................................... 86
7.4 Results ........................................................................................................................... 87
7.4.1 Land cover and forest type classification result and accuracy assessment for multi-
temporal imagery .................................................................................................................. 87
Table 7.4 Confusion matrix of multi-temporal image classification result by SVM ....................... 88
Table 7.5 Confusion matrix of multi-temporal image classification result by Maximum likelihood 89
Table 7.6 Classification accuracy assessment of Maximum likelihood and SVM........................... 89

viii

7.4.2 Land cover and biophysically adjusted forest type image classification result and
accuracy assessment .............................................................................................................. 90
Table 7.9 Post-classification accuracy assessment of Maximum likelihood and SVM.................... 91
7.4.3 Dominant forest type, larch and cedar mapping accuracy assessment ......................... 92
7.5 Discussion ..................................................................................................................... 94
7.5.1 Classification accuracy improvement in multi-temporal imagery ............................... 94
7.5.2 Land cover mapping and accuracy ............................................................................. 95
7.5.3 Dominant forest type mapping and accuracy .............................................................. 95
7.6 Conclusion .................................................................................................................... 96
Part II. Summary ........................................................................................................................... 97
References .................................................................................................................................... 99
Appendixes ................................................................................................................................. 110

ix

List of Figures
Fig. 0.1 Framework of Part I study .................................................................................................. 3
Fig. 1.1 Model comparison for the study plot .................................................................................. 7
Fig. 2.1 dbh (left) and h (right) model comparisons for pine forests, study plots P1-P3. The
histogram represents the observed distribution and the short dashed line (Burr), long dashed
line (Dagum), and solid line (J ohnson SB) show the estimated distributions. ........................ 14
Fig. 3.1 Model comparison for the Spruce-Larch mixed forest. The histogram represents the
observed diameter at breast height (dbh) and height (h) distributions with simple distribution
model (solid line), mixture distribution model (dashed line). a) dbh of spruce-larch mixed
trees fit by simple distribution, Burr (4P) and fit by mixture distribution, Burr (4P) and
J ohnson S
B.
b) dbh of spruce trees fit by simple distribution, Burr (4P) and fit by mixture
distribution Burr (4P) and J ohnson S
B
. c) dbh of larch trees fit by simple distribution, Dagum
(3P) and fit by mixture distribution, double J ohnson S
B.
d) h of spruce-larch mixed trees fit by
simple distribution J ohnson S
B
and fit by mixture distribution double J ohnson S
B
. e) h of
spruce trees fit by simple distribution, Dagum (3P) and fit by mixture distribution Dagum (4P)
and J ohnson S
B
. f) h of larch trees fit by simple distribution, Dagum (3P) and fit by mixture
distribution Dagum (3P) and Johnson S
B
. ............................................................................. 23
Fig. 3.2 Spruce and larch trees a. D and b. H structures. a. Spruce trees are dominant in D 2-10 cm
while larch trees are dominant in mid D classes around 12-16 cm and then equally distributed
around 18-24 cm and around 26-30 cm only spruce trees inventoried. b. Spruce trees are
dominant 3-11 m H while larch trees are dominant in 12-19 m. ............................................ 27
Fig. 4.1 The diameter distribution of the study plots in larch forests. The histogram represents the
observed dbh distribution, the dashed line represents the derived simple distribution model
and the solid line represents the derived mixture distribution model. .................................... 36
Fig. 4.2 Curvilinear relationships of tree height to dbh of larch forests. The relationship is
exponential for plots 1-2 and sigmoid for plots 3-6. .............................................................. 39
Fig. 5.1 Tree species of study area, Mongon morit soum, Tov aimag, Mongolia. ........................... 44
Fig. 6.1 Spectral reflectance of healthy Red cypress (Chamaecyparis formosensis), green vegetation
for the wavelength interval 350-2500 nm. The dominant factors controlling leaf reflectance
are the various leaf pigments in the palisade mesophyll (e.g., chlorophyll a and b, and -
carotene), the scattering of near-infrared energy in the spongy mesophyll, and the amount of
water in the plant. The chlorophyll absorption regions are at 430-450 nm and 650-660 nm and
the water absorption regions occur at 970, 1190, 1450, and 1940 nm. .................................. 56

x

Fig. 6.2 Additive reflectance of Cinnamomum camphora leaves. 1, 3 and 9 layers. Near-infrared
reflectance increases as the leaf layers in the canopy increases. Direct relationship exists
between response in the near-infrared region and biomass measurements. ............................ 56
Fig. 6.3 Yearly pattern of Red cypress (Chamaecyparis formosensis) trees absorption change in the
blue spectral region, 400-550 nm (Tsogt et al. 2008). ........................................................... 57
Fig. 6.4 Framework of aboveground biomass (AGB) and leaf area index (LAI) study relationship
with vegetation indices (VIs). ............................................................................................... 58
Fig. 6.5 Relationship between Landsat TM bands 3, 5 and 7 and aboveground biomass (AGB) (left),
and relationship between Landsat TM bands 2, 3, 4 and 5 and leaf area index (LAI) (right).
Correlation coefficients (r) are given for bands and AGB and/or LAI. .................................. 64
Fig. 6.6 Relationship between VIs and AGB for forest stands. Log transformed Brightness index
and aboveground biomass (AGB) (top left), log transformed Surface albedo and AGB (top
right), visible atmospherically resistant index and AGB (middle left), NDVIc and AGB
(middle right), and SWIR corrected NDGVI and AGB (lower left). The SWIR correction
substantially reduced NDVI values over low AGB. .............................................................. 67
Fig. 6.7 Relationship between VIs and LAI. Logarithm transformed (log) Brightness and
exponential transformed (exp) leaf area index (LAI) (top left), log-Surface albedo and exp-
LAI (top right), VARI and log-LAI (middle left), exp-NDVIc and log-LAI (middle right), and
exp-NDGVIc and log-LAI (bottom left). .............................................................................. 68
Fig. 7.1 Framework of image processing and forest type map........................................................ 79
Fig. 7.2 Biophysical information rules of forests distributions along elevation for dominant forest
types post classification. ....................................................................................................... 85
Fig. 7.3 Alternative, accuracy assessment for image post-classification. ........................................ 86
Fig. 7.4 Multi-temporal image composite differences between SVM and Maximum likelihood
classification results before and after applying ancillary biophysical information in dominant
fores type, larch and cedar classes. a and c - SVM and Maximum likelihood classification
results; b and d - SVM and Maximum likelihood classification results adjusted by forests
biophysical distribution characteristics along altitudinal-belt zones ...................................... 94

xi

List of Tables

Table 1.1 Parameter estimates of the three distribution models for the study plot ............................ 6
Table 1.2 Summary of empirical diameter distribution for larch forest (=0.05) .............................. 6
Table 2.1 Inventory information of the study site .......................................................................... 11
Table 2.2 Descriptive diameter and height statistics of the study site ............................................. 11
Table 2.3 Parameter estimates of the dbh and h distribution models for pine forests, P1-P3 ........... 12
Table 2.4 Goodness-of-fit and ranking (rank in parentheses) of Burr, Dagum and J ohnson SB
distributions for the empirical dbh and h distributions as measured by MLE criterion ( = 0.05)
............................................................................................................................................ 15
Table 3.1 D and H characteristics of the spruce-larch forest in the study site ................................. 19
Table 3.2 Parameter estimates of diameter at breast height distribution models for the spruce-larch
mixed, spruce and larch trees (=0.05) ................................................................................. 24
Table 3.3 Parameter estimates of tree height distribution models for the spruce-larch mixed, spruce
and larch trees (=0.05) ....................................................................................................... 25
Table 3.4 RMSE and
2
test of the simple and the mixture models for the spruce-larch mixed,
spruce and larch trees (N/ha) ................................................................................................ 26
Table 4.1 Descriptive statistics of the study plots. ......................................................................... 32
Table 4.2 Coefficients of tree volume equation (with bark) by diameter at breast height and height
............................................................................................................................................ 33
Table 4.3 Measures of the goodness of fit tests for the unimodal approach derived dbh distribution
models. ................................................................................................................................ 35
Table 4.4 Measures of the goodness of fit tests for the multimodal approach derived sub-group dbh
distribution models............................................................................................................... 37
Table 4.5 Measures of the goodness of fit tests for the empirical dbh-h models. ............................ 38
Table 4.6 A comparison of the accuracy of volume estimation using the simple dbh models and the
mixture dbh models
*
. ........................................................................................................... 40
Table 5.1 General characteristics of studied plots .......................................................................... 45
Table 5.2 Four main elemental compositions of larch and birch tree species in Mongolia .............. 47
Table 5.3 Aboveground biomass and carbon of larch, birch and mixed forest stands in Eastern
Khentii, Mongolia (AGC Aboveground carbon) ................................................................ 47
Table 6.1 Vegetation indices (VIs) used in this study .................................................................... 60

xii

Table 6.2 Biomass of tree components and LAI of study plots (0.09 ha, regards to Landsat TM pixel
size 30x30 m)....................................................................................................................... 62
Table 6.3 Correlations of the biomass components of forest stand and leaf area index (LAI) against
the Landsat TM band reflectance ......................................................................................... 64
Table 6.4 Correlations of the aboveground biomass of forest stand and leaf area index against the
band ratios. .......................................................................................................................... 65
VIs. ...................................................................................................................................... 66
Table 6.6 Regression models for aboveground biomass (AGB) and leaf area index (LAI) ............. 69
Table 6.7 Correlations of the AGB of forest stand and LAI against the VIs. .................................. 69
Table 7.1 Training pixels and reference pixels used for image classification and validation ........... 82
Table 7.2 Forests altitudinal belt zones in Eastern Khentey province, Mongolia ............................ 84
Table 7.3 Conditional distributions of forest types by elevation ..................................................... 86
Table 7.4 Confusion matrix of multi-temporal image classification result by SVM ....................... 88
Table 7.5 Confusion matrix of multi-temporal image classification result by Maximum likelihood 89
Table 7.6 Classification accuracy assessment of Maximum likelihood and SVM........................... 89
Table 7.7 Confusion matrix of post-classification result by SVM. ................................................. 90
Table 7.8 Confusion matrix post-classification result by Maximum likelihood. ............................. 91
Table 7.9 Post-classification accuracy assessment of Maximum likelihood and SVM.................... 91
Table 7.10 Accuracy assessment for the larch and cedar cover type map of the study area with
comparison of inventory map-data ....................................................................................... 92
Table 7.11 Accuracy assessment for the larch and cedar cover type (that was adjusted based on
forests distribution characteristics along elevation) map of the study area with comparison of
inventory map-data .............................................................................................................. 93

1

Part I. Modeling forest stand structures for volume estimation

Part I. Background
Volume estimation of forest stands is an important task in the accounting of forest stocks as
well as carbon storage and sequestration. Modeling diameter at breast height (dbh) structure and
dbh-height (h) relationship is essential in order to carry out the volume and carbon accounting task.
In past decades, the mainstream approach to forest volume estimation generally operates at the
stand level (Wiant et al. 1989, Hamilton and Brack 1999) and therefore is also named stand-level
approach. It uses the stand diameter and height to obtain the information needed for forest
management such as the calculation of forest productivity, annual or periodical yield, and economic
benefits. The stand-level approach lacks accuracy because generalized stand parameters may not be
able to represent the natural variation of a larger area of forest stands and hence might cause a high
uncertainty of prediction. Recently, a large variety of research has been conducted based on using
forest structures, such as dbh, height and age to overcome the inaccuracy of the stand-level
approach. These kinds of methods generate more suitable models based on the probability density
functions (pdf) of the inventoried forest samples to present more accurately the dbh and height
structure of a forest stand.
In a forest, there are always variations in tree size. These variations are caused by biological
reasons and by spatial locations across stands, and by microclimatic conditions. These variations
easily can be seen by tree diameter and height.
In forestry, we often use probability density functions to reveal forest stands structures. Based
on forest stand structure, we can understand forests development stages, phases and dynamics. But,
depending on development stages, forests size structural shapes are different. Then it is necessary to
use various theoretical distributions to fit stands trees dbh distributions to explain their development
cycle or phase. It is also possible to build growth models. to predict stand changes and dynamics.
For instance, dbh distribution models can be used to complete missing dbh class of life table
information.
Modeling dbh distributions of different development stages of forests is important to define
for theoretical and practical forestry. Depend on development stages, forest structural distributions
possible to be explained by different distributions. Then it is able to define theoretical forests by
distribution shape. Especially, if the stand trees are highly skewed or irregular shaped, diameter
distribution modeling is very important in order to choose proper strategy management plan.

2

Forest volume measures expressed by per hectare area, but researchers often do not develop a
hectare area for study and we divide volume stock by study area to transfer for a hectare. But, each
sample plots has got unique structural characteristics. This could bring huge bias to forest planning
and an end product. One of the advantages of using pdf to predict forest structure is that they could
predict forest structure by their general tendency.
pdf s can manipulate to produce detailed diameter distribution of forest stand with statistically
promised probability. Traditional dbh distribution models falls between one sided an exponential
distribution and a symmetrical normal distribution (or Weibull distribution). There are more
sensitive and flexible distribution functions are available to model dbh distributions. Modeling dbh
distribution becomes easy task with the availability of specified statistical software packages.
Forest is a complex dynamic system which may consist of several individual patches with
different variations. If the variations are similar among the patches, the forest size distribution is
regular shaped (unimodal) and could be modeled by simple distribution function. If the variations
are different among the patches, the forest structure is irregular shaped. Then simple distribution
functions are no longer sufficient for modeling irregular shaped diameter and height distributions.
Therefore, mixture distribution models are known better for structurally heterogeneous forests.
In Fig. 0.1 shows general framework of Part I study. Chapters 1 and 2 represent simple
distribution modeling while Chapters 3 and 4 compares the results of simple and mixture
distribution modeling. Chapter 1 shows suitable distribution model selection for forest diameter
structure modeling in case of a pure larch (Larix sibirica) forest stand. Chapter 2 shows also
distribution model selection, but for both diameter and height structures in case of pure pine (Pinus
sylvestris) forest stands. Rather than distribution model selection, Chapter 3 shows a comparison of
simple and mixture distribution modeling in structural parameters, both diameter and height in case
of spruce-larch (Picea obovata and Larix sibirica) mixed forest. Chapter 4 shows flexible mixture
distribution modeling in case of larch (Larix sibirica) forest stands and further diameter-height
relationships discovered for volume estimation. In chapter 5, we are concentrated on one of the
important tasks of quantifying forests AGB.

3

Fig. 0.1 Framework of Part I study

4

Chapter 1
1 Diameter structure analysis of a larch forest stand and selection of suitable model

1.1 Abstract
Ecologically and economically it is important to understand how many tree stems are in each
diameter at breast height (dbh) class. The purpose of this study was to find larch forest (Larix
sibirica) dbh distribution model among Weibull, Burr and J ohnson S
B
distributions. Inventory was
conducted near Gachuurt village, Ulaanbaatar, Mongolia. The goodness of fit test were
accompanied with Kolmogorov-Smirnov (K-S), Anderson-Darling (A-D) and Chi-Square (
2
) tests
for distribution models. Study result shows J ohnson S
B
distribution gave the best performance in
terms of quality of fit to the diameter distribution of larch forest.

1.2 Introduction
Detailed information of forest stand is crucial for forest research and planning. This
information used for input of ecosystem modeling and/or forest growth and yield models. In the
analysis of stand dynamics, detailed data for all trees on a plot is often lacking. In such case, we
may generate missing data using various theoretical at breast height (dbh) distributions. For many
years there were various activity and interest in describing the frequency distribution of dbh
measurements in forest stands using probability density functions. First study of dbh distribution
mathematical description was negative exponential (DeLiocourt 1898), and since then, researchers
used various distributions.
All distribution models have their advantage and sensitive in specific shape. Weibull
distribution able to describe Exponential, Normal and Lognormal distribution shapes (Bailey & Dell,
1973; Lin et al., 2007), while Burr distribution cover much larger area of skewness and kurtosis
plane than the Weibull distribution (Lindsay et al., 1996). Moreover, it is closely approximate with
above mentioned distributions plus Gamma, Logistic and several Pearson type distributions.
J ohnson S
B
distribution cover different region of skewness and kurtosis plane than the Burr
(J ohnson, 1949; Hafley & Schreuder, 1977), and it is closely approximate Beta and generalized
Weibull distributions.
In case of Mongolian forests, Khongor et al. (2011a) published the birch forest dbh study
using Weibull and Lognormal distributions and compared the accurateness of these models. For
larch forest dbh distribution, Khongor et al., (2011b) used Exponential, Lognormal and Gaussian
(or Normal) distributions, but they did not used Weibull, Burr and J ohnson S
B
before.

5

The purpose of this study is to investigate the suitability of the Weibull, Burr and J ohnson S
B

distributions for modeling dbh distribution of larch forest (Larix sibirica).

1.3 Materials and Methods
1.3.1 Field measurement
Study plot was selected near the Gachuurt village in the vicinity of Ulaanbaatar city,
Mongolia, located at 4800'18.9''N and 10713'23.1''E with altitudinal elevation 1607-1627 m above
sea level. The forest consisted of natural stands and any management activity had taken previously.
Inventory was conducted in summer of 2009. Composition of the stands is pure larch. Plot size was
0.2 ha, i.e. 40 x 50 m in area. D were measured for all trees >1.3 m, and totally 275 stems were
counted. Average dbh of tree stands in plot was 14.6 cm with standard error mean 0.497 cm. Dbh of
tree stems ranges from 2 to 32 cm. Dbh distribution skewness value was 0.38 indicating that the tail
on the right side of the probability density function is longer than the left side and kurtosis value -
0.97 indicating statistically flattered peak.

1.3.2 Data analysis
The goodness of fit of empirical D distribution was tested using three theoretical distributions:
J ohnson S
B
, Weibull and Burr (Appx 1). A suitable model was chosen for further application based
on the results of the goodness of fit test for each plot. EasyFit 5.5 Professional software was used
to estimate the parameters and to examine the goodness of fit of the distribution models mentioned
above. Empirical models of the derived stand dbh structure were examined by the Kolmogorov-
Smirnov (K-S), Anderson-Darling (A-D) and chi square (
2
) tests. The K-S test is a nonparametric
test for the equality of a continuous, one-dimensional probability distribution that can be used to
compare a sample with a reference probability distribution. It is based on the largest vertical
difference between the theoretical and the empirical cumulative distribution function. The K-S
critical values used in this study are based on the table published in statistical literature (DAgostino
1986). The A-D test is also a nonparametric test of whether there is evidence that a given sample of
data did not arise from a given probability distribution. It is more sensitive to the tails of a
distribution than the K-S test. The
2

test

is used for binned data and checks if the sample data came
from a specific distribution. So the value of the test statistic depends on how the data is binned. For
this study, the dbh data grouped into intervals of equal probability. Hypothesis tests of dbh structure
model were carried out by examining the p-value that is associated with a goodness of fit statistic.
When the p-value is less than the pre-defined critical value or the significant probability level, the

6

null hypothesis is rejected and it is concluded that the data did not come from the specified
distribution.

1.4 Results
The parameter estimates of the three models are given in Table 1.1 Parameter estimates of the
three distribution models for the study plot. The predictions from each model were compared with
observed frequencies. The K-S, A-D and
2
tests and P value for K-S and
2
tests were computed for
each model (Table 1.2 Summary of empirical diameter distribution for larch forest (=0.05)). All
tested distribution models were statistically fitted with observed dbh distribution and among them
J ohnson S
B
distribution was more flexible than Weibull and Burr distributions.

Table 1.1 Parameter estimates of the three distribution models for the study plot
J ohnson S
B
Weibull Burr
k
0.38167 0.68757 31.863 1.81 1.8296 16.335 1474.1 1.8545 841.49
1
By the definition, the area under the pdf graph must equal 1, so the theoretical pdf values have to be multiplied by the
total number of stems to match the
1
histogramand the D coverage of bin width to calculate the number of stems in each
dbh class.

Table 1.2 Summary of empirical diameter distribution for larch forest (=0.05)
Distribution
Kolmogorov-Smirnov
(critical value 0.08189)
Anderson-Darling
Chi-Squared
statistic P-value statistic statistic P-value
Johnson S
B
0.03106 0.94589 0.18795 7.6044 0.47303
Weibull 0.06891 0.14004 1.8136 10.535 0.22946
Burr 0.07506 0.08567 1.909 13.562 0.09392

Tree stems are smoothly distributed in dbh classes and it is statistically unimodal. It is easy to
fit such distribution, but here flattered peak is problem that causes under/over prediction. Though all
models passed on goodness of fit test, Weibull and Burr models over-predict dbh classes around 10-
16 cm and lower-predict 4-6 cm and 24-30 cm classes. It is evident that the J ohnson S
B
model was
more flexible in fitting flattered dbh distribution of larch forest stand (Fig. 1.1 Model comparison
for the study plot).

7

Fig. 1.1 Model comparison for the study plot

1.5 Discussion and Conclusion
Weibull and Burr theoretical distributions fit the best for right tailed dbh distributions whilst
J ohnson S
B
distribution has ability to represent equally well right and left tailed distributions. With
this reason we have chosen the Weibull, Burr and J ohnson S
B
distributions to test their suitability
and flexibility for larch forest dbh distribution. Then, our study result suggest J ohnson S
B

distribution for larch forest and that is would not be necessary to believe about the Johnson S
B

distribution is the best for all over larch forest in Mongolia.
Forests stand dbh distributions are different depending on the site quality, climatic condition
and history of natural or human disturbances. Supposedly, empirical distribution models would
accurately work in big scale if the geographic and climate conditions are same. But, random
disturbances, such as forest fire, insect invasion or selective logging are change the forest structure
and shape in different forms. Specially, every forest ever influenced with forest fire in Mongolia
and near urbanized areas all forests under danger of illegal timber logging.
However, it is still important that stand specific forest structure information for model
development and research or management planning in small scale forest area. If we needed bigger
scale as regional forest dbh structure, we have to collect more stand dbh data to fit general dbh
distribution. The required amount of stem numbers or sample plots for regional dbh distribution
study would be defined by stability of a chosen model. If the one fails we need to collect more stand
samples and do it again until it become statistically stable. Westphal (2006) suggested that for the
0
5
10
15
20
25
30
35
40
2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32
N
u
m
b
e
r

o
f

s
t
e
m
s
Diameter (cm)
Observed diameter distribution
Johnson SB distribution
Weibull distribution
Burr distribution

8

regional scale dbh distribution is reverse J shaped because of many small stems and relatively fewer
big stems.
Strong intensity disturbances or high intensity regeneration may change dbh structure as
bimodal. If disturbance happened repeatedly in same forest, then the dbh distribution would forms
multimodal shape. In this study, we used unimodal dbh distribution. However, it may not be
sufficient when a frequency distribution is reverse J with hump, bimodal or multimodal, and
therefore, irregular shaped dbh distributions should have tested by mixture distribution (Zhang &
Liu, 2006).

9

Chapter 2
2 Diameter and height distributions of natural even-aged Pine forests

2.1 Abstract
The purpose of this study is to find a suitable probability density function (pdf) to model diameter
breast height (dbh) and height (h) distributions of even-aged pine (Pinus sylvestris L.) forests. For
the study, three different age-class (AG) pine forests were used. Burr, Dagum, and J ohnson SB
distributions were applied due to their flexible properties. Result showed that dbh distributions of
the 10~15 yr (AG1) and the 40~45 yr (AG2) stands are left tailed and the 60~65 yr (AG3) stand is
normally skewed. h distribution of AG1 and AG3 are left tailed while AG2 show no obvious
distribution shape, due to its discrete h distribution. Distribution study reveals that in the left tailed
forests, dbh and h distribution shapes were the best approximated by the Dagum distribution. In
case of normal distribution shape, Johnson SB is better than the Burr and Dagum. Based on the
results, it is concluded that dbh distributions of even-aged AG1 and AG2 forests are heavy left
tailed and forest structure tend to normal at the AG3. The h distribution is left tailed (AG1 and AG3)
if a forest h growth is not constrained for space, while it will become discrete in a high density stand
(AG2).

2.2 Introduction
Forest managers often require information concerning the size-class distribution of a forest
stand, usually in the form of a tabulation of numbers of trees by dbh and h class. This dbh and h
distribution information is often used to predict volume production, the primary variable that forest
managers are interested in. dbh and h distribution models are also important in forecasting the range
of products that might be expected from a stand. The forest volume estimates are usually based on
the dbh distributions and for improved forest volume estimates, tree hs are used. Unmanaged forests
are used as a standard for comparison of different types of managed stands. Thus, detailed modeling
of the variation of dbh and h classes is required.
A wide range of pdfs have been used in forestry to model tree dbh and h structural
distributions (Bailey and Dell 1973, Hafleyand Schreuder 1977, Gove et al. 2008, Wang et al. 2010)
as well as age distribution (Lin et al. 2007).
Distribution having unimodal shape (single high point) is easy to simulate by traditional pdfs.
However, regular shaped distributions may take different shapes therefore one needs to choose
appropriate theoretical distribution function that can express the empirical distribution shape. In

10

forestry, more often flexible pdfs are used for forest size-structure modeling. For instance: Beta,
Burr, Dagum (inverse Burr), Gamma, J ohnson SB and Weibull.
Weibull is generally the most favored distribution used in forestry as it is easy to use and able
to describe both positive and negative skewness. Comparison study of lognormal and Weibull
distribution on a regular shaped birch forest dbh distribution structure show that Weibull was
superior (Tsogt et al. 2011a). However these studies did not include any other distribution
functions. According to Lindsay et al. (1996) study Burr distribution outperforms Weibull
distribution in its ability to drive dbh distribution and Dagum has the ability to fit rotated sigmoid
dbh distribution (Gove et al. 2008). Both Weibull and Burr distributions are family of Dagum
distribution and the Dagum distribution has a wider region of applicability then either of two
(Lindsay et al, 1996).
Hafley and Schreuders (1977) study showed that J ohnson SB distribution gave the best
performance, while normal, lognormal and gamma distributions were inferior to the Weibull and
beta distributions in terms of their general performance over variety of even-aged stands. Generally
beta was the second best fitting distribution and Weibull was third best. From the viewpoint of
practical application, they believed J ohnson SB distribution has important advantage over the beta
distribution. The reason being that it spans a slightly broader range of the skewness-kurtosis space
than the beta distribution (ie., it covers, in addition, the region between the lognormal and the
gamma).
Based on pdfs, few studies have paid attention to different species (birch and larch) of
Mongolian forests dbh and h (Tsogt and Lin 2012, Tsogt et al. 2011a, b, c) distribution structures.
But, no one has studied dbh and h distribution of pine forests. The purpose of this study is to find
the appropriate distribution for modeling dbh and h distributions of even-aged pine forests using
Burr, Dagum and J ohnson SB distributions.

The distribution models investigated in this paper were fitted to dbh and h measurements of 3
pine forest plots: P1 and P2 in Khuder soum Selenge aimag and P3 in Shariin gol soum Darkhan-
Uul aimag. The species found in 3 forest-plots were pure pine (Pinus sylvestris L.). dbh and h were
measured for all trees higher than 1.3 m. General characteristics of the study plots are given in
Table 2.1 and 2.2

11

Table 2.1 Inventory information of the study site
Plot #
Year of
Plot size
Number of trees Average age
inventory by plot area in a hectare of plot trees
P1 2005 50 x 40 m 239 1195 10~15
P2 2005 40 x 40 m 250 1563 40~45
P3 2001 50 x 20 m 96 960 60~65

Table 2.2 Descriptive diameter and height statistics of the study site
Plot ID Variables Mean Standard deviation Min-Max Skewness Kurtosis
P1 dbh (cm) 16.10 6.03 2.50-30.00 -0.23 -0.75
h (m) 14.06 3.58 3.30-19.66 -0.94 0.21
P2 dbh (cm) 15.80 6.18 2.50-30.00 -0.17 -0.82
h (m) 11.88 3.25 6.19-15.02 -0.82 -0.91
P3 dbh (cm) 22.10 5.58 8.30-36.80 -0.02 -0.1
h (m) 15.89 2.58 6.11-21.62 -1.36 3.67

P1 and P2 forests result from the expansion of the forest area caused by favorable site
conditions and abundant parent material (seed sources and advance regeneration). The P1 stand
does contain old trees, however they are not inside our plot area. P2 is the neighbor of an older
forest stand. P3 has no obvious seed source. Perhaps it had already died and/or been removed from
the stand.
All 3 plots, including dbh and h, were negatively skewed but P3 h is very close to zero. Which
means P3 h would be a symmetric distribution and the rest of the distributions are left tailed. In
addition, P1 and P2 dbhs are relatively short tailed and the hs for P1, P2, and P3 are long tailed.
Four of 6 kurtosis values are negative. P1, and P2 dbh, and P2 h are more flat than the normal
distribution according to their kurtosis value. The P3 dbh is less flat and similar to normal
distribution The P1 and P3 h are peaked distribution. In fact, P3 h is high peaked distribution.

2.3.2 Data analysis
This study intended to model dbh distribution of the forest plots with 3 functions (Appx 1),
including Burr, Dagum, and J ohnson SB. The goodness-of-fit test descriptions are given in heading
1.3.2, chapter 1.

12

2.4 Results
The parameters, estimated by likelihood with the measured data are shown in Table 2.3. All
parameters were successfully computed. The observed and estimated frequency distributions of
each plot are shown in Fig. 2.1 and Table 2.4. The dbh distributions are plotted by widths of 2 cm
classes and the h distributions by 1 m classes. The histogram represents the observed distribution
and the curves represent the estimated distributions. Clearly, the Dagum is superior to the other two
distributions in terms of their general performance over the left tailed distributions. The Burr is
second best and it was not failed at any test statistics of goodness-of-fit but in the case of P2 h all 3
models failed. Because of discrete distribution characteristic, none of them could model the P2 h
distribution among the 3 pdfs. The fact that Burr and Dagum lines fall rather close to each other in
comparison with the J ohnson SB distribution helps to explain why sets of data can often be fitted
equally well (or equally poorly), by either of these distributions. For the P3 dbh, J ohnson SB
distribution showed the best result. In this study, either Burr and Dagum or J ohnson SB were the
best fitted models. In each case the Weibull model did not show the best or the second best
performance of goodness-of-fit test (results not shown) and generally was similar to the Burr
distribution results. Therefore we decided to not include the Weibull distribution in this study.

Table 2.3 Parameter estimates of the dbh and h distribution models for pine forests, P1-P3
Plot #dbh h
Burr Dagum J ohnson
SB
Burr Dagum J ohnson
SB

P1 k =96.872 k =0.12207 = -0.33446 k =3422.4 k =0.10296 = -1.0667
= 4.3718 = 16.329 = 1.0467 = 489.29 = 32.122 = 0.83265
= 70.738 = 24.025 = 30.387 = 1328.3 = 18.134 = 18.118
=-6.4707 = 0 = -1.0791 = -1290.7 = 0 = 0.79728
P2 k =108.79 k =0.09438 = -0.23789 k =1045.9 k =0.00778 = -0.55124
= 3.8956 = 15.688 = 0.99499 = 4.7765 = 442.93 = 0.27967
= 76.963 = 22.283 = 29.794 = 56.121 = 15.101 = 9.1613
= -5.0039 = 2.3031 = -0.52346 = 0 = 0 = 5.5655
P3 k =8.253 k =0.43323 = -0.30612 k =3.0013 k =0.398 = 0.51806
= 4.7694 = 9.8894 = 4.3887 = 1.4860E+6 = 147.07 = 3.5092
= 36.962 = 25.513 = 99.144 = 2.4671E+6 = 123.03 = 84.116
= 0 = 0 = -29.188 = -2.4671E+6 = -105.52 =-16.76

13

2.5 Discussion
In our study, young pine forests dbh distributions are heavy left tailed for P1 and P2, and
normal for P3. h distribution of P1 and P3 are left tailed while P2 show discrete distinct h groups.
P1 and P2 dbh distributions are identical even though their geographical locations and ages are
different. That means, their regeneration and growth patterns are same, according to their dbh
distribution. In P1 and P2, most stems were established shortly after the disturbance, and this still
left enough growing space for stems to become established later, during 1990-1995 (P1) and 1960-
1965 (P2). During the 5 yr many seedlings successfully grew and survived after regeneration; those
are 10 yr old in P1 and 40 yr old in P2. They are recognizable on the dbh distribution P1 and P2 of
Fig. 2.1, where plateaus are on the left side of the peak. The left tailed dbh distributions indicate that
growth space is still sufficient (in free growth situations) for the major trees which occupy the main
canopy of the stand and the dbh distribution structure will not change until the forest reaches the
maximum capacity of stem number and mean size ratio. Once, a forest reaches maximum tree
density-size ratio, individual tree growth can continue only if the number of individuals is reduced.
Thus, the forest dbh distribution structure may change with different distribution shapes depending
on the size of tree stems removed from the stand.
P3 dbh distribution indicates that some trees dominate the major population while some are
suppressed. In theory, dominant trees are located in more favorable microclimate conditions than
suppressed trees or they may have just inherited good genetic materials. Either way, dominant trees
grow bigger, faster and stronger while other trees will grow more slowly and they will be
suppressed. However, still the dbh structure of main population is well normal in P3 forest.
h distributions of P1 and P2 are very different. The plateau is clearer on h distribution of P1
but P2 is not showing any distribution shape. P1 is just in crown closure stage, so the explanation is
same as for the dbh distribution of P1. Because P2 stem density is higher than P1 and P3, the crown
development is strained. When a tree respiration and foliage and root growth consume all
photosynthates, normal h growth is not maintained. Tree h repression occurs first in trees with small
crowns. In our case, those are 6 and 7 m high trees in P2 h distribution. It can occur in individual
trees as they become suppressed or in whole stands as they approach stagnation (Eversol 1955).
This explains why the P2 average h is shorter than P1. h distribution structure of P3 indicates that
the major trees h growth has not yet stopped, while a few trees h growth is constantly suspended.
That results in long left tailed distribution structure.

14

Fig. 2.1 dbh (left) and h (right) model comparisons for pine forests, study plots P1-P3. The
histogram represents the observed distribution and the short dashed line (Burr), long dashed line
(Dagum), and solid line (J ohnson SB) show the estimated distributions.

0
50
100
150
200
2 4 6 8 1012141618202224262830
N
U
M
B
E
R

O
F

S
T
E
M
S
,

h
a
DBH, cm
P1 dbh
0
50
100
150
200
250
3 5 7 9 11 13 15 17 19
N
U
M
B
E
R

O
F

S
T
E
M
S
,

h
a
H, m
P1 h
0
50
100
150
200
250
2 4 6 8 1012141618202224262830
N
U
M
B
E
R

O
F

S
T
E
M
S
,

h
a
DBH, cm
P2 dbh
0
90
180
270
360
450
6 7 8 9 10 11 12 13 14 15
N
U
M
B
E
R

O
F

S
T
E
M
S
,

h
a
H, m
P2 h
0
40
80
120
160
8 1012141618202224262830323436
N
U
M
B
E
R

O
F

S
T
E
M
S
,

h
a
DBH, cm
P3 dbh
0
50
100
150
200
250
6 8 10 12 14 16 18 20 22
N
U
M
B
E
R

O
F

S
T
E
M
S
,

h
a
H, m
P3 h

15

Table 2.4 Goodness-of-fit and ranking (rank in parentheses) of Burr, Dagum and J ohnson SB
distributions for the empirical dbh and h distributions as measured by MLE criterion ( = 0.05)
Plot #
Distribution
Kolmogorov-Smirnov Anderson-Darling Chi-Squared
Structure statistic statistic statistic
P1 dbh
critical value 0.08784 2.5018 14.067
Burr (4P) 0.06589 (3) 1.1113 (2) 7.9501 (2)
Dagum (3P) 0.04047 (1) 0.40807 (1) 7.7654 (1)
J ohnson S
B
0.04462 (2) 4.3858 (*) -
P1 h
critical value 0.08784 2.5018 14.067
Burr (4P) 0.07234 (2) 1.4021 (2) 13.7770 (2)
Dagum (3P) 0.06231 (1) 0.95645 (1) 8.4542 (1)
J ohnson S
B
0.07286 (3) 38.763 (*) -
P2 dbh
critical value 0.08589 2.5018 14.067
Burr (4P) 0.06888 (3) 1.3299 (2) 9.136 (2)
Dagum (4P) 0.03904 (1) 0.26976 (1) 4.6169 (1)
J ohnson S
B
0.04521 (2) 4.4621 (*) -
P2 h
critical value 0.08589 2.5018 7.8147
Burr (3P) 0.25849 (*) 22.676 (*) 30.646 (*)
Dagum (3P) 0.26938 (*) 25.977 (*) 14.728 (*)
J ohnson S
B
0.16567 (*) 142.13 (*) -
P3 dbh
critical value 0.13675 2.5018 12.592
Burr (3P) 0.04506 (2) 0.20612 (2) 3.8621 (3)
Dagum (3P) 0.05222 (3) 0.24601 (3) 2.7432 (2)
J ohnson S
B
0.04361 (1) 0.18449 (1) 0.9937 (1)
P3 h
critical value 0.13675 2.5018 12.592
Burr (4P) 0.04699 (2) 0.3492 (2) 1.8253 (2)
Dagum (4P) 0.04166 (1) 0.22303 (1) 1.6842 (1)
J ohnson S
B
No fit

* assumption rejected at = 0.05

16

2.6 Conclusion
Naturally generated, even-aged young pine forest dbh distribution is left tailed in Western
Khentii. Few distribution have the ability to simulate left skewed distribution. All the 3 pdfs, which
applied in this study, can simulate both left and right skewness. However, the Dagum distribution
was superior according to test statistics. The Burr, generally, was good enough to simulate the
distributions. The J ohnson SB was only the best in case of normal distribution. In fact, J ohnson SB
is good at simulating heavy tailed shapes which is not given in this study. In addition, forest h
distribution is more sensitive than dbh distribution to reveal forest structural relationships within
forest stands.

17

Chapter 3
3 Modeling diameter and height distributions of a spruce-larch mixed forest

3.1 Abstract
We derived diameter breast height (D) and height (H) structural distribution model for
Siberian spruce and Siberian larch (Picea obovata and Larix sibirica) mixed and uneven aged stand.
The aim of this study is to fit distribution models for irregular shaped uneven aged mixed stands D
and H distributions. For this reason we apply simple unimodal and mixture multi modal distribution
models. More than dozen probability density functions are used to describe the D and H
distributions of the species groups and entire forest stand. Moreover, we tried to explain their
distributional structures from ecological perspective. Inventory was conducted in river valley of
near Gachuurt village, Ulaanbaatar, Mongolia. The goodness of fit test were accompanied with
Kolmogorov-Smirnov, Anderson-Darling and Chi-Squared (
2
) tests for distribution models. Root
mean square error, and chi-squared tests were used for comparing how closely predict given D and
H distributions can predict by simple and mixture distribution models. The study result shows that
all distribution models goodness of fit test accepted. Mixture distributions are more similar to
observed distributions than simple distribution models according to the root mean square error
(RMSE) and
2
test. Both RMSE and
2
tests values are bigger in simple distribution case than in
mixture distribution case. The overall study result indicates that the mixture models were suitable
for modeling irregular and bimodal structural distributions. Mixture models have a potential to
characterize tree structural diversity.

3.2 Introduction
Siberian spruce grows on marshy soils around rivers and bogs (Dilis, 1981), and is able to
withstand strong shading (Kellomaki, 1987). It can regenerate under canopies of all forest types
(Kujala, 1924). Often spruce forests grow in riverside of mountain valley and either grow by pure
stand or mixed with larch in Mongolia (Tsedendash, 2007). Forest fire frequently occurred during
spring due to its dry climate in Mongolia. Spruce seedlings are fire intolerant. The study plot of
spruce-larch mixed forest grows separately from main forest in river valley. Therefore, spruce trees
successfully develop in river valley. But, still spruce trees are easy to die in river valley after
nutrient layer of soil is washed away by flooding and their thin roots are exposed in sun.
Spruce forests are less occur in the country and separated within each other as patches in
Mongolia. However, mixed spruce-larch forests are common in Western Khentii (Dugarjav, 2006).

18

The spruce forest grows along the southern boundary of its areal distribution in Khentii ridge.
Spruce forest productivity and natural regeneration become low into transition zone. So, the spruce
forests in transition zone are in developing regress succession. Spruce trees mix with pine and larch
forests and sometimes only juvenile spruce trees occur in forests. This can be explained by the
forest composition change. However, this is also a sign of the transition zone effect. Spruce and
larch forests belong to cool taiga and areal distributions follow permafrost region. And as a
consequence of global warming and anthropogenic negative impact, the areal distribution is now
reducing. Therefore, elimination of their distribution caused by dry steppe effect and recent general
dryness.
Studying spruce-larch mixed forest size structure may offer better understanding about what
is happening in the mixed stand in transition zone. This understanding could lead us to develop a
proper forest management policy for uneven aged mixed forest.
Strong intensity disturbances or high intensity regeneration may change regular D structure to
bimodal structure. If disturbance happens repeatedly in same forest, then the D distribution would
forms multimodal shape. In forestry, the probability density functions, such as Weibull (Bailey and
Dell, 1973; Lin et al., 2007), Gamma (Nelson, 1964), Burr (Lindsay et al., 1996), Johnson S
B

(Hafley and Schreuder, 1977) etc. are well known and broadly used. These distribution functions
have unimodal shape and are weak in multimodal or irregular shaped distributions.
Mixture model is better at modeling multimodal forest horizontal and vertical structural size
distributions. J aworski and Podlaski (2011) found the usefulness and applicability (Zhang and Liu,
2006) of two-component and three-component models for describing D distributions in mixed stand
(Shunzhong et al., 2006). A frequency distribution made up of two or more component distributions
is defined as a mixture distribution (Liu et al., 2002; Zasada and Cieszewski, 2005; Zhang et al.,
2001; Zhang and Liu, 2006).
Researchers focused on understanding the structure, function, and productivity of larch
forests(Danilin, 1995; Osawa et al, 2010; Tsogt, 1993), and D distribution of larch forests (Khongor
et al., 2011a and b), but spruce-larch mixed forest D and H distributions have not been studied in
Mongolia. Therefore, it is necessary to model and explain D and H distributions of spruce-larch
mixed forests.
The objective of this study was to investigate the suitability of either simple or mixture
functions in modeling the D and H distributions of spruce-larch mixed forest in river valley of
forest-steppe zone.

19

3.3 Material and Methods
Study plot was selected from Gachuurt village 10 km to North East in river valley surrounded
by grassland, in green zone Ulaanbaatar, Mongolia and the geographical location is 4800'30.9'' N
and 10712'49.5'' E with altitudinal H 1495 m above sea level. The forest trees appear in multi
cohort and double stratum. Composition of the stand is mixed spruce and larch with few birch and
willow. 60% of live trees is spruce, 40% is larch and less than 1% is birch and willow. Birch and
willow trees are not included in the study. Spruces in the first stratum are 160 years old and larches
are 80-100 years old. And in the second stratum, both species are 50 years old. There are few larch
trees are 180-200 years old. The inventory was conducted summer in 2009. Composition of the
stand is mixed spruce and larch with few birch and willow. For D and H distribution modeling only
live trees were analyzed (Table 3.1). Plot size is 0.16 ha (40x40 m). This forest stand is so small and
there are some these kind of spruce-larch mixed forests grow along that river valley but occurrence
of formed stand trees along the river valley are distant around 100 m to 1 km. The distance between
study plot and the nearest forest is 374 m in mountain slope. It is pure larch forest. The forest is
natural stand, and no management activity had been taken previously. But, because livestock
grazing do not give chance to survive young seedlings around the study area, even if seedlings were
regenerated.

Table 3.1 D and H characteristics of the spruce-larch forest in the study site

Number of
observation
Diameter (cm) Height (m)
Mean Min-max Mean Min-max
Spruce
149 live
95 dead
9.8
3.0
2.2-30.2
1.0-18.5
9.2 3.0-18.5
Larch
99 live
12 dead
13.1
7.5
2.2-24.2
2.2-25.9
12.8 3.0-19.0

3.3.2 Data analysis
Choosing appropriate distribution model is problematic. In order to do that first you have to
understand the distribution shape and depend on the shape of the distribution you could choose the
model. There are similar distributions but sensitive in specific subtle skewness and kurtosis. So, it
will ask you to know about distribution models characteristics. However, statistical software
packages provide letting us to choose distribution models without prior knowledge.

20

There is only few distribution models used in forestry and have their biological meanings and
vast amount of distribution models which used in engineering field have no biological meaning. But,
sometimes distribution models which have no biological meaning could describe forest structural
distributions better than broadly used one according to their goodness of fit test. In that case we
could use those distribution functions to model forest structures with higher fitness.
We have chosen the best fitted distribution function among Burr, Dagum and J ohnson S
B
functions (Appx 1) according to their goodness of fit test for modeling dbh and h distributions
(Table 3.2 and 3.3). The goodness-of-fit test descriptions are given in heading 1.3.2, chapter 1.

3.3.3 A flexible mixture model for dbh structure modeling
A typical modeling method of the dbh structure of a forest stand is single distribution
modeling or unimodal approach (SDM approach). It assumes that the dbh distribution has a single
peak and can be fitted with a particular distribution function. If the tally dbh histogram has
obviously more than one peak, the dbh structure or distribution is irregular and can be divided into
several sub-groups. A modeling of irregular dbh distribution is mixture (multi-single) distribution
modeling or multimodal approach (MDM approach). It applies first a single model to fit each of the
sub-group dbh distribution and then integrates all of the single models into a multi-single model or
mixture model. It is supposed that a simple distribution model is potentially not able to cover the
high diverged dbh histogram while mixture models have better estimation ability because a mixture
model can cover the detailed outstanding features of the dbh variation.
The proposed multimodal approach is a four-step modeling process. First, to examine the
breakpoint(s) or gap(s) that occurred in the whole-range dbh histogram of a sample plot. A
continuous pdf curve is theoretically a common feature of any arbitrary distribution function. Thus
an apparent discontinuous point existing in the curve can be defined as a breakpoint. Next, the
number of sub-groups is determined by the number of breakpoints plus one. The corresponding dbh
of a breakpoint is assigned as a boundary of a sub-group pdf curve. Third, a sub-group is considered
as a histogram with a particular distribution and is fitted with each of the five theoretical pdf
functions. A suitable pdf function is suggested based on the goodness of fit test. Finally, a mixture
model for the whole range dbh histogram is integrated directly by the suggested single model of
each sub-group histogram.
Suppose a multimodal dbh distribution is composed of k sub-groups and each of the sug-
group has m dbh classes. Then the generalized mixture distribution model, f(x) can be expressed as
Eq. 3.1,

21

= = =
= =
k
j
m
i
ij ij j
k
j
j j
x f n w x f w x f
1 1 1
) ( ) ( ) ( Eq. 3.1
where
ij
n is the number of stems of the i
th
dbh class in j
th
sub-group (adopted from Zhang and Liu
2006). By the definition of probability, the area under the pdf curve of a single distribution function
equals 1. So, the total number of stems would be scaled to 1. Once a single distribution model is
derived successfully, the predicted number of stems in each diameter classes of that specific sub-
group can be determined by multiplying the estimated probability and the observed total number in
that sub-group.

3.3.4 Model comparison
RMSE and
2
tests were used for comparing how closely predict by unimodal and mixture
distribution models for a given D and an H distributions. RMSE is a measure of the differences
between values predicted by a model and the values observed from the measurement.

3.4 Results
The stand is assumed to consist of two individual generations which are spruce and larch trees.
It is feasible to say that there are two generations of trees in both species according to D and H
structures of stand in Fig. 3.1. By D class, the first generation of spruce trees are 14-30 cm and larch
trees are 10-26 cm, and the second generation of spruce trees are 2-12 cm and larch trees are 2-8 cm
(Fig. 3.1b and 3.1c). By H class, the first generation of spruce trees are 13-19 m and larch trees are
9-19 m, and second generation of the spruce trees are 3-12 m and larch trees are 3-8 m (Fig. 3.1e
and 3.1f). In Fig. 2a and 2d we can see that the first generation of both spruce and larch stems
clustering distributed around 12-30 cm and the second generation stems clustering distributed
around 2-11 cm by D class; by H class the first generation clustering distributed around 12-19 m
and the second generation of stems clusteing distributed around 3-11 m. If we look separately,
species by species, both spruce and larch D and H distributions are also bimodal. But, their stem
clustering appeared in different places in the D clustering. D distribution of spruce stems is right
tailed. It is very similar to reverse J shape distribution. The major stem clustering is around 4-10 cm
and then slowly decreasing until 30 cm except gape at 12 cm. H distribution of spruce trees is right
tailed and the first generation of stems clustered at two distinct positions which are 14 m and 18 m.
The second generation of spruce trees D distribution structure seen that the right tail is heavier than
left tail. Both D and H right tailed distributions show that regeneration status of spruce trees is good.

22

For instance, we can see in Fig. 1b and 1e D distribution that there are many smaller trees and big
stems which gradually declined in number.
D distribution of larch trees is left tailed, and stem clustering is around 4-8 cm and 14-16 cm.
However even though larch trees regeneration peaked at 6 cm, it is still smaller than first generation.
H distribution of larch trees is left tailed and also both humps are left tailed. The first generation of
larch stems holds the most proportion of all the stems. From both D and H distribution shape, larch
trees regeneration status is not so good (Fig. 3.1c and 3.1f).
The parameter estimates of the simple and mixture models are given in Table 2 and 3, Noting
that spruce-larch forest D and H distributions are bimodal. The predicted frequencies by D and H
classes were obtained from simple and mixture model. Plus, single and mixed species are tested in
these models. The predictions from each model were compared with the observed frequencies. The
RMSE, and
2
test were computed for each model and each condition (spruce-larch mixed, spruce
and larch) (Table 3.4). The observed frequency distribution (histograms) and the simple and
mixture prediction curves are illustrated for each condition (Fig. 3.1).
The simple models were definitely not flexible enough to fit the distribution at all. These
simple functions missed the second peak as well as the valley between the two peaks. However all
models passed their test statistics for Kolmogorov-Smirnov,
2
and Anderson-Darling (Table 3.2
and 3.3). The mixture models fit to the plot data better, while single mode models were not enough
to characterize the distribution (Table 3.4). Fig. 1 shows that simple models under-predict small and
large trees and over-predict middle sized trees.

23

Fig. 3.1 Model comparison for the Spruce-Larch mixed forest. The histogram represents the
observed diameter at breast height (dbh) and height (h) distributions with simple distribution model
(solid line), mixture distribution model (dashed line). a) dbh of spruce-larch mixed trees fit by
simple distribution, Burr (4P) and fit by mixture distribution, Burr (4P) and J ohnson S
B.
b) dbh of
spruce trees fit by simple distribution, Burr (4P) and fit by mixture distribution Burr (4P) and
J ohnson S
B
. c) dbh of larch trees fit by simple distribution, Dagum (3P) and fit by mixture
distribution, double J ohnson S
B.
d) h of spruce-larch mixed trees fit by simple distribution J ohnson
S
B
and fit by mixture distribution double J ohnson S
B
. e) h of spruce trees fit by simple distribution,
Dagum (3P) and fit by mixture distribution Dagum (4P) and J ohnson S
B
. f) h of larch trees fit by
simple distribution, Dagum (3P) and fit by mixture distribution Dagum (3P) and Johnson S
B
.
0
50
100
150
200
250
300
350
2 4 6 8 1012141618202224262830
N
u
m
b
e
r

o
f

s
t
e
m
s

(
N
/
h
a
)
Diameter (cm)
0
40
80
120
160
200
240
3 4 5 6 7 8 910111213141516171819
N
u
m
b
e
r

o
f

s
t
e
m
s

(
N
/
h
a
)
Height (m)
0
50
100
150
200
250
300
2 4 6 8 1012141618202224262830
N
u
m
b
e
r

o
f

s
t
e
m
s

(
N
/
h
a
)
Diameter (cm)
0
30
60
90
120
150
180
3 4 5 6 7 8 910111213141516171819
N
u
m
b
e
r

o
f

s
t
e
m
s

(
N
/
h
a
)
Height (m)
0
30
60
90
120
150
180
2 4 6 8 10 12 14 16 18 20 22 24
N
u
m
b
e
r

o
f

s
t
e
m
s

(
N
/
h
a
)
Diameter (cm)
0
20
40
60
80
100
120
3 4 5 6 7 8 910111213141516171819
N
u
m
b
e
r

o
f

s
t
e
m
s

(
N
/
h
a
)
Height (m)
a d
c
b e
f

24

Table 3.2 Parameter estimates of diameter at breast height distribution models for the spruce-larch mixed, spruce and larch trees (=0.05)
Species and model type
Diameter at
breast height
range(cm)
Distribution
Parameters
Kolmogorov-Smirnov Chi-Square Anderson-Darling
Statistic
Critical
value
P value Statistic
Critical
value
P value Statistic
a

Simpledistribution
model of spruce-larch mixed
trees
2-30 Burr (4P) 0.08474 0.08623 0.05353 21.071* 14.067 0.00367 2.278
Mixturedistribution
trees
2-11 Burr (4P) 0.05083 0.11015 0.80804 10.628 14.067 0.15569 0.48412
12-30 Johnson S
B
0.04745 0.1319 0.96166 2.3259 12.592 0.88742 0.27905
Simpledistribution
model of sprucetrees
2-30 Burr (4P) 0.08483 0.11125 0.22104 23.76 14.067 0.00126* 1.5427
Mixturedistribution
2-12 Burr (4P) 0.05504 0.13067 0.88093 1.3878 12.592 0.96659 0.29153
13-30 Johnson S
B
0.07312 0.20517 0.96614 0.97413 11.07 0.96463 0.14632
Simpledistribution
model of larch trees
2-24 Dagum(3P) 0.05769 0.13469 0.87759 3.8462 12.592 0.69748 0.37686
Mixturedistribution
2-10 Johnson S
B
0.10927 0.22425 0.75696 1.5339 7.8147 0.67448 0.39013
11-24 Johnson SB 0.06094 0.15755 0.93698 3.7463 12.592 0.71096 0.27327
a
- the critical value of Anderson-Darling statistic used is 2.5018 for all distributions at =0.05
* - assumption rejected at =0.05

25

Table 3.3 Parameter estimates of tree height distribution models for the spruce-larch mixed, spruce and larch trees (=0.05)
Trees height
range(m)
Distribution
Parameters
Statistic
Critical
value
P value Statistic
Critical
value
P value Statistic
a

Simpledistribution
trees
3-19 Johnson S
B
0.0588 0.08623 0.34451 - - - 19.896*
Mixturedistribution
trees
3-11 Johnson S
B
0.07333 0.11396 0.4102 5.8852 14.067 0.55322 0.78883
12-19 Johnson S
B
0.06157 0.1319 0.79315 5.723 12.592 0.45492 0.27435
Simpledistribution
3-19 Dagum(3P) 0.06995 0.11125 0.43954 6.9528 14.067 0.43381 1.1546
Mixturedistribution
3-12 Dagum(4P) 0.06203 0.12663 0.74418 4.864 12.592 0.56137 0.61725
13-19 Johnson S
B
0.07644 0.21273 0.96341 - - - 4.1025*
Simpledistribution
3-19 Dagum(3P) 0.06385 0.13469 0.79034 2.3502 12.592 0.88484 0.3465
Mixturedistribution
3-8 Burr (3P) 0.10555 0.2749 0.93632 0.22348 5.9915 0.89428 0.35917
9-19 Johnson SB 0.06748 0.14868 0.8301 3.5696 12.592 0.73468 0.28456
a
- the critical value of Anderson-Darling statistic used is 2.5018 for all distributions at =0.05
* - assumption rejected at =0.05

26
Table 3.4 RMSE and
2
test of the simple and the mixture models for the spruce-larch mixed,
spruce and larch trees (N/ha)
Diameter Height
RMSE Chi-Square test Bias RMSE Chi-Square test Bias
Spruce-Larch model 17.81 130.00 4.34 13.94 -3.50 -3.38
Mixture Spruce-Larch model 5.97 22.08 0.48 9.39 38.68 1.38
Spruce model 10.12 131.52 4.14 8.63 134.68 2.86
Mixture Spruce model 5.22 21.11 1.45 7.02 53.97 1.58
Larch model 6.65 40.35 0.53 4.76 41.59 0.26
Mixture Larch model 5.00 13.37 -0.05 3.77 13.15 0.18

For each condition, mixture models produced satisfactory fitting results. All mixture
distribution models fit to observed D and H frequency distributions better than simple distribution
according to the RMSE and
2
tests (Table 3.4) and yield similar predictions across tree Ds and Hs.
On the other hand, the simple model did not fit the forest D and H distribution well and definitely
missed the valley portion of the distribution (Fig. 3.1).

3.5 Discussion
For a mixed conifer forest, size distribution is often a better predictor of future forest
composition (Veblen, 1992). In this sense, we can guess the future stand composition on Fig. 3.2.
Species composition remained relatively constant in two strata. The current size structure suggests
that the stem composition would be remained relatively constant in the future. Understory size
classes support this conclusion. Comparing with larch, many of the spruce trees were died and most
of them took place in small D class around 3 cm (Table 3.1). However, there might be changes in
the stem composition of the study plot, those changes would occur in the future under a continued
policy of herbivores suppression. On the study plot, recruitments around 100 years later appeared
according to two species age intervals. The first stratum of spruce trees were grown 110 years ago
and this stratum might have been the second stratum when they were 50 years old. The oldest ones
among the age-known trees of larch in this plot are 180-200 years and regeneration lapse is 80-120
years. Also these first-stratum trees might have been second strata when they were 50 years old and
after new mass regeneration, those are occupied the first stratum. The current mortality rate of
spruce trees in the second stratum suggests that larch trees will increase their dominance of the
stand in the future.

27
28 24 20 16 12 8 4
40
30
20
10
0
Diameter (cm)
N
u
m
b
e
r

o
f

s
t
e
m
s
Spruce diameter distribution
Larchdiameter distribution
18 15 12 9 6 3
25
20
15
10
5
0
Height (m)
N
u
m
b
e
r

o
f

s
t
e
m
s
Spruce height distribution
Larchheight distribution

Fig. 3.2 Spruce and larch trees a. D and b. H structures. a. Spruce trees are dominant in D 2-10 cm
while larch trees are dominant in mid D classes around 12-16 cm and then equally distributed
around 18-24 cm and around 26-30 cm only spruce trees inventoried. b. Spruce trees are dominant
3-11 m H while larch trees are dominant in 12-19 m.

H distributions of two species show that nursing benefit from larch trees to spruce trees (Fig.
3.2b). Siberian spruce trees are shade tolerant and they can regenerate under strong shading
(Kellomaki, 1987) canopies of all forest types (Kujala, 1924). Structural distributions of both D and H
indicated that larch trees are followed by spruce trees. Probably, though live spruce trees are
majority of the stems, still larch trees are dominant in big size H distribution and mortality rate is
higher in spruce 39% of died stems. Because of their shade tolerance, they can normally generate
under closed canopy and then go through the process of self thinning (Kimmins, 2004; Yoda et. al.,
1963) in overcrowded trees due to shortage of nutrients, water or light. Specially, competition for
survival is more significant among juvenile trees.
Siberian spruce crown shape is conical with drooping branchlets. While the larch is having
conic crown when young, and then crown become broader with age. The main branches are
horizontal to upswept and the side branches often pendulous. In the understory level two species,
they use same source of light and mostly diffuse light or low angle of light. Dominant and
codominant trees in lower strata often have large crowns even though they are in high shade. These
dominant trees do not grow much while in the understory; however, if they have large crowns with
many living buds, their crowns can quickly expand and they grow rapidly in H if provided by space
(Hoyer, 1980; Oliver, 1976) and they do not killed by the shock of open space release.
Larch trees are sun trees. Therefore they do not die once released in open space where
exposed direct sun light. Instead of that, their crown developed to become broader and bigger.
While spruce trees suffered by sudden exposure of high angle sun light, therefore the trees die in
sudden light exposure. Their general H growth strategy is in codominant place for reduced angle of

28
sun light. However, the gradually developed H growth of spruce trees is resistant to direct sun light.
This conclusion can be seen in Fig. 3.2b in H distribution of spruce and larch trees 15-17 m.

3.6 Conclusion
This study showed that how one species of trees affect the patterns of another species using
distribution models. To drive forest distributional structure, first, it is necessary to understand that
whether the distribution is simple unimodal or irregular shaped multimodal. Appropriate simple or
mixed distributions could apply depend on the distribution structure. In our study the mixture model
was more flexible in fitting various irregular D and H distributions of uneven-aged mixed forest
stand, while the simple distribution models missed the irregular fluctuation of the both D and H
distributions due to different generations. Furthermore, simple distribution models are not useful to
explain ecological relationship of two species in forest stand while appropriate mixture models are
good to explain the relationship of mixed forest structure pattern.

29
Chapter 4
4 A flexible modeling of irregular diameter structure for the volume estimation of larch
forest stands

4.1 Abstract
Simple distribution models (SDM) have deficiencies in portraying irregular forest diameter
structure. This paper introduces mixture distribution models (MDM) to improve the estimation
accuracy of stand volume of Siberian larch Larix sibirica Ledeb. forests. Stand volume was
estimated by combining the suitable diameter mixture model and diameter-height model.
Appropriate mixture models are derived by integrating multiple SDMs of Burr III and XII, J ohnson
S
B
, Weibull or lognormal probability density functions (pdf) that satisfied the criteria of goodness
of fit tests. Results showed that the average bias of volume estimation for all of the study plots
using SDM and MDM approaches are under estimated by 6.93 and 2.42 m
3
,

respectively. Each of
the estimates is equivalent to an estimation error of 25.5918.18% and 7.082.97%. This suggests
that MDM approach is a more flexible and suitable modeling technique for forest volume
estimation, in particular for forests that have been frequently disturbed by natural events.

4.2 Introduction
During the past decades, the Weibull pdf has been widely applied to derive the frequency of
dbh classes of forest stands (Bailey and Dell 1973, Lindsay et al. 1996a) because the parameters of
this function can cover a large-variant frequency distribution. The scale and shape parameters of
Weibull distribution, for example, can be adjusted to describe the Gaussian distribution (the typical
even-aged forest) and the reverse J -shape distribution (the typical un-even aged forest). The Weibull
function particularly can be applied not merely for the growth or volume estimation but also for the
assessment of forest regeneration (Lin et al. 2007). The Burr types III (Burr 1942) and XII (Zimmer
and Burr 1963) distributions have been introduced to forest research by Lindsay et al. (1996b).
These distributions are inherently more flexible because they can cover a much larger area of the
skewness-kurtosis plane than the Weibull distribution (Lindsay et al. 1996b, Rodriguez 1977,
Tadikamalla 1980, Gove et al. 2008). That is because the Burr XII distribution can approximate to
the normal, gamma, lognormal, exponential, logistic, and several Pearson-type distributions
(Rodriguez 1977, Tadikamalla 1980). The J ohnson S
B
distribution has also been introduced into
forest research by Hafley and Schreuder (1977) and later its parameterization method is improved
by Rennolls and Wang (2005). The J ohnson S
B
distribution covers a region of the skewness-

30
kurtosis plane different from the Burr distributions (J ohnson 1949, Hafley and Schreuder 1977), and
it is very similar to the Beta and generalized Weibull distributions in nature.
In general, the modeling of dbh structure is to fit the inventoried dbh sample data with a
particularly suitable simple distribution function. But, the dbh structure of a forest stand is easily
influenced by factors such as the natural succession (growth competition among the species and
individuals), the disturbances caused by naturally climatic events (e.g. strong wind) and/or
anthropogenic activities (e.g. forest fire and illegal logging). As a result, the dbh structure may be
changed into a highly skewed irregular shape, and hence causes a potential problem of model
inadequacy for the simple distribution model. This might induce a significant bias of forest volume
estimation.
A frequency distribution consisting of two or more component distributions is defined as a
mixture distribution. Mixture distribution models (MDMs) have a better ability to present the
multimodal dbh structure of a forest stand. Zhang and Liu (2006) demonstrated the suitability of
mixture model for multimodal structure of mixed-species stands, highly skewed and highly
irregular shapes of uneven aged stands, and rotated sigmoid form of old growth stands. Podlaski
(2010a, 2010b) and J aworski and Podlaski (2012) also found the usefulness and applicability of
two-component and three-component distribution models for the characterization of stand structure.
In contrast to the tree dbh measurement, tree height is relatively more difficult and time
consuming to obtain in a forest environment. Fortunately an appropriate dbh-h model can be
derived based on the inventoried sample data to predict the other trees in the forest stands. Many of
the dbh-h models presented in the literature use a linear and/or a log-linear model form (Curtis 1967,
Wykoff et al. 1982). However, some researchers have found that the residuals of these log-linear
models are not normally distributed (Larsen and Hann 1987). On the contrary, the dbh-h
relationship works more like a nonlinear one. Fang and Bailey (1998) and Huang et al. (2000) have
successfully demonstrated the dbh-h relationship as concave-shape and S-shape models. It is
difficult to explain all dbh-h relationship by a single model (Dorado et al. 2006).
The purpose of this study is to find a flexible dbh distribution model and dbh-h model in order
to calculate a better estimation of forest volume. The dbh structure modeling is conducted in two
parts, unimodal distribution modeling and multimodal distribution modeling. Then each of the dbh
distribution models will be integrated with the derived dbh-h model of the forest stand to estimate
volume stock, and finally the volume estimates are then validated using the measured volume.

31
More than 70% of the Mongolia forest is made up of Siberian larch Larix sibirica Ledeb.
forests. These are distributed along the southern boundary of the Siberian boreal forests. The larch
forests are located in Mongon Morit soum, Tov aimag, Mongolia was selected as study site because
it has been established for natural forest inventory and forest dynamic research. This study site is
geographically distributed within the Kherlen river watershed which is located between
48
o
01855-20558N, 108
o
23493-46322E and ranges from 1523 to 1897 meters above the sea
level. Mean annual air temperature is -3.7
, monthly average air temperature is -27.7
in
J anuary, and 17.0
in J uly. Mean annual precipitation and humidity is around 252 mm and 64%
respectively. The forest types are mesophyte herbaceous, steppe characterized herbaceous,
Rhododendron dahuricum, herbs-Vaccinium vitis-idaea, and forest-meadow herbaceous larch
forests. Forests grow in permafrost soils seasonally or annually and forest stands grow on forest-
soddy soils and forest-meadow soils.
Two criteria designated for the determination of plot size are (1) a minimum area size of 0.04
ha (20 m by 20 m) for high density stand (stems >1500 in a hectare) and (2) a minimum number of
stems for low density stand (stems >500 in a hectare). As a result, the six plots used for this study
with an area between 0.04-0.20 ha (Table 1) should be able to represent the characteristics of the
larch forests in the Kherlen river watershed area. Ground survey data were collected in the summer
of 2008. The dbh of each tree in the plots were inventoried and tallied in accordance with every 2-
cm diameter class. The height of the trees was then measured using suunto clinometer for each dbh
class in the plots. According to the dbh distribution shown in Fig. 1, the larch stands studied can be
classified as one-age cohort stand (plot 4), two-age cohort stand (plots 1-3 and 6), and multi-age
cohort stand (plot 5). Based on the stem analysis with the disc sampled by Tsogt et al. 2009, the
average age of trees in the main canopy layers of those forest stands is between 55 and 300 years
old (Table 4.1).

32
Table 4.1 Descriptive statistics of the study plots.
Plot
Average age of
the main layer
Plot area
(ha)
Stems
counted
Stand density
(N/ha)
Average
dbh (cm)
dbh
range (cm)
Average
height (m)
1. 70 0.05 118 2360 11.6 2-64 13.1
2. 60 0.20 144 720 19.8 1-40 14.2
3. 55 0.10 135 1350 11.6 2-65 11.6
4. 60 0.06 258 4300 5.4 1-46 6.3
5. 300 0.20 120 600 25.2 1-56 9.2
6. 90 0.04 61 1525 16.5 2-77 15.5

4.3.2 Data analysis
A large variant of distribution functions have been used for the purpose of empirical data
modeling in the field of forestry. Burr III and XII, J ohnson S
B
, Weibull and lognormal distributions
were selected (Appx 1) for this study to model dbh structure due to their flexible and specific region
of applicability behavior. Characteristics of different distributions might be suited to specific plots
due to their varying degrees of skewness and kurtosis. The goodness-of-fit test descriptions are
given in heading 1.3.2, chapter 1.

4.3.3 dbh-h relationship modeling and assumption tests
The relationship between dbh and height of trees was analyzed using regression techniques.
Based on the sample data, exponential function (Eq. 4.1) and sigmoid function (Eq. 4.2) were
adopted to derive a suitable dbh-h empirical model for each of the plots. Three basic assumptions
related to the error (residual) term of the regression model were being examined. They are
normality distribution, independent, and constant variance (C-V). In this paper, the K-S test is used
for the normality examination; the Durbin-Watson (D-W) statistic is used to check if an
autocorrelation exists among the residuals. If the residuals are not correlated, the D-W statistic will
be 2, while a value below 1.50 or above 2.50 shows autocorrelation. C-V statistic is used to verify if
the variance of the error is constant across observations.

( ) ( ) x x f | o = exp 1 ) ( Eq. 4.1
|
|
.
|
\
|
+
=
|
o
0
exp 1
) (
x x
x f Eq. 4.2

33
4.3.4 Volume estimation and accuracy assessment
Generally, the volume of individual trees is determined using the Schumacher-Hall equation (vol=a
dbh
b
h
c
Eq. 4.3) with the generalized coefficients (
Table 4.2). These generalized parameters were well developed by Semechkin and Tetenkin
(1980) and have been widely applied to the larch forests and other tree species in Mongolia.

vol=a dbh
b
h
c
Eq. 4.3

Table 4.2 Coefficients of tree volume equation (with bark) by diameter at breast height and height
Coefficient
Larch
Larix
sibirica
Pine
Pinus
sylvestric
Cedar
Pinus
sibirica
Fir and Spruce
Abies sibirica and
Picea obovata
Birch
Betula
platyphilla
a 0.229067 0.356149 0.422366 0.407192 0.121815
b 1.75631439 1.91061796 1.93511358 1.94199866 1.79633106
c 1.04530318 0.96807707 0.94402384 0.94744137 1.24762399

The derived suitable dbh distribution models and dbh-h models were applied to calculate total
volume (vol
T
) estimation of forest stands in this study. Suppose the range of a whole dbh
distribution can be divided into m classes. The estimates of stand volume (evol
T
) were determined
by integrating the value of each dbh class (
i
dbh ) and its corresponding tree height (
i
h ), which is
derived by dbh-h model, and then times the number of stems (
i
f ) in the i
th
dbh class using

= =
= =
m
1 i
c
i
b
i i
m
1 i
i i T
) h dbh (a f ) (vol f evol Eq. 4.4. Since the real number
of stems and diameter for each tree in the plots are inventoried in advance, these measurements can
be used to determine the height of each individual tree and then summed up to determine the
measured volume of a stand (mvol
T
). Using mvol
T
as a standard value, the accuracy of evol
T

estimation by the MDM approach derived models and the SDM approach derived models were
evaluated and compared. The deviation of volume estimates (evol
T
-mvol
T
) and error percentage
(deviation/mvol
T
*100) were used as accuracy indicators.

= =
= =
m
1 i
c
i
b
i i
m
1 i
i i T
) h dbh (a f ) (vol f evol Eq. 4.4

34
4.4 Results
4.4.1 Simple dbh distribution models of larch forest
Among the study plots in Fig. 1, only the histogram of plot 4 can be well fitted by an ideal
single pdf model, i.e., lognormal distribution function. But the measures of goodness of fit tests
indicated that the Weibull and Burr types distribution functions are not able to yield satisfactory
results to the histogram of plot 4.
Though a double peaks pattern can be found in plots 1-3 and 6, there still exists a little
difference between these plots. For example, plots 1, 3, and 6 are similarl with regards to high
frequent small dbh stems, but plots 1 and 6 have a small amount of stems in the middle dbh classes
and plot 3 shows a small amount of stems at larger dbh classes revealing a little raised right tail.
These variations caused the rejection of a single model such as exponential, lognormal and Weibull
distributions. The failure to account for middle or large dbh stems might cause a serious shortage of
stand volume estimation.
The stems of plots 2 and 5 are obviously very low while they have a larger number of average
diameter stems than the other plots. This is probably due to more space in the stand being available
for the trees to support their growth. Plot 2 looks like a normal distribution with high kurtosis but it
still failed to pass the goodness of fit test due to the redundant stems in middle dbh class. The dbh
distribution of plot 5 is the most irregular one among the study plots because the plot has been
repeatedly disturbed by forest fire. According to our analysis, the Weibull and the normal
distribution functions are obviously not able to explain the multi-cohort distribution.
In fact, Burr types of distributions are available to draw five of the irregular shaped dbh
distribution plots. Although all of the plots have been examined and passed by the tests of K-S and
A-D method, plots 1 and 2 were unfortunately rejected by the
2
test. This result suggested that
plots 1 and 2 potentially seem not to be matched very well with the fitted models. The modeling
with other distribution functions were unable to produce results superior to these however. So, even
though the
2
test was not satisfactory the K-S and A-D tests did agree that the plots 1 and 2 are
with Burr XII distribution. The Burr types distribution functions can describe those irregular shaped
dbh distributions very well. This is potentially due to their skewness and kurtosis plane being much
larger or more flexible than the Weibull distribution. The unimodal approach based Burr XII and
Burr III models are therefore considered to be relatively suitable for the volume estimation of the
plots 1-3 and 5-6. Detail of the goodness of fit for the single pdf model (unimodal approach) can be
seen in Table 4.3.

35
Table 4.3 Measures of the goodness of fit tests for the unimodal approach derived dbh distribution
models.
Plot Model type
Statistic
Critical
value
p value Statistic
Critical
value
p value Statistic
a

1. Burr XII 0.08 0.13 0.42 13.18 12.59* 0.04 0.80
2. Burr XII 0.09 0.11 0.17 23.64 14.07* 0.00 0.93
3. Burr III 0.07 0.12 0.44 11.80 14.07 0.11 1.07
4. Lognormal 0.07 0.09 0.12 11.34 15.51 0.18 1.21
5. Burr III 0.07 0.12 0.55 10.60 12.59 0.10 0.60
6. Burr III 0.10 0.17 0.54 5.60 11.07 0.35 0.53
a
: the critical value of Anderson-Darling statistic used is 2.5 for all distributions at =0.05.
* : assumption rejected at the probability level =0.05.

4.4.2 Mixture dbh distribution models of larch forest
Multi-cohort features are evident in the plots 1-3 and 5-6. A multimodal approach based
method is more flexible to derive a suitable mixture model that fit such plots having dbh structure
with irregular shape or multiple peaks by integrating adequate single pdf models. Fig. 1
demonstrated the curve of each derived mixture models (real line). The unimodal approach derived
single model (dash line) is also shown in Fig. 4.1 for comparison. The breakpoints assigned to the
multi-cohort plots are: 19 cm for plots 1 and 2, 23 cm for plot 3, 19, 25, and 42 cm for plot 5, and
29 cm for plot 6. Table 4.4 shows the statistics of the tests K-S, A-D and
2
tests for each of the
single pfd model of the plots. The results of goodness of fit tests completely agree that every
derived single model of the sub-group distribution is the same as the tested theoretical distribution
function.

36

Fig. 4.1 The diameter distribution of the study plots in larch forests. The histogram represents the
observed dbh distribution, the dashed line represents the derived simple distribution model and the
solid line represents the derived mixture distribution model.

37

Table 4.4 Measures of the goodness of fit tests for the multimodal approach derived sub-group dbh
distribution models.
Plot
Range
of
dbh
class
Model
number
Model type
Kolmogorov-Smirnov Chi-Square
Anderson-
Darling
Statistic
Critical
value
p value Statistic
Critical
value
p value Statistic
a

1. 2-19 1.1 Burr III 0.08 0.13 0.44 7.94 12.59 0.24 0.70
1. 19-64 1.2 Burr III 0.14 0.36 0.98 0.07 3.84 0.79 0.27
2. 1-19 2.1 Burr XII 0.13 0.17 0.19 6.36 9.49 0.17 1.89
2. 19-40 2.2 Johnson S
B
0.12 0.14 0.13 5.60 12.59 0.47 1.50
3. 2-23 3.1 Burr III 0.07 0.12 0.59 11.60 12.59 0.07 0.49
3. 23-65 3.2 Johnson S
B
0.20 0.45 0.86 - - - 0.29
5. 1-19 5.1 Johnson S
B
0.07 0.18 0.97 0.44 11.07 0.99 0.31
5. 19-25 5.2 Johnson S
B
0.17 0.36 0.79 1.08 3.84 0.30 0.43
5. 25-42 5.3 Weibull 0.10 0.22 0.10 0.92 11.07 0.97 0.27
5. 42-56 5.4 Weibull 0.12 0.28 0.89 0.23 5.99 0.89 0.59
6. 2-29 6.1 Lognormal 0.06 0.19 0.98 1.47 11.07 0.92 0.20
6. 29-77 6.2 Burr XII 0.13 0.36 0.96 0.07 3.84 0.78 0.40
a
: the critical value of Anderson-Darling statistic used is 2.5 for all distributions at =0.05

4.4.3 dbh-h relationship of larch forest
The relationship between tree dbh and tree height of the study sites were drawn by 1)
exponential raise (plots 1 and 2) and sigmoid (plots 3, 4, 5, and 6). Fig. 4.2 demonstrates the
graphical fitness of those dbh-h models. The corresponding statistics of the regression assumption
tests are also shown in Table 4.5. Although the C-V test for plot 2 suggested a non-constant
situation existed (p value <0.01), the other plots actually do satisfy the constant variance rule (p
value >0.05). Even though we tried to use root-squared and log transformation to overcome this
problem, the constant variance still not satisfied.

38

Table 4.5 Measures of the goodness of fit tests for the empirical dbh-h models.
Plot Model R
2
Durbin-Watson Statistic
b
Kolmogorov-Smirnov Constant Variance
p value Statistic p value
1. Exponential 0.93 1.91 0.1558 0.5422 0.6463
2. Exponential 0.61 1.08 0.1232 0.6343 <0.0001**
3. Sigmoid 0.98 2.41 0.0808 0.9807 0.1448
4. Sigmoid 0.93 1.58 0.1641 0.5555 0.0742
5. Sigmoid 0.73 2.22 0.1905 0.5602 0.2284
6. Sigmoid 0.94 1.50 0.1436 0.8926 0.6856
b
: the critical value of D-W statistic is 1.5-2.5
** : assumption rejected at the probability level = 0.01

The result of C-V test indicates that the growth of tree height is not equally responding to the
increment of tree dbh for plot 2. In particular, a bigger variation of tree heights occurred among the
smaller diameter trees in this plot. This might be an effect caused by disturbance or natural variation.
In Fig. 1, plot 2 showed that a disturbance most likely happened at the 18cm dbh class causing a
sharp reduction of stems in the adjacent smaller dbh classes. The main canopy space was probably
removed by the disturbance and thus improved the height growth of smaller trees that occupy the
under storey of forest canopy. In addition, the stand density of plot 2 is 720 stems/ha. A lower
density stand offers a bigger space for tree growth, and particularly the height growth of younger or
smaller trees would not be constrained by the surrounding taller trees. The non-constant variance of
plot 2 is considered reasonable based on the low density and hence their dbh-h model is still
acceptable.
The R
2
of the derived dbh-h models (Table 4.4) for plots 1-6 ranges from 0.61 to 0.98. They
are statistically significant at the 0.05 probability level. This indicated that the prediction ability of
the tree height using the dbh-h models varies dramatically among the forest stands. We therefore
suggest that a forest classification (especially for the site index or site productivity) should be
conducted in advance of the dbh-h modeling in order to upgrade the quality of height prediction
using the dbh of trees and thus to improve the accuracy of tree volume estimation.

39

Fig. 4.2 Curvilinear relationships of tree height to dbh of larch forests. The relationship is
exponential for plots 1-2 and sigmoid for plots 3-6.

40

4.4.4 Accuracy assessment of the volume estimation for the models derived by unimodal
approach and multimodal approach
Table 4.6 summarized the accuracy assessment of the plot volume estimation for all of the six
study sites. The SDMs have plot volume estimates with a deviation from -0.5 to -11.6 m
3
while the
MDMs have deviations from -1.2 to -5.6 m
3
. Under estimation is the common feature of these two
kinds of estimation methods. The percentage of prediction bias using the MDM and the SDM
approaches is between 4.67% and 11.69% and 3.01% and 46.77% respectively. Briefly, the MDM
method has plot volume under estimated with average bias 2.42 1.07m
3
or error percentage 7.08
2.97% which is better than the SDM method whose average bias is 6.03 3.53m
3
or error
percentage 23.52 17.47%.

Table 4.6 A comparison of the accuracy of volume estimation using the simple dbh models and the
mixture dbh models
*
.
Plot mvol
T
(m
3
) evol
T
by
SDM (m
3
)
evol
T
by
MDM (m
3
)
Bias of evol
T

by SDM (m
3
)
Bias of evol
T

by MDM (m
3
)
Error% of
evol
T
by SDM
(%)
Error% of
evol
T
by
MDM (%)
1. 23.7 14.8 21.7 -8.9 -2.0 -37.55 -8.44
2. 38.3 30.2 36.3 -8.1 -2.0 -21.15 -5.22
3. 24.8 13.2 21.9 -11.6 -2.9 -46.77 -11.69
4. 12.4 11.6 n.a. -0.8 n.a. -6.45 n.a.
5. 74.6 72.3 70.6 -2.3 -4.0 -3.08 -5.36
6. 25.7 15.8 24.5 -9.9 -1.2 -38.52 -4.67
Average 33.25(37.42) 26.32 35.00 -6.93 -2.42 -25.59 -7.08
SD 21.86(21.61) 25.01 20.78 3.53 1.07 18.18 2.97
*: n.a. means not available; the number with a sign - indicates under estimation; the values in
parenthesis are calculated without the volume of plot 4.

4.5 Discussion and Conclusion
A mixture dbh model which integrates sub-group distribution functions definitively can draw
a more realistic distribution of dbh structure and is therefore able to make a more accurate
estimation of forest volume.
Among the six plots collected from different larch stands of this study, five of them are better
fitted with a mixture distribution dbh model (one of them was regular shaped), which integrates two
or multi component distribution functions. A largest deviation of plot evol
T
for the simple
distribution models is around 11.6 m
3
or 47% error. This deviation can be reduced down to 2.9 m
3

or 12% by the mixture distribution models. For a general comparison, the average volume estimates

41

of the multi-cohort plots also shows that the average volume prediction bias and error percentage of
simple distribution model is 8.16 and 26% while they can be improved to 2.42 m
3
and 8% by the
mixture distribution models. Since the simple distribution model and the mixture distribution model
based methods used the same dbh-h relationship formula in stand evol
T
, it is concluded that the
proposed multimodal approach could be achieved with a satisfactory evol
T
.
The mixture model is more flexible in describing the diameter structure in particular of those
stands with irregular or atypical even-aged/uneven-aged distribution. It can be applied to enhance
the forest growth and yield model and moreover to explore the stages of forest development. Forest
development is a result of many factors including growth, competitions, disturbances, and global
warming. The mixture model has a high potential for exploring the effects of those factors in forest
succession. In order to obtain a more accurate measure of forest volume, forest surveyors should
note that mixture modeling should be considered if a histogram of diameter structure is different
from typical distributions.

42

Chapter 5
5 Forests aboveground biomass and carbon storage estimation in Eastern Khentii,
Mongolia

5.1 Abstract
Direct relationships exist between forest volume, biomass and carbon stock. Due to the laborious
work and time consumption, it is common to use regression models to predict forest biomass. In this
paper biomass and carbon of pure larch, birch and mixed larch-birch forests in Eastern Khentii,
Mongolia were calculated. Allometric equations (Usoltsev 2002) were used to predict the biomass
components including stem, branch and foliage. Total biomass were later translated into carbon
stock by using elemental composition of wood sample result.

5.2 Introduction
Biomass and carbon monitoring becomes important when economic and political meaning of
biomass and carbon estimation facts are used in international treaties. For instance, forest biomass is
used as a renewable energy source and carbon is an internationally tradable commodity.
Diameter and height are the main key parameters to describe the forest structure. Tree volume,
biomass and carbon are easily related and measured by these variables. Because direct measurement
of forest biomass on the field is destructive and therefore, often, observations of biomass are
normally based on allometric models, which approximate the biomass of the main components of
tree (stem, branch and foliage) and the total biomass, according to measurable variables, such as
diameter and height.
Because forest biomass study requires intensive labor work and time, researchers often use
allometric equations to calculate forest stand biomass partly and/or fully. These equations
determined relationship between forest descriptive parameters (forest age, diameter, height, number
of stems) and main components of tree (stem, foliage, branch, and root) biomass.
The first information of biomass study in forests of Mongolia were noted by Krasnoschekov
(1982), Savin and Dugarjav (1983), Tsogt (1993) and Danilin (1995).
Usoltsev (2001 and 2002) collected more than 5 thousand plots including other researchers
data to build biomass base for Europe and Eurasian forests and developed species-specific and
region-specific mathematical model for 13 species in 50 regions. In that work, Mongolian 8 larch
plots data were included (Krasnoschekov 1982, Savin and Dugarjav 1983, Danilin 1995).

43

Previously, we published (Dorjsuren and Tsogt 2005, Dorjsuren et al. 2007) larch forests
biomass tables by using Usoltsev model in two different provinces. This work has been based on
stand growth table of our result (Dorjsuren et al. 2003 and Tsogt et al. 2006). According to Tsogt
(1993) study, biomass share of tree components were changed as stand age increases in case of 16-
70 years old larch forests in Eastern Khentii, Mongolia. The study shows that tree crown (foliage
and branch) biomass were bigger than stem biomass at age of 16 years. This ratio reduced around
30-40 years and stem biomass reached 70-80% at age of 70.
Objective of this study is to calculate 14 forest-stands biomass by using Usoltsev model and
then to calculate their carbon stock by using the wood sample analysis result.

5.3.1 Study area
Forests of Eastern Khentii sub-province, Khentii province, Southern Transbaikal region in
Mongolia have been selected for this research and all tree genera which grow in Mongolia are in the
forests of this region. The following are included: Siberian larch (Larix sibirica), Scotch pine (Pinus
sylestris), Siberian cedar/pine (Pinus sibirica), Siberian spruce (Picea obovata), Siberian fir (Abies
sibirica), birch (Betula plathyphilla), aspen (Populus tremula) and poplar (Populus laurifolia) etc.
Eastern Khentii covers territory along the basin of the rivers Onen, Kherlen and Tuul. Elevtations of
the research plots range from 1000 m (upper river Onon) up to 2500 m along the watershed areas.
Mean annual air temperature is -3.5 to -3.8
o
C. Monthly average air temperature in J aunuary is -27.4
to -27.9
o
C. In J uly, it is +16.9 to +17.1
o
C. The sum of temperature above +5
o
C is 1876 to 1891
o
C
and above +10
o
C is 1754 to 1595
o
C. Length of vegetation period is 143 to 146 days, mean annual
precipitation is 240 to 264 mm, and mean annual relative air humidity is 62 to 65 %. The lowest
temperature difference occurs during winter season in a day but the highest ones occur during
summer season.
In the study area, forests are natural which generates itself and produced naturally. The
compositions of forests are mainly conifer and broadleaf mixed forests. Shrubs grow on river basins
and mountain valleys. According to the report of National Forest and Water Inventory Center
(2007), the relative occupied areas of the species to the total natural forests area is that larch is
59.2%, pine is 2.6%, cedar is 28.4%, fir is 0.3%, birch is 6.4%, aspen is 0.3%, poplar is 2.4% and
willow is 0.4% in Mongon-morit soum, Tov aimag (Fig 5.1). Most of these areas are dominated by
larch (Larix sibirca Ledeb.) stands with herbaceous ground layers, growing usually in the lower
portions of mountain slopes. Totally, six pure larch, one pure birch (Betula platyphylla), two larch

44

dominated, two birch dominated, one birch-larch mixed and one birch-aspen mixed stands were
inventoried in this study and their general characteristics are given in Table 5.1.
The most widespread forest types among subtaiga larch forest are mesophyte herbaceous,
steppe characterized herbaceous, Rhododendron dahuricum, herbs-Vaccinium vitis-idaea, and
forest-meadow herbaceous larch forests. The most common soil type is mountain derno-taiga deep
permafrost soil (Savin et al. 1988, Tsedendash 1993, Dugarjav 1995, 2006, Dorjsuren 2006).
According to forest law (Forest law 2012), with consideration of forest protection, application
type and ecology-economy interest, the forests of Mongon-morit soum are classified as forests of
protection and application zone (Otgonsuren et al. 2007).

Fig. 5.1 Tree species of study area, Mongon morit soum, Tov aimag, Mongolia.

45

Forest inventory were conducted in total 14 plots of Larix sibirica, Betula platyphylla and
mixed forests in year 2008. The latitude and longitude of each plots were recorded with global
positioning system (GPS) in four corner. Originally, the study plots were established by Russian-
Mongolian Complex Biological Expedition for forest ecosystem structure and other features of the
main forest types of larch forest and their dynamics under the influence of disturbance factors
studies. Selection, notation and description of the sample plots were carried out according to the
common methodological instructions given in the research works of Anuchin (1977) and Sukachev
and Zonn (1961).

Table 5.1 General characteristics of studied plots
Plot
ID
Plot location
Species
composition
a

Average age
of main
canopy, year
Average
diameter,
cm
Average
height, m
Number
of stems,
N/ha
Volume
stock,
m
3
/ha
P1
N48
o
20.571
E108
o
35.331
Birch 10 75 12.03 11.67 1244 95.03
Larch + 70 20.50 16.20 67 25.55
Total 1311 120.58
P2
N48
o
20.567
E108
o
35.278
Larch 8 190 11.64 10.89 2000 447.06
Birch 2 70 11.11 11.78 360 23.69
Total 2360 470.76
P3
N48
o
20.277
E108
o
35.125
Larch 10 60 19.80 13.98 720 183.42
P4
N48
o
20.295
E108
o
26.324
Birch 8 25 1.81 3.19 2967 2.06
Willow <1 25 1.91 2.75 175 0.08
Larch + 80 35.75 18.09 33 22.10
Total 3175 24.24
P5
N48
o
20.317
E108
o
26.453
Larch 6 60 11.84 9.68 840 208.55
Birch 4 55 11.35 11.74 500 32.27
Total 1340 240.81
P6
N48
o
20.264
E108
o
26.439
Larch 8 80 11.64 11.14 790 75.58
Larch + 300 63.53 26.87 40 131.41
Birch 2 70 13.70 13.10 190 21.82
Total 1020 228.81
P7
N48
o
20.260
E108
o
26.214
Birch 10 12 1.46 3.49 1492 0.91
P8
N48
o
16.676
E108
o
27.046
Birch 6 12 1.15 2.14 8900 1.36
Aspen 4 12 1.24 2.24 7400 1.80
Total 16300 3.16

46

Plot
ID
Plot location
Species
composition
a

Average age
of main
canopy, year
Average
diameter,
cm
Average
height, m
Number
of stems,
N/ha
Volume
stock,
m
3
/ha
P9
N48
o
11.900
E108
o
27.046
Birch 6 50 10.70 12.33 1424 83.98
Larch 4 60 3.30 4.16 1056 4.02
Total 2480 88.00
P10
N48
o
26.160
E108
o
50.880
Larch 10 190 30.30 20.00 315 209.10
P11
N48
o
01.861
E108
o
46.067
Larch 10 60 5.39 6.35 9200 245.97
P12
N48
o
02.171
E108
o
46.335
Larch 10 300 23.89 17.14 600 378.05
P13
N48
o
01.870
E108
o
46.099
Larch 10 70 9.02 9.16 363 25.47
P14
N48
o
01.868
E108
o
44.526
Larch 10 90 16.78 14.33 1525 646.17
a
Species mixture of stand composition is expressed by total 10 score, which are ratio between
number of stems of an observed species and total stems in plot area. For instance, two species
percentages were 70 and 30, it will count as 7 and 3, relatively. Pure forest expressed as 10.

5.3.3 Biomass calculation
The general models of Usoltsevs (2002) stem biomass Eq. 5.1, and other part of tree
components biomass Eq. 5.2 are given below. The equations coefficients are species-specific as
well as region-specific.

ln(Ps)=f(lnA, lnD, lnH, lnN, lnM) Eq. 5.1
ln(Pi/M)=f(lnA, lnD, lnH, lnN) Eq. 5.2
where Ps stem biomass, P
i
a main component of tree biomass (P
F
foliage, P
B
- branch) t/ha;
A forest age, D average tree diameter, H average tree height, N - number of stems, N/ha; M
forest stock, m
3
/ha

The biomass of tree main components of species coefficients, determination coefficient (R
2
)
and standard error (SE) given in Appx 2.

47

5.3.4 Carbon and other elemental composition estimation
Carbon and other main elemental compositions (oxygen, hydrogen and nitrogen) were
measured by laboratory analysis, primarily to get an estimation of carbon content in dry matter of
tree species (Table 5.2). Stem discs and branches taken for wood samples and analyzed at Forest
Soil and Tree Mycorrhiza laboratory, Department of Forestry, National Chung Hsiung University in
Taiwan. Those selected tree species are the main tree species in Mongolia including Populus
tremula and P.laurifolia (Dugarjav, 2006). Biomass to carbon conversion was made based on the
percentage of the elemental composition by using simple linear equation. For carbon calculation of
broadleved tree species Populus laurifolia and Salix were translated as Betula platyphylla carbon
ratio.

Table 5.2 Four main elemental compositions of larch and birch tree species in Mongolia
Species C% O% H% N%
Larix sibirica 46.64 46.85 6.49 0.02
Pinus sibirica 47.45 45.80 6.71 0.04
Pinus sylvestris 48.25 44.84 6.86 0.05
Picea obovata 46.93 46.42 6.61 0.04
Picea abies 47.32 46.14 6.52 0.03
Betula platyphylla 46.56 46.97 6.44 0.04

5.4 Results and Discussion
Forest descriptive parameters (Table 5.1) used as inputs to the model and calculated 14
sample plots biomass and carbon (Table 5.3). Species biomass accumulation characteristics are
different depending on age and morphology of each tree species. Stem has the largest share in total
stand biomass while foliage and branch biomass are the smallest share in stands regardless of
species. Around 50-75 years old birch forest stands foliage biomass was higher than branch
biomass, P1 and P9 while larch trees branch biomass was higher than foliage biomass, because of
our age range of larch stands is older and it is 60-300 years old. The species morphology differences
result in biomass accumulation differences in parts of tree components between larch and birch
trees. This can be seen by comparing foliage and branch biomass of birch and larch species, higher
and lower, relatively.

Table 5.3 Aboveground biomass and carbon of larch, birch and mixed forest stands in Eastern
Khentii, Mongolia (AGC Aboveground carbon)

48

Plot ID
Species
composition
Aboveground biomass, t/ha Total AGC
t/ha Stem Foliage Branch Total
P1
Birch 50.887 14.922 10.723 76.533 35.634
Larch 12.883 0.229 1.403 14.514 6.769
Total 63.770 15.151 12.126 91.047 42.403
P2
Larch 213.200 3.696 18.333 235.229 109.711
Birch 13.068 3.641 3.176 19.886 9.259
Total 226.269 7.337 21.510 255.115 118.970
P3 Larch 88.960 1.303 9.130 99.392 46.356
P4
Birch 1.119 0.223 0.475 1.817 0.846
Willow 0.045 0.016 0.501 0.562 0.262
Larch 11.253 0.124 1.323 12.700 5.923
Total 12.417 0.363 2.299 15.079 7.301
P5
Larch 100.892 2.374 11.854 115.120 53.692
Birch 17.684 4.997 4.133 26.814 12.485
Total 118.576 7.371 15.987 141.934 66.177
P6
Larch 36.733 0.833 3.608 41.173 19.203
Larch 66.316 0.319 6.473 73.109 34.098
Birch 12.158 3.532 3.051 18.741 8.726
Total 115.207 4.684 13.132 133.023 62.027
P7 Birch 0.502 0.118 0.274 0.893 0.416
P8
Birch 0.734 0.165 0.431 1.330 0.619
Aspen 2.140 0.579 9.513 12.232 5.695
Total 2.874 0.744 9.945 13.562 6.314
P9
Birch 44.892 12.874 8.831 66.597 31.008
Larch 1.969 0.151 0.307 2.427 1.132
Total 46.861 13.026 9.138 69.024 32.140
P10 Larch 102.487 0.873 8.287 111.647 52.072
P11 Larch 115.198 4.407 12.432 132.037 61.582
P12 Larch 183.310 1.669 15.671 200.650 93.583
P13 Larch 12.557 0.405 1.476 14.439 6.734
P14 Larch 308.845 4.517 25.545 338.908 158.067

Often, studies concentrate on the commercially valuable species rather than all tree species
biomass. It is still important for broadleaved tree species biomass to be calculated. Therefore, we
have to build more sample plots in different species such as, age and density forest stands to check

49

precision of Usoltsevs coefficients of regression equations and adjustment or to find accurate
coefficients.
Generally, biomass estimates multiplied by 0.5, a standard factor for converting woody
biomass to carbon (Garzuglia and Saket 2003, Tsolmon 2003). The overall result of the wood
chemical analysis is that the content of any component of forest woody biomass is on average about
50% carbon, 6% hydrogen, 44% oxygen, less than 1% nitrogen and trace amounts of several
elements, which justifies further application of these values (Table 5.2). The results of carbon and
other main element proportion, according to the relative position of samples in the tree, were not
stated. Neither was any conclusions drawn in that sense predominantly due to a relatively small
number of sampled trees. Although, the chemical analysis study did not perform any statistical
measure, the study shows the potential approach to estimate carbon and other main elemental
compositions.

5.5 Conclusion
Calculations of 14 plots biomass and carbon estimation results have been presented in this study.
Relationships between forest volume stock, biomass and carbon give an insight in the possible
application of remote sensing in biomass and carbon monitoring and forest management projects for
spatially vast areas. Furthermore, the biomass and carbon calculation result could be used as input
to image and biomass/carbon study.

50

Part I. Summary
Volume stock of a forest stand is generally predicted by an area approach that involves
dividing the plot volume by plot area to get the volume per hectare (m
3
/ha) and then this is
multiplied by the area of a forest stand. Sampling technique, sample representative, and plot size are
the general factors that might influence the accuracy of volume estimation. In the design of forest
inventory sampling methods, the method used to derive plot volume becomes a key factor
controlling the quality of forest volume/biomass/carbon storage estimates.
dbh distribution modeling (stand structure based method) offers a better way of deriving the
dbh structure of forest stand. If the number of stems in each dbh class that are predicted by a
distribution model are closer to the real value, then the volume estimation will be more accurate. In
the past, a unimodal based distribution function was usually applied for dbh structure modeling. The
derived simple distribution model has potentially violated the basic assumptions of the model or
deviated far from the real situation of stems frequency along the dbh classes. As a result, a poor
estimation of forest volume occurred. This is obviously evident in some of the larch forest shown in
Chapter 4 study.
Forest is complex population system containing all sizes of tree stems. Therefore, in forestry
we have used many different distribution functions to represent their structure. The choice of
distribution function is also dependent upon the purpose of study. In general cases, simple
distribution functions should be better for a regional study and if forest structure is irregular,
mixture distributions should be used for smaller spatial area to express forest structural distribution.
It is also obvious that for specific management plan or ecological studies, simple distribution
functions are not good enough especially if the forests are irregular shaped. Even if the forest
diameter distribution structure is regular, there are still variations within trees due to their genetic
inheritance and environmental factors. Therefore, researchers need to take more care while
choosing appropriate distribution model.
Distributions are different from each other, covering different regions of skewness and
kurtosis area which is their display of unique characteristics. So, foresters are better to see the two
tails as well as the peak of the fitting distribution. One can choose the best available distribution
among suggested distributions depend on the study goal. For instance, if regeneration status is
important task then left tail, if merchantable trees are important then right tail and if improving
forest stock is important task, then mid diameter classes should be taken more care while drawing
forest diameter distribution.

51

Spatial scale is also one of important factors to define forest trees size distribution structure
which is rarely considered in this study. For this kind of study, sample plots have to be chosen
randomly according to equal probability. But, the study plots are not random plots; they have been
selected for long term stand dynamic analysis, therefore they are carefully chosen for uniformity
before this study. However, plots do satisfy stand characterization which is able to explain pattern
of forest species vertical and horizontal distributions, size of tree parts, tree age, and combinations,
respectively to surrounding forest area. For the consideration of our study, we sampled forest stand
by plot trees that is medium of certain range of forest area. This is the opposite of random sampling,
but the goal is the same to find the forest stand that can express general structure of total forest area.
In another way, random sampling needs more plots/area to cover. Directly choosing medium forest
stand helps to reduce labor work and time. Moreover, these plots can be used as surrogates of pure
pixels in remote sensing because of their general characteristics.
To model simple distribution for broad range of forest area, the left tail of distribution
constrained only by 0, because, tree height which is often shorter than breast height is excluded
from inventory. Right tail of the distribution should not be constrained if one does not know the size
of biggest tree. There is always the highest peak on the frequency distribution and if the distribution
shape is irregular and the heights of peaks are exactly the same then they can be thought as one.
To model mixture distribution, there are two possible ways. First method is based on forest trees
statistical characteristics. Second way is based on exclusive parameters (dominance/health status,
Kraft class, and bonitet class so on) which are collected during forest inventory. In our study case,
mixture distribution modeling is based on stand trees statistical characteristics. Advantage of the
mixture distribution is accurate at small scale forest area but could be easily biased if the stem
numbers and the spatial area slightly changed.

52

Part II. Aboveground biomass estimation using Landsat TM imagery in Mongolia

Part II. Background
Mongolian forests distribute in transition zone and grow on Northern part of mountainous
area, the forests hugely influenced by Asian dry steppe, and is in extreme condition, therefore
Southern boundary of Siberian forest distribution limits there. But, the climate condition shapes the
forests structures and distributions and distinguishes them from Siberian forests. The forests are
easily influenced and regressed by forest fire, insect outbreak and anthropogenic negative activity.
Direct measurement of biomass on the ground is time consuming (expensive), and repeated
measurements, if they occur at all, are generally limited to ten year intervals according to what is
stated in Mongolian forest law where forest inventory revisit time frame is eighteen years.
Therefore, it is certainly important to use remote sensing techniques to update forests information
including forest distribution, composition, structure and biomass.
Wide range of approaches has been developed for quantifying biomass and leaf area index
(LAI) using active and passive remote sensing systems. Forest biomass study is well developed yet
hard to quantify and asked the well prepared ground data. By the different remote sensors, bring
different results would be found, depending on research area, sensor capability and good ground
truth. Several earth observation satellites (Landsat, MODIS and SPOT etc.) give us potential ways
to monitor vast forest areas. The Landsat TM is the most ideal satellite used for forest research
because of its relatively wide spatial path which is 185 km, temporal resolution which is 16 days
and 7 wavelength channels in the visible, near-infrared (NIR) and showtwave-infrared SWIR
portions of the spectrum.
In chapter 6, we developed regression models for AGB and its surrogate of ground measured
LAI by using their relationship with satellite image based vegetation indices (VI). Quantification of
AGB and LAI on regional scale, accurate land cover and forest type map needed. For that, in
chapter 7 we studied land cover and forest type mapping. Because separability of spectral signatures
within classes increased in seasonal composite images due to plant phenology changes, different
seasonal composites, multi-temporal imageries bring higher classification accuracy than single-data
imagery and satellite image data has been widely and successfully used to map land cover and
forest type. In addition, ancillary biophysical information can improve classes separation according
to a detailed level of forest distributional characteristics. In the case of our study, forest distribution
characteristics of altitudinal-belt zone information were applied in province level and only dominant
tree species were concerned.

53

Chapter 6
6 Regression models for forest aboveground biomass and leaf area index in North-Eastern
Mongolia

6.1 Abstract
Due to relatively optimal spatial and temporal characteristics Landsat TM data have been
dominantly used in forest researches. In this study, we developed aboveground biomass (AGB) and
leaf area index (LAI) models using vegetation indices (VIs). Beforehand, single, double, triple band
ratios calculated to find correlation with AGB and LAI. Triple band combination, green, red and
shortwave-infrared (SWIR) was showing the best correlation with AGB; while band combination,
green red and near-infrared (NIR) was showing the best correlation with LAI. For the AGB a new
spectral index is presented integrating normalized difference green vegetation index (NDGVI) with
SWIR optimization (NDGVIc) method and for LAI Brightness and surface albedo tested (SA). The
relationship between NDGVIc and AGB is defined at R
2
=0.71; while Brightness of tasseled cap
transformation and (SA) were showed promising result to predict LAI in R
2
=0.83 and 0.81,
relatively. These VIs were using combination of all available bands given from Landsat sensor from
blue to SWIR region. That why they were successful. Due to limited number of stands regression
model was not available to be assessed. However, the relationships between VIs and AGB and/or
LAI were tested on other two images which are acquired in same season. The results were
optimistic; therefore the criteria of choosing appropriate VIs were developed among tested VIs with
consideration of different band correlation circumstances in this study.

6.2 Introduction
6.2.1 Background
Since vegetation indices (VIs) are proportional to the value of biophysical parameters such as
LAI, green vegetation fraction, net primary productivity, and fraction of absorbed
photosynthetically active radiation, it is commonly used to indicate vegetation dynamic and amount.
VIs intended to enhance the vegetation signal, while minimizing solar irradiance and soil
background effects. Each of the VIs is designed to accentuate a particular vegetation property.
Many are functionally equivalent in information content (Tucker 1979a, and 1979b, Perry and
Lautenschlager 1984) and some provide unique biophysical information (Qi et al. 1995). The
simplest form of VI is a ratio between two digital values from separate spectral bands.
Band rationing is divides the pixels in one band by the corresponding pixels in a second band.
The reason for this is twofold: one is that differences between the spectral reflectance curves of

54

surface types can be brought out (Satterwhite 1984), and the second is that illumination, and
radiance, may vary, then the ratio between an illuminated and a not illuminated area of the same
surface type will be the same. Thus, this will aid image interpretation, particularly the near-
infrared/red (NIR/R) band ratio (Birth and McVey 1968). The ratio is high for healthy vegetation,
but is lower for stressed or yellowed vegetation (lower near infrared and higher red values) and for
nonvegetated areas. In addition, the lower the correlation is between the bands, the greater the
information content of the band-ratioed image could present (J ensen 2005). Usually a more
reflective band is used as the numerator of the ratio, so that the occurrence of the target material
yield higher ratio values (greater than 1.0) and appears bright in the ratio image. Although such
ratios have been shown to be powerful tools for studying vegetation, they must be used with care if
the values are going to be rigorously (rather than qualitatively) interpreted. The IR/R ratio is one of
many measures of vegetation vigor and abundance. The green/red (G/R) ratio is based upon the
same concepts used for the IR/R ratio, but it is considered less effective. In principle, this index
conveys same kind of information as the IR/R and G/R ratios.
VIs have been compared to field measured AGB and LAI. AGB is represented by the weight
of vegetative tissue and LAI is represented by the area of leaf surface per unit area of soil surface in
image data. No single VI seems to be equally effective for AGB and LAI, as well as for all plants
and ecosystem conditions.
Leaf optical variability has the largest effect on canopy reflectance in the NIR (Asner 1998).
Therefore, most VIs combine information contained in two spectral bands, the red and NIR. One of
the most widely used VIs are known as the simple ratio (SR) and normalized difference vegetation
index (NDVI). NDVI is normalized verion of SR, therefore it is considered better than SR in bigger
area. But, NDVI is less sensitive to chlorophyll content, biomass and LAI (Buschmann and Nagel
1993). NIR reflectance decrease as a result of a change in leaf orientation, from predominantly
horizontal to predominantly vertical, at a certain stage in the growth cycle (Colwell, 1974, J ackson
and Ezra 1985, Suits 1972). At a certain phenological stage of deciduous tree development, , the
leaves of trees change color and to turn yellow or reddish. The result is a decrease in the number of
actively reflecting leaf layers, with a consequent decrease in NIR reflectance (Leamer et al. 1980).
Thus, Gitelson et al. (2002) designed VARIgreen and proposed to use only visible range of the
spectrum to estimate AGB.
Changes in the amount of leaf water can be detected by in SWIR reflectance (Gausman et al.
1970 and 1978, Thomas et al. 1971, Tucker 1980). Crist and Kauth (1986) derived the visible, NIR
and SWIR coefficients for transforming Landsat TM imagery into brightness (B), greenness and
wetness. It is called cap transformation because of its cap shape. Chen et al. (2002) modified the SR

55

(RSR) and Nemani et al. (1993) improved NDVI by using SWIR reflectance as a correction
(NDVIc) factor. The surface albedo (SA) is extrapolated by Richter (2000a). These works include
water related information through the inclusion of SWIR band. A leaf relative water content study
(Hunt et al. 1987) show that SWIR-2 (band-7 in case of Landsat TM) had a larger percent change
than SWIR-1 (band-5). The absolute change was small so that SWIR band-5 is preferable for
determining the absorptance by water (Tucker 1980, Rock et al. 1985).
It is important to understand how the reflectance is related to various leaf physiological
properties to use vegetation indices.

6.2.2 Spectral characteristics of vegetation
The spectral characteristics of vegetation depend on many factors such as causing absorption,
transmission, and reflection of incoming solar radiation. In the visible region of the spectrum (0.4
0.7 m), most of the radiation is used for photosynthesis and is absorbed by foliar pigments, like
different types of chlorophyll, carotene and xanthophyll (Waring et al. 1995) (Fig 6.1). Much of the
near-infrared (0.7 1.3 m) energy is reflected by foliage (Knipling 1970, Gausman 1977). Near-
infrared reflectance is determined by differences in internal leaf structure of plants (Gausman et al.
1969) and absorption features of other biochemical contents, like lignin and cellulose (Peterson
1989). Leaf optical variability has the largest effect on canopy reflectance in the NIR (Asner 1998).
In a healthy green canopy, the near-infrared reflectance increases dramatically (Fig 6.2) and leaf
transmittance characteristics depend on spongy mesophyll cells. The reflectance of midinfrared
region (1.3 2.5 m) is generally higher and within this region, strong water absorptions occur.
Spectral reflectance and characteristics are different, depending on species (Cochrane 2000,
Schmidt and Skidmore 2001 and 2003). Furthermore, the amount of bare soil and dead organic
material also influence the spectral signature. Phenologic changes in plant species and in the
vegetation during the growing season do affect the spectral properties of vegetation types. Tsogt et
al. (2008) present absorption changes of blue spectra around a year in case of Red cypress
(Chamaecyparis formosensis), while red light absorption was not changed. The blue spectral
absorption was high in summer and low in winter which follows the seasonal pattern Fig 6.3.

56

2700 1940 1450 1190 970 660 440
0.5
0.4
0.3
0.2
0.1
0.0
Wavelength (nm)
R
e
f
l
e
c
t
a
n
c
e
|leaf pigments -------------------water content-------------------| |---cell structure---|

Fig. 6.1 Spectral reflectance of healthy Red cypress (Chamaecyparis formosensis), green vegetation
for the wavelength interval 350-2500 nm. The dominant factors controlling leaf reflectance are the
various leaf pigments in the palisade mesophyll (e.g., chlorophyll a and b, and -carotene), the
scattering of near-infrared energy in the spongy mesophyll, and the amount of water in the plant.
The chlorophyll absorption regions are at 430-450 nm and 650-660 nm and the water absorption
regions occur at 970, 1190, 1450, and 1940 nm.

Fig. 6.2 Additive reflectance of Cinnamomum camphora leaves. 1, 3 and 9 layers. Near-infrared
reflectance increases as the leaf layers in the canopy increases. Direct relationship exists between
response in the near-infrared region and biomass measurements.

57

Fig. 6.3 Yearly pattern of Red cypress (Chamaecyparis formosensis) trees absorption change in the
blue spectral region, 400-550 nm (Tsogt et al. 2008).

6.2.3 Objective of the research
There is an increased interest in estimating the AGB and LAI of forests in short period of time
with high accuracy via using remote sensing in Mongolia where national forest inventory data re-
updated in more than ten year time period. If we want to use satellite data to map the AGB and LAI
of the mountain forests, it is necessary to understand how these variables relate to the satellite
observed reflectance and/or VIs. The objective of the study is to develop forest AGB as well as LAI
models, based on relationship of ground collected data points and VIs of remote sensing imagery in
East-Northern Mongolia (Fig 6.4)
20
30
40
50
60
1 51 101 151 201 251 301 351
Days of the year
A
b
s
o
r
p
t
i
o
n

a
r
e
a

58

Fig. 6.4 Framework of aboveground biomass (AGB) and leaf area index (LAI) study relationship
with vegetation indices (VIs).

6.3.1 Study area and field measurement
The descriptions of study area and field measurement are given in heading 5.3.1 and 5.3.2,
Chapter 5.

6.3.2 Abovegound biomass estimation
Traditionally, stand biomass estimates are derived by conversion of stem volume estimates
from the forest inventory database (Aldred and Alemdag 1988) In the world, biomass estimates also
developed through forest covertype volume tables (Brown and Lugo 1984). In Mongolian case,
Dorjsuren and Tsogt (2005) and Dorjsuren et al. (2007) published larch forests biomass tables for
two different regions by using volume tables. The estimate begins with a tree stem analysis by
species and site type. The appropriate local allometric equations are developed to partition the
estimate into foliar, branch, stem, and root biomass estimates, or perhaps into two components:
aboveground and belowground woody biomass components (e.g., Tsogt 1993, Danilin 1995).
Usoltsev (2002) developed species-specific and region-specific general biomass model for 13 main

59

species in 50 regions, including Mongolian forests. By using Usoltsev model, we measured AGB of
14 forest stands used in this study. The AGB study result is given in chapter 5.

6.3.3 LAI-2000 Plant canopy analyzer (Li-COR)
LAI index was measured by using LAI-2000 pland canopy analyzer (PCA). Determination of
the LAI with LAI-2000 PCA is based on measurements of light transmittance for five zenith angles
assuming an exponential model of light extinction (further information can be found in the
manual Li-COR Inc 1992). However, due to the influence of stems and branches in the image
detected by the sensor the LAI-2000 PCA measurement is rather a Plant area index (PAI) than LAI.
The exponential model of light extinction is based on the assumptions that (i) the leaves are black
i.e. they do not absorb nor reflect radiation, (ii) the leaves in the canopy are randomly distributed,
(iii) the leaves are relatively small compared to the area of view of each ring and (iv) the azimuths
of the leaves are randomly distributed (Li-COR Inc 1992). When using the LAI-2000 PCA, certain
measuring conditions must be fulfilled. The most important of these is the sky condition for which
uniform overcast conditions are preferred. Making measurements at bright sunlight should be
avoided, as this gives an underestimation of the LAI. The incident sunlight onto the leaves is being
scattered in all directions, consequently the sensor does not see the sunlit leaves and the LAI is
underestimated.
A 45
o
view of cap was used in order to block human induced shadow and the LAI-2000 PCA
sensor was always directed towards to the center of plots. The optical sensor was leveled
horizontally and held at breast height. As only one instrument was available, the free field
measurements were conducted in a nearby clearing and immediately followed by the measurements
in the forest. Attention was paid to the sky conditions, measurements were only carried out under
uniform overcast sky.
LAI was determined during the active vegetation growth period of J uly, 2009. We applied the
correction factor for larch stands (1.48) provided by Gower and Norman (1991) since we do not
perform direct LAI measurement.

6.3.4 Landsat TM data
Landsat TM satellite data was acquired on September 07, 2009, orthorectified and the image
georeferenced to a UTM Zone 49 projection based on the WGS 84 datum. Because, cloud free
images which are available on VI study were during fall season. Two other Landsat TM fall
imageries were used to verify the relationships found in this study. They are September 15, 2006
and September 02, 2007 images. An atmospheric correction procedure was done by ATCOR3

60

algorithm (Richter 2000b) to all 7 bands of Landsat TM. The ATCOR3 model calculates the ground
reflectance in rugged terrain with the Lambertian assumption, i.e. assuming an isotropic reflectance
law. SA is derived from a value adding product of the ATCOR3 module.

6.3.5 Vegetation indices
Most of available VIs are applied in this study (Appx 3) to find relationship between AGB
and or LAI of study plots. Some specific indices which are better correlated with AGB and LAI are
given in Table 6.1. They are: brightness (B), reduced simple ratio (RSR), SWIR corrected
normalized difference vegetation index (NDVIc), visible atmospherically resistant index
(VARIgreen) and surface albedo (SA). Landsat TM bands 1, 2, 3, 4, 5, and 7 were used as blue,
green, red, NIR, SWIR1, and SWIR2 band in the VIs, respectively.

Table 6.1 Vegetation indices (VIs) used in this study
VI Equation Reference
B 0.2909IH1+0.2493IH2+0.4806IH3+0.5568IH4
+0.4438IH5+0.1706IH7
Crist and Kauth
(1986)
RSR
NIR
RcJ
_1
SwIR SwIR
mn
SwIR
mux
SwIR
mn
]
Chen et al.
(2002)
NDVIc NIR RcJ
NIR +RcJ
_1
SwIR SwIR
mn
SwIR
mux
SwIR
mn
]
Nemani et al.
(1993)
VARI 0rccn RcJ
0rccn +RcJ Bluc

Gitelson et al.
(2002)
SA

p(x)dx
2.Sm
0.3m
dx
2.Sm
0.3m
; - reflectance value and d - proportion of diffuse
illumination (please see for more details Richter (2000b))
Richter (2000b)

The ratioed bands which have significant correlations with AGB and LAI are further tested
with VIs which are composed of the same band combinations. For a band combination G/R/NIR
B and SA, and for band combinations G/R/SWIR1 or G/R/SWIR2 - SWIR corrected normalized
difference green vegetation index (NDGVIc) (Eq 6.1).

N0IIc =
Luccn-Rcd
uccn+Rcd+L
[1
SwIR-SwIR
min
SwIR
mcx
-SwIR
min
Eq. 6.1

61

where L is an adjustment factor that accounts for differential green and red extinction through the
canopy, SWIR
min
and SWIR
max
are shortwave infrared reflectance from the study plots.
In a senescent canopy, green extinction through a canopy will exceed that in the red-light
resulting more energy reflecting off the canopy background relative to the green. Specially, green
part of the light spectrum become less and other compounds which are chemicals called
anthocyanins (reds) and carotenoids (yellows) become the dominant pigments in the leaves in fall
season. This spectrally dependent, secondary signal is not amenable to simple ratioing and a
background correction becomes necessary. The L helps to keep the index above 0 and magnifies the
index range. Therefore, its use can only be justified if either it improves the regression fit or it
normalizes the regression errors (Draper and Smith 1966). While the SWIR correction acts as a
scalar for the canopy water content, and varies from zero to one depending on the amount of water
which directly related to the AGB of stand trees. Nemani et al (1993) suggest that we could not use
observed max/min values from the scene, as they tend to occur over water bodies and exposed rock
or roads respectively. Consequently, one must be cautious about the choice of max/min SWIR
values. If we choose the max/min values among the plots of observed biomass estimation, there
would have problem to estimate biomass of lower or higher reflectance values beyond the known
range. Therefore, 1000 random points were produced on forest areas to get the max/min values in
order to avoid this problem.

6.3.6 Data analysis
Pearsons correlation analysis used to check interrelationships between two sets of variables.
In this study case, forest variables which are AGB and LAI, and multispectral Landsat TM bands
investigated. Pearsons correlation measures the strength of the linear relationship between the X
and Y variables on a probability plot. The correlation will range between 0 and 1, with higher
values indicating a better fitting distribution. p-values for the hypothesis test of the correlation
coefficient being zero. If p-values are smaller than 0.05, there is sufficient evidence at alpha level
0.05 that the correlations are not zero.
Due to spatial coverage of Landsat TM imagery 6 of 20 plots which are on the edge of
imagery were not available for modeling AGB and LAI of forests, therefore, only 14 plots AGB and
LAI are used for modeling (Table 6.2). The modeling was employed to examine the linear and non-
linear relationships between AGB and/or LAI and spectral features, band and VIs using correlation
analysis, and linear and non-linear by using regression analysis.
The data transformations, logarithm and exponential were used this study. Data
transformations enhance the linear correlation of nonlinear relationship, and therefore an

62

appropriate transformation functions would improve a nonlinear correlation. After testing linear or
nonlinear functions on the transformed data, if the result does not improve the relationship, and
violate the test statistics, original data values will be used.

Table 6.2 Biomass of tree components and LAI of study plots (0.09 ha, regards to Landsat TM pixel
size 30x30 m)
Plot species LAI Volume
Aboveground biomass
Stem Foliage Branch Total
1 birch 2.26 10.85 5.74 1.36 1.09 8.19
2 larch 2.24 42.37 20.36 0.66 1.94 22.96
3 larch 1.96 16.51 8.01 0.12 0.82 8.95
4 birch 2.00 2.18 1.12 0.03 0.21 1.36
5 larch 2.03 21.67 10.67 0.66 1.44 12.77
6 larch 1.96 20.59 10.37 0.42 1.18 11.97
7 birch 1.92 0.08 0.05 0.01 0.02 0.08
8 birch 1.76 0.28 0.26 0.07 0.90 1.22
9 birch 2.02 7.92 4.22 1.17 0.82 6.21
10 larch 2.10 18.82 9.22 0.08 0.75 10.05
11 larch 2.27 22.14 10.37 0.40 1.12 11.88
12 larch 1.89 34.02 16.50 0.15 1.41 18.06
13 larch 2.20 2.29 1.13 0.04 0.13 1.30
14 larch 2.67 58.16 27.80 0.41 2.30 30.50

Take those few data plots used in this study into consideration, additional research should
however be taken to verify the significance of the relationship model of reflectance-biomass.
However the relationship found in this study could be tested by using other imageries acquired in
another dates. Therefore, we used two other additional images to check whether there is the same
kind of relationships that exist.
AGB estimation on image is determined by relationship between AGB and VIs in
corresponding locations (Eq. 6.2). This is the same in case of LAI and VI.

AGB or LAI =f(VI) Eq. 6.2

63

If the AGB and/or LAI estimation had done in different time than the image acquired date, the
time factor (year) should have considered (added) to the relationship model between AGB and/or
LAI and VIs (Eq 6.3).

AGB or LAI =f(VI, time) Eq. 6.3

6.4 Results
6.4.1 Correlations between biomass of tree components and/or LAI and Landsat TM bands
Some spectral bands were spectrally correlated with biomass of tree components and/or LAI
(Table 6.3). Band to band correlation analysis result is given in Appx 4. The correlation patterns
were similar for stem biomass and AGB whilst foliage biomass and branch biomass were not.
However, the correlations were somewhat stronger with LAI than with stem biomass and AGB.
Band-5, corresponding to the SWIR1, was the band showing the strongest correlation with AGB
(r=-0.650). And red band-3 which is the band showing the strongest correlation with LAI (r=-0.808).
Also bands 3 and 7, corresponding to the red and SWIR2 reflectance, were correlated with stem
biomass and AGB; TM Bands 2, 4 and 5, corresponding to the green, NIR and SWIR1 reflectance,
were correlated with LAI. No correlation is found in foliage biomass and branch biomass in the
range of given spectral bands. Therefore, further study will not perform on foliage and branch
biomasses.
The AGB and the LAI are plotted against the correlated reflectance bands in Fig 6.5. The
relationship between bands and AGB and/or LAI is curvi-linear. The scatterplots of the spectral
bands versus stem biomass were very similar to those shown in Fig 6.4 AGB (Appx 5). They could
be surrogate of each other, due to strong and positive correlation between stem biomass and AGB
(r=0.998, Appx 6). Hence, stem biomass is not investigated and only total AGB is studied.

64

Table 6.3 Correlations of the biomass components of forest stand and leaf area index (LAI) against
the Landsat TM band reflectance
Band
Abovegroung biomass
LAI
Stem Foliage Branch Total
Band 1
Band 2
Band 3
Band 4
Band 5
Band 7
-0.450
-0.436
-0.649*
-0.401
-0.648*
-0.549*
-0.019
0.061
-0.134
0.147
-0.344
-0.311
-0.262
-0.286
-0.487
-0.307
-0.491
-0.405
-0.435
-0.402
-0.641*
-0.385
-0.650*
-0.552*
-0.292
-0.731*
-0.808*
-0.601*
-0.610*
-0.531
* Assumption accepted at the probability level alpha 0.05

Fig. 6.5 Relationship between Landsat TM bands 3, 5 and 7 and aboveground biomass (AGB) (left),
and relationship between Landsat TM bands 2, 3, 4 and 5 and leaf area index (LAI) (right).
Correlation coefficients (r) are given for bands and AGB and/or LAI.

6.4.2 Correlation between AGB and/or LAI, and band ratios
Band ratioing was applied to investigate all possible double and triple band combinations into
single indices (Appx 7). According to the result of Pearsons correlation, 4 band ratios given in
Table 6.4 were significantly correlated (p<0.05) with AGB, and green/red/SWIR1 which is the band
ratio showing the strongest correlation 0.797; while 6 band ratios were significantly correlated with
LAI and band combinations green/red/NIR shows the best correlation 0.837.
The correlation analysis showes that some band ratios have a stronger correlation with AGB
than single bands; while the correlation of LAI and band ratios were not much different from some
single bands which are those green and red bands.
-5
0
5
10
15
20
25
30
35
0 5 10 15 20 25
A
b
o
v
e
g
r
o
u
n
d

b
i
o
m
a
s
s

(
t
/
h
a
)
Reflectance (%)
Band 3, r =-0.641 Band 5, r =-0.650 Band 7, r =-0.552
1.5
2
2.5
3
3.5
4
4.5
0 5 10 15 20 25
L
e
a
f

a
r
e
a

i
n
d
e
x

(
m
2
/
m
2
)
Reflectance (%)
Band 2, r =-0.731 Band 3, r =-0.808
Band 4, r =-0.601 Band 5, r =-0.610

65

band ratios.
Band ratio
Aboveground biomass Leaf area index
Pearsons correlation P value Pearsons correlation P value
Band 2/Band 3 0.774 0.001 0.684 0.007
Band 3/Band 5 -0.344 0.228 -0.741 0.002
Band 3/Band 7 0.287 0.320 -0.679 0.008
Band 2/Band 3/Band 4 0.756 0.002 0.837 0.000
Band 2/Band 3/Band 5 0.797 0.001 0.703 0.005
Band 2/Band 3/Band 7 0.709 0.005 0.643 0.013

6.4.3 Correlation between AGB and/or LAI, and VIs
Variety of VIs used to predict AGB and LAI (Appx 8). VIs which have same band
combinations with those ratioed bands which have significant correlations showed better result
according to correlation analysis. They are B, SA, VARI and NDGVIc (Table 6.5). The most
common used VIs, SR, NDVI SAVI etc. were not correlated with AGB and LAI at alpha level 0.05.
But, RSR, NDVIc and some other VIs were significantly correlated with AGB and LAI. However
they do not perform strong correlation hence were not included in this study result. VARI and
NDGVIc show the strongest correlation with AGB which are 0.751 and 0.803, respectively; while
B and SA were showing strongest negative correlation with LAI which are -0.837 and -0.747,
respectively. Since NDVIc is normalized version of RSR, we use only NDVIc for comparison
reason with NDGVIc. The AGB and LAI are plotted against those significantly correlated VIs in
Fig 6.6 and 6.7.

66

VIs.
Vegetation index
Aboveground biomass Leaf area index
Pearsons correlation P value Pearsons correlation P value
B -0.709 0.005 -0.837 0.000
SA -0.640 0.014 -0.747 0.002
VARI 0.751 0.002 0.643 0.013
NDVIc 0.649 0.012 0.609 0.021
NDGVIc 0.803 0.001 0.694 0.006

6.4.4 Models for AGB and LAI
Either simple linear or nonlinear regression equation was used to develop AGB and LAI
models. The regression models for AGB and LAI, and coefficients of determination (R
2
) are shown
in Table 6.6. Complete test statistic results are given in Appxs 9-18, where parameter estimation of
coefficients, analysis of variance (ANOVA), Durbin Watson (D-W) Kolmogorov-Smirnov (K-S)
and Constant variance (CV) test statistics and other statistical measures reported. The variables (VIs,
AGB and LAI) were transformed before the analysis and checked whether their linear relationship
is improved. And only the best regression models (for each VI and AGB/LAI) are presented in this
paper among those transformed-linear/nonlinear and non-transformed-linear/nonlinear versions.
Generally, transformations enhance the linear correlation, but in some cases, they do not improve
the correlation, significantly. Also, if the test statistics fail after transformation, original data values
were applied for regression equations.
The highest R
2
values resulted from using the NDGVIc, as expected in case of AGB; while, B
and SA gave the highest R
2
values for the LAI.

67

Log(Brightness)
1.20 1.25 1.30 1.35 1.40 1.45 1.50
A
b
o
v
e
g
r
o
u
n
d

b
i
o
m
a
s
s

(
t
/
h
a
)
-10
0
10
20
30
40

Log(Surface albedo)
2.0 2.1 2.2 2.3
A
b
o
v
e
g
r
o
u
n
d

b
i
o
m
a
s
s

(
t
/
h
a
)
-10
0
10
20
30
40

Visible atmospheric resistand index
-0.2 -0.1 0.0 0.1
A
b
o
v
e
g
r
o
u
n
d

b
i
o
m
a
s
s

(
t
/
h
a
)
-5
0
5
10
15
20
25
30
35

NDVIc
0.0 0.2 0.4 0.6 0.8
A
b
o
v
e
g
r
o
u
n
d

b
i
o
m
a
s
s

(
t
/
h
a
)
0
5
10
15
20
25
30
35

NDGVIc
0.00 0.05 0.10 0.15 0.20 0.25
A
b
o
v
e
g
r
o
u
n
d

b
i
o
m
a
s
s

(
t
/
h
a
)
0
5
10
15
20
25
30
35

Fig. 6.6 Relationship between VIs and AGB for forest stands. Log transformed Brightness index
and aboveground biomass (AGB) (top left), log transformed Surface albedo and AGB (top right),
visible atmospherically resistant index and AGB (middle left), NDVIc and AGB (middle right), and
SWIR corrected NDGVI and AGB (lower left). The SWIR correction substantially reduced NDVI
values over low AGB.

68

Log(Brightness)
1.20 1.25 1.30 1.35 1.40 1.45 1.50
e
x
p
(
L
e
a
f

a
r
e
a

i
n
d
e
x
)
0
10
20
30
40
50
60

Log(Surface albedo)
2.0 2.1 2.2 2.3
e
x
p
(
L
e
a
f

a
r
e
a

i
n
d
e
x
)
0
10
20
30
40
50
60

Visible atmospheric resistant index
-0.2 -0.1 0.0 0.1
L
o
g
(
L
e
a
f

a
r
e
a

i
n
d
e
x
)
0.2
0.3
0.4
0.5
0.6
0.7

exp(NDVIc)
0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2
L
o
g
(
L
e
a
f

a
r
e
a

i
n
d
e
x
)
0.2
0.3
0.4
0.5
0.6
0.7

exp(NDGVIc)
1.0 1.1 1.2 1.3
L
o
g
(
L
e
a
f

a
r
e
a

i
n
d
e
x
)
0.2
0.3
0.4
0.5
0.6
0.7

Fig. 6.7 Relationship between VIs and LAI. Logarithm transformed (log) Brightness and
exponential transformed (exp) leaf area index (LAI) (top left), log-Surface albedo and exp-LAI (top
right), VARI and log-LAI (middle left), exp-NDVIc and log-LAI (middle right), and exp-NDGVIc
and log-LAI (bottom left).

69

Table 6.6 Regression models for aboveground biomass (AGB) and leaf area index (LAI)
Band or index AGB model R
2
LAI model R
2

B 159.1833 -111.3617*log(x) 0.54 exp(-1.3711/(1 - 0.8183*log(x))) 0.83
SA 199.4698 - 91.7518*log(x) 0.48 exp(-0.9390/(1 - 0.5157*log(x))) 0.81
VARI 10.3865 +84.0871*x 0.56 log(0.4157/(1 - 2.4598*x)) 0.48
NDVIc exp(4.3349*x) 0.49 log(0.0400+0.2394*exp(x)) 0.40
NDGVIc 3.7505/(1 - 3.9873*x) 0.70 log(-0.7597+1.0489*exp(x)) 0.50

6.4.5 Index validation
Regression model could not test considering limited number of forest stands. However the
correlations found in this study have been tested by using two other imageries. We run correlation
analysis between AGB and/or LAI and VIs. Study results show that AGB is correlated with B, SA
and NDVIc on fall 2007 image, and VARI and NDVGIc have significant correlation on fall 2006
image; while LAI correlates with B and SA on fall 2007 image and all those VIs have significant
correlation on fall 2006 image (Table 6.7).

Table 6.7 Correlations of the AGB of forest stand and LAI against the VIs.
VI
Sep 15, 2006 Sep 02, 2007
AGB LAI AGB LAI
B
-0.461 -0.744 -0.707 -0.788
0.097 0.002 0.005 0.001
SA
-0.423 -0.677 -0.730 -0.782
0.132 0.008 0.003 0.001
VARI
0.566 0.589 0.090 -0.023
0.035 0.027 0.759 0.938
NDVIc
0.481 0.635 0.533 0.460
0.081 0.015 0.050 0.098
NDGVIc
0.605 0.590 0.402 0.251
0.022 0.026 0.155 0.387
Cell Contents: Pearson correlation
P-Value
6.5 Discussion
6.5.1 Relationship between AGB, LAI, Landsat TM bands and band ratios
Red band (band 3) was the single band with a strong correlation with AGB and the strongest
correlation with LAI. The chlorophyll pigment in the green leaves absorbs radiation in the red

70

wavelengths and red reflectance is thus inversely related to the quantity of chlorophyll present in
the canopy (Tucker and Seller 1986). Dahlberg (2001) reported that the red bands of the Landsat
TM, SPOT XS and IRS LISS were the best predictors of the AGB and LAI in birch forests. The
high correlation between the red bands and forest variables has also been observed in the more
productive deciduous stands (Hame et al. 1997, Eklundh et al. 2003) and in the broadleaved stands
in the savannes (Xu et al. 2003). Furthermore, the red band has been among the most correlated
bands with the forest variables in the coniforiuous stands (Spanner et al 1990, Hame et al. 1997).
Green band (band 2) was the second best correlated band with LAI but was not correlated
with AGB. Fassnacht et al (1997) observed the poor performance of the red band in the hardwood
forests and identified the green band to be the most correlated one with LAI in the visible range.
NIR band (band 4) showed a weak negative correlation with LAI and is not correlated with
AGB. By theory, in the NIR range reflectance should be high (Tucker and Seller 1986), which
increases the canopy reflectance as LAI increases. Dahlberg (2001) found that in the mountain birch
forests the correlation between the NIR band and LAI was relatively strong compared to the
relationship between NIR band and AGB. The direct relationships between NIR band and other
forest variables have also been reported for other types of deciduous stands (Hame et al. 1997,
Eklundth et al. 2003), in contrast to the the typically inverse relationship observed in the coniferous
stans (Nilseon and Peterson 1994, Hame et al. 1997, Eklundth et al. 2003). Broadleaved trees with
large trees reflect effectively, but in th coniferous canopies the reflectance I reduced due to
increasing multi-layer absorption as the height of the canopy increases (Hame et al. 1997).
SWIR1 band (band 5) has the strongest negative correlation with AGB but weak negative
correlation with LAI. This is in agreement with the data from mountain birch forests (Dahlberg
2001, Heiskanen 2006) but in contradiction to the data from other types of deciduous stands
(Eklundh et al 2003). However, a strong correlation has been found between reflectance in SWIR
bands and forest variables in coniferous stands (Ardo 1992, Eklundh et al. 2003). In the SWIR
wavelengths, the reflectance will decrease with increasing leaf area as consequence of increasing
absorption due to the water in the canopies (Tucker and Sellers 1986). Shadowing is also likely to
decrease reflectance in all bands (Ardo 1992).
Band composition correlation analysis is important factor to choose VIs among those number
of available VIs. We found a composite image containing TM bands G/R/SWIR1 to be the most
optimum for distinguishing stands that contain different AGBs and bands G/R/NIR to be the most
optimum for LAI. A band combination, G/R was not employed in the case of AGB model, because
a correlation was improved in a triple band combination, G/R/SWIR1, from r=0.744 to r=0.797.
Also, it does not show a good correlation in the case of LAI. The reason, a band combination

71

R/SWIR1 being not employed in this study is that the correlation is reduced from single band R, -
0.808 to double band -0.741, in the case of LAI. Also the combination was not significant in the
case of AGB.

6.5.2 Relationship between AGB, LAI and VIs
During the course of the growing period, when the vegetation fraction ranges from 60% to
100%, canopy absorption in the NIR range increases and transmittance decreased in the case of
wheat (Stark and Gitelson 1999, 2000). Perhaps, during fall season, such behavior of reflectance,
absorption and transmittance in the NIR range is a clue of the relationship between AGB and/or
LAI and VARI in our study case. But, the relationships were not strong enough for further
application.
Several studies have found that SWIR modification of the SR and NDVI improve the
correlation with LAI in the coniferous forests (Nemeni et al. 1993, Brown et al. 2000, Chen et al.
2002, Stenberg et al. 2004) and in the deciduous forests (Chen et al. 2002). But, in the comparison
of Eklundh et al. (2003) and Heiskanen (2006), RSR (which is the SWIR modification of the SR)
performs poorly in the deciduous stands. In this study, RSR shows a weak linear relationship
against the forest variables, but could improve the correlation of the SR. The result of Chen et al.
(2002) suggests that the relationship approaches exponential if higher values of LAI occur. NDVIc
has a stronger linear correlation with the forest variables than NDVI, and the relationship with
NDVI was not significant. Both RSR and NDVIc are sensitive to the SWIRmin and SWIRmax
values that used (Nemani et al. 1993, Brown et al. 2000), which is also likely to explain the
differences in the model performance in the different studies.
NDGVI is just the same as NDVI but sensitive in visible region. Since the correlation analysis
of ratioed band suggests that G/R/SWIR is better correlated with AGB, we integrated NDGVI with
SWIR correction factor. The newly integrated NDGVIc achieves the highest correlations related to
the VIs, with AGB. The low water contents (associated with low biomass in less dense canopies)
are responsible for the high SWIR response (Nemani 1993).
The relations were between VIs, B and SA with LAI. Because to drive B and SA, all six
bands blue, green, red, NIR and SWIR were used. Inclusion of all spectral bands enable the VIs, B
and SA to take account on each light region, especially when the visible region become more
sensitive than NIR, yet still include NIR and SWIR bands in fall season. In the midlatitudes,
deciduous trees shed leaves, while evergreens experience much more elaborated phenological
cycles; individual leaves may experience senescence separately (i.e. trees do not necessarily shed all
leaves simultaneously), and individual trees or branches of trees may shed leaves on cycles quite

72

distinct from others in the same forest (Koriba 1958). During the onset of senescence, deterioration
of cell walls in the mesophyll tissue produces a distinctive decline in infrared reflectance; an
accompanying increase in visible brightness may be the result of decline in the abundance and
effectiveness of chlorophyll as an absorber of visible radiation (Campbell 2006).
The soil line adjusted VIs have been employed successfully in some studies (Broge and
Leblanc 2000, Nemani 2006). SAVI2 was the best predictor of LAI and the least affected index by
the background reflectance in the sensitivity analysis of Broge and Leblanc (2000). In this study,
soil line adjusted VIs which are SAVI, SARVI and OSAVI are not correlated with AGB and LAI,
significantly.

6.5.3 AGB and LAI estimation using regression equation
The exponential transformation was the best to represents negative relationship and the log
transformation was the best to represents positive relationship in case of LAI, while no
transformation functions were improved AGB on either positive and negative relationships. In the
case of AGB, rational model with non-transformed NDGVIc variable produced the highest R
2
. In
the case of LAI, rational model was with both log transformed VIs, B and SA produced the highest
R
2
. Band ratioing enables the statistical comparison of the single band indices with multiple band
effect, and suggests which spectral combination was better. Cohen et al. (2003) combine several
VIs into single index and Heiskanen (2006) finds that combining bands is also useful to develop
regression model. Because of limited study plots, species specific correlations were not checked.

6.5.4 Effect of the undergrowth vegetation and background reflectance
This study is concerned with the forest AGB and/or LAI and excludes grass which is
photosynthetically active material. This material may have contributed to the poor relationship
observed. For example, as it was observed in the field, a stand contains few trees would have more
grass and herbs. Therefore, that stand would have a small quantity of AGB and/or LAI whereas the
photosynthetic activity of vegetation in the same plot is high, hence high VI. Thus, this may explain
negative correlation between NIR band (band 4) and LAI.
The satellite sensor directly measures the reflectance of the foliage and not the woody
portions of the tree. Woody biomass does not contribute much to spectral reflectance. As such a
relationship between woody biomass and foliage has to be sought first before biomass could be
related to spectral reflectance. However, this relationship did not found in our study.
Where there is more biomass, one would expect high photosynthetic activity but this is not the
case because stem biomass accumulates or is made over time, and photosynthetic activity is

73

measured at an instant, hence the correlation between stem biomass and VIs are low. With increase
in age of forest there is a steady decrease in chlorophyll accumulation (Thenkabail 2004). For
example, a young birch stand has high photosynthetic activity and low biomass whereas an old
larch stand may have a low photosynthetic activity and high biomass. In this study, some forest
stands were contained with few large trees around dbh>50 cm which significantly increase stand
biomass while their spectral contribution to the pixel may be less.

6.5.5 Model verification
The objective of this study was to develop regression model between AGB and/or LAI and
VIs. The results suggest that the best models explain AGB 70% by NDGVIc and LAI 83% by
Brightness index.
A reason for relatively good results obtained is that the field plots consist of only two
different tree species, which is in contradiction to the other types of evergreen stands in the Khentii
forests area typically consisting of multiple species. These two species helped to elongate the data
distribution for the regression analysis.
The disadvantage of the model is that errors and variations are not tested due to limited study
plots. It is important to remember the possible overprediction, due to the saturation (nonlinear
function) on both AGB and LAI when using the best VIs, NDGVIc and Brightness, relatively. But,
linear version of VIs could predict relatively stable regardless of extreme low and max values and
also saturation would not appear in the relationships as long as linear model is used. The linear
versions of the best models explain the relationships between on AGB and/or LAI and VIs around
59% and 65%, relatively. The study results are consistent, and sometimes better than results of
previous studies done in Mongolian forests by using Landsat TM imagery (Tsolmon 2003, Ariunzul
2008).
Considering the few data plots used in this study, additional research should however be taken
to verify the significance of the relationship model of VI and AGB and/or LAI.

6.5.6 Suggestions
We use two different date images to test our newly developed VI index NDGVIc and other
VIs to check whether they have same kind of relationship with AGB and LAI. AGB and LAI data
sets used in this study were sensitive in light regions which are visible and SWIR.
We run correlation analysis between AGB and/or LAI and VIs. Study result show that
biomass data and NIR and SWIR1 bands B, SA and NDVIc have significant correlation on 15 Sep
2006 image, while SWIR1 band VARI and NDVGIc have significant correlation on 2 Sep 2007

74

image (Appx 19 and Table 7). Based on the results of these three fall images certain AGB and/or
LAI and behavioral characteristics of trees reflectance can be pointed out. These are useful to
decide under what circumstances either NDVIc, NDGVIc, B or SA can be chosen based on
correlation analysis of within bands.
- if NIR band correlate with AGB and/or LAI, NDVIc shows better result. But NDGVIc is
better if NIR band does not correlate with AGB and/or LAI while SWIR band shows
significant correlation with these variables.
- if NIR band (band 4) has correlation with Red (band3) and SWIR band, NDVIc would be
better. But if SWIR does not correlated with either any of the bands, traditional NDVI
would give better result to predict AGB and LAI.
- if Green band (band 2) correlate with Red band (band 3) and not related with NIR band,
NDGVI would give better prediction.
- if SWIR bands (bands 5 and 7) correlate with visible bands, SWIR correction would
contribute better result and help to improve prediction in either case NDVIc and NDGVIc. It
seems SWIR bands do not correlate with NIR band during fall season, therefore SWIR
correction factor gives more effect on NDGVI than NDVI.
- if Green, Red and SWIR bands correlate, then NDGVIc would give better result than non
SWIR corrected VIs. And if deduction result (Green-Red) gives minus sign, the adjustment
coefficient of L should be used. This coefficient has no physical meaning but helps to bring
the values bigger than zero and helps to increase the index range. Therefore, one should
choose appropriate adjustment value depending on the index values.
- if the deduction result (Green-Red) gives plus sign L adjustment coefficient is not necessary.
But according to our result, L coefficient still contributes better performance to the index.
- if either Red or Green band does not relate with NIR band, then Brightness and SA indices
give better result to predict AGB and LAI.
These indices were tested only in fall images where chlorophyll absorption lessened and green
light extinct. However their result was promising to predict AGB and LAI in the Boreal forests.
Therefore, these indices can be a useful tool to determine vegetation phenology.

6.6 Conclusion
The potential of the visible to shortwave infrared satellite data for estimating stand AGB and
LAI in mountain larch, birch and mixed forests was examined by using Landsat TM data. The
results show some statistical dependence between the stand AGB, LAI and satellite data. This
relationship was modeled by regression analysis. Keep the accuracy of the estimates in mind, the

75

developed models need to be tested in another ground collected stand biomass data. However,
understanding the relationship pattern which found in this study is the first step to develop
relationship model and this result could be employed to map AGB and LAI in the study area in the
future. The factors affecting the reflectance of mountain larch and birch, and mixed as well as
another species stands should be studied further by sensitivity analysis, by using a suitable canopy
reflectance model. Particularly, the causes of the saturation in the highest biomass and LAI values,
and the effect of the undergrowth vegetation on the canopy reflectance in a continuum of canopy
closure should be studied. The sensitivity analysis would also provide the optimal foundation for
the selection of the most suitable VIs for AGB and LAI estimation in mountain forests. The SWIR
band has a good effect to improve AGB and/or LAI relation with VIs acting as correction factor in
Landsat image on fall season around brown shift. The spectral features of larch and birch forests
reflectance revealed in this study would conceivably be comparable with other forest types. It
should be stressed, however, that the applicability of the proposed techniques to other forest types
must yet be verified. Reflectance from green vegetated surface in the visible region of the spectrum
is often quite low and the difference in the levels of reflectivity among the visible bands is much
smaller than the difference between red and NIR. Despite the subtle differences in response from
vegetation in the visible range, the sensitivity of our newly suggested indices is better than other
suggested indices. Index validation results of AGB estimation, using data acquired by different
dates, show that the same relationship was found and proposed index was useful. Nevertheless, we
recognize that the difference in species, illumination, canopy architecture, soil properties and many
other factors may potentially decrease the accuracy of the developed indexes as well as other tested
indexes. In order to devise comprehensive algorithms for remote monitoring of AGB and LAI,
additional research dealing with the sensitivity of the newly suggested indices to these external
factors is required.

76

Chapter 7
7 Mapping land cover and forest types using multi-temporal Landsat imageries with
biophysical information of dominant forest types in mountainous area

7.1 Abstract
Forest is a critical component of the global carbon cycle, and species classified forest cover map
allows for assessing the impacts of different tree species on carbon dynamics. Earth observation
satellite data sets are important source of information that allows systematic monitoring of the entire
region to work. A method of North-Easthern Mongolian forest cover and species mapping which
use Landsat imagery has been developed and tested in East Khentii by summer, fall and winter
images acquired in different years. The approach employs a different seasonal compositing
methodology adapted for improving land class and species separation by using seasonal variation of
forest species-reflectance. Then adjusted tree species upper or lower distributions based on forest-
vegetational biophysical characteristics of altitudinal-belt zones within the study area. Support
vector machine (SVM) and Maximum likelihood classifier were applied to the seasonal image
composites to characterize land cover and larch cedar and broadleaf species separation over the
study area. Classification results of single-date Landsat imageries range overall accuracy of 91-93%,
Kappa of 84-93. While multi-temporal image classification results had higher agreement than
single-date imagery in both SVM and Maximum likelihood cases (overall accuracy of 96-97%,
Kappa of 0.92-0.95), respectively. Then applying ancillary information of forest-vegetational
biophysical characteristics of altitudinal-belt zones improved the separation of forest type
approximately by 30% for larch in exchange 5% decreasing accuracy for cedar. Our result
highlights a potential method of tree species separation based on seasonal differences of reflectance
characteristics and forest-vegetational biophysical characteristics of altitudinal-belt zones in
Mongolia.

7.2 Introduction
Taiga forest is a critical component of the global carbon cycle, storing the largest amount of
carbon of all global biomes (IPCC 2000, Kasischke and Stock 2000). The taiga forests contain
approximately twice as much carbon per unit area as that found in tropical forests. About 44% of
boreal forest remains as intact landscapes, isolated from industrial management practices (FAO
2006, Potapov et al. 2008, Yaroshenko et al. 2009). These intact forest ecosystems have
accumulated carbon for centuries. Extension of logging and human-caused fire of the boreal biome,
coupled with possible changes in natural disturbance regimes due to climate change, may lead taiga

77

forests to be a significant source of carbon to the atmosphere (Kurz et al. 1998, Kasischke and
Stocks 2000, Luyssaert et al. 2008, Bradshw et al. 2009, Pimm et al. 2009).
The best classification accuracies in any land cover classification are usually obtained with
multiple imagery that captures different periods of the growing season (multi-seasonal
imagery)(Civco 1989, Oetter et al. 2001, Prishchepov 2012, Rodriguez-Galiano et al. 2012, Wagner
et al. 1993, Wolter et al. 1995). Multi-seasonal imagery has proven particularly important when
classifying agricultural land use due to the different sowing and harvesting times among crops
(Guerschman et al. 2003, Homer et al. 2004, Kalensky 1975, PaxLenney and Woodcock 1997) and
monitoring regional scale forest cover change (Potapov et al. 2011) as well as forest species
composition (Wolter and Townsend 2011).
SVM is widely used in classification of multisensor data (Camps-Valls et al. 2006 Hill et al.
2005, Waske and Benediktsson 2007, Wolter and Townsend 2011). SVM is also used for
biogeographic modeling (Guo et al. 2005, Drake et al. 2006, Pouteau et al. 2011 and 2012.) It can
be an optimal statistical tool for our study case since it is adapted to numerous and heterogenous
variables and mixtures of qualitative and quantitative variables. In Waske and Benediktsson (2007),
a range of the most commonly used classifiers are compared for multisource classification, namely
maximum likelihood, decision trees and support vector machines. According to their results, SVM
gives the best accuracy. Artificial neural networks is another widely used classifiers but shows lacks
face to SVM in other studies (Foody and Mathur 2004, Pal and Mather 2004 Nemmor and Chibani
2006). It was unclear how the use of other robust classifiers (e.g. random forest classification
algorithm) to identify subtle signal changes could overcome image data dependence.
SVM is a critical tool for fusion of multisource data. They Waske and Benediktsson (2007)
compare two fusion schemes: the single SVM and the fused SVM (In our consideration single
date image is single SVM and multi-date images are fused SVM). In the single SVM method, an
overall SVM is trained on all sources together whereas in the fused SVM, a single SVM is trained
on each source separately, and then an additional SVM is applied as a decision rule on all
monosource rule images together. The fused SVM method outperforms the single SVM and appears
to be easier to implement than the classification scheme in Hill et al. (2005) or the use of composite
kernel proposed in Camps-Valls et al. (2006). Consequently, we use hereinafter the fused SVM for
its accuracy and ease of implementation trade-off.
The Maximum Likelihood classification method is well known for the analysis of the
satellite images (e.g. Poulin and Careau 2002, Hagner and Reese 2007). So far, satellite image
interpretation which use the Maximum Likelihood approach was mostly applied for land cover
classification (Huang et al 2007) and monitoring of land-use changes (Shalaby and Tateishi 2007),

78

as well as for soil salinity mapping (Katawin and Kotrapat 2005) and peat land mapping (Haapanen
and Tokola 2007) showing overall high accuracies (mostly over 80%).
Our goal is to identify an approach to accurately map land classes and forest type by using
30-m resolution Landsat TM/ETM+images. To accomplish this goal, we composite different
seasonal imageries to improve the effects of discrimination for land cover and forest types and then
adjusting forest type distributions by ancillary information of altitudinal-belt zone on resulting
spatial patterns of land classes and tree species (Fig 7.1).

7.3.1 Study area
Description of study area is given in heading 5.3.1, chapter 5.

7.3.2 Landsat TM data
Landsat TM/ETM+images that capture different phenological stages (i.e. the beginning and
middle of the growing season) are not always available in single year due to issues such as clouds
and haze. We have used five images: one summer, three falls and one winter images (J uly 26, 2002;
September 15, 2006; September 02, 2007; September 07, 2009; December 12, 2009). They were
only cloud-free available images over study area. We found that, there existed not a single foot-
print for which cloud-free images recorded in spring. Three fall images were selected, because
elevated red reflectance due to autumn foliar senescence in deciduous forest and before dormant
state of evergreen forest species. However, timeframe of deciduous larch and birch species
senescence are unknown for this study area. We limited cloud contamination to be less than 5%
because cloud cover and cloud shadows in multiple images are typically additive, resulting in the
removal of half of images. All five Landsat images were downloaded from the USGS Earth
Resources Observation and Science Center (EROS) web site (source:http://glovis.usgs.gov/). The
images were orthorectified and georeferenced to a UTM Zone 49 projection based on the WGS 84
datum (source: www.landcover.org). The spatial resolution of image from the Landsat TM/ETM+is
30 m for all spectral bands except for band 6, which is 60 m. The spectral resolution involve visible
- bands 1 (0.45-0.52 um), 2 (0.52-0.60 um) and 3 (0.63-0.69 um), NIR - band 4 (0.76-0.90 um),
SWIR bands 5 (1.55-1.75 um) and 7 (2.08-2.35) and thermal infrared (TIR) band 6 (10.4-12.5
um). The temporal resolution is 16 days.

79

Fig. 7.1 Framework of image processing and forest type map

80

7.3.3 Image preprocessing
An atmospheric correction procedure was done by ATCOR3 algorithm (Richter 2000a) to all
7 bands of each Landsat image. ATCOR3 employs digital elevation model (DEM) for topographic
correction and spatial resolution of a 1 arcsecond (approximately 30 m) ASTER DEM was used.
The image processing framework is given in Fig 7.1.

7.3.4 Classifiers
The ranges of the most commonly used classifiers are compared for multi-seasonal
classification, namely maximum likelihood, decision trees and support vector machines. Artificial
neural networks and random forest classifications are other widely used classifiers which are not
compared in the later study. On a seasonal base, relationships between informational classes and
spectral classes are not constant and relationships defined for one image cannot be extended to
others. Therefore, multiple images represent different seasons would give general distribution
boundary of any species.

7.3.4.1 Support vector machines (SVM)
SVM, originally introduced as a binary classifier (Vapnik and Chervonenkis 1998) and uses
two classes of training samples within a multi-dimensional feature space to fit an optimal separating
hyperplane. In this way, it tries to maximize the margin which is the distance between the closest
training samples, or support vectors, and the hyperplane itself.
SVM consists of projecting vectors into a high dimensional feature space by means of a
kernel trick then fitting the optimal hyperplane that separates classes using an optimization function.
For a generic pattern x, the corresponding estimated label is given by Eq 7.1.

=sign[f (x)] =sign[sum(i from 1 to N)y
i
.
i
.K(x
i
, x) +b] Eq.7.1

wherein N is the number of training points, the label of the ith sample is y
i
, b is a bias parameter,
K(x
i
,x) is the chosen kernel and
i
denotes the Lagrangian multipliers.
SVM, introduced by Vapnik and Chervonenkis (1998), is extensively described by Burges
(1998), and Scholkopf and Smola (2002). Several kernels are used in the literature and generally,
the Gaussian radial basis function (RBF) has advantages over other well working kernels (Hsu et al.
2009). Therefore, we used radial kernel (Eq 7.2) in this study.

K(x
i
, x) =(-||x
i
T
, x||
2
), > 0 Eq. 7.2

81

Only
i
belonging to support vectors s
i
has no null value so the classification function is
actually Eq 7.3 (Pouteau et al. 2012).

=sign[f (x)] =sign[sum(i from 1 to P
s
)y
i
.
i
.K(x
i
, x) +b] Eq. 7.3

7.3.4.2 Maximum Likelihood
Maximum likelihood classification assumes that the statistics for each class in each band are
normally distributed and calculates the probability that a given pixel belongs to a specific class.
Each pixel is assigned to the class that has the highest probability (that is, the maximum likelihood
Eq 7.4). ENVI implements maximum likelihood classification by calculating the following
distriminant functions for each pixel in the image (Richards 1999):

g
(x) lnp(
)
1
2In|l
i
|
1/ 2(x m
)
t
X
-1
(x m
) Eq. 7.4

where i class, x n dimensional data (where n is the number of bands), p(
i
) probability that
class
i
occurs in the image and is assumed the same for all classes, |
i
| - determinant of the
coviarance matrix of the data in class
i
,
i
-1
its inverse matrix and m
i
mean vector

7.3.5 Region of interest (ROI) for classification
In our ground truth methodology, we used unsupervised classification for homogenous ROI.
Unsupervised classification can be defined as the identification of natural groups, or structures,
within multispectral data. The notion of the existence of natural, inherent groupings of spectral
values within a scene may not be intuitively obvious, but it can be demonstrated that remotely
sensed images are usually composed of spectral classes that are reasonably uniform internally in
respect to brightness in several spectral channels. Unsupervised classification is the definition,
identification, labeling, and mapping of these natural classes (Campbell 2006). After preprocessing,
an unsupervised classification is applied on all the thirty five spectral bands.
In this way, we can easily detected homogeneous pixels to choose region of interest (ROI)
since it is proved to have a good accuracy/confusion trade-off. When a homogeneous land cover is
detected, two ROI are sampled in each class to train a supervised classification.
A ROI has to be good enough to be representative and pure enough to be homogenous. The
optimal ROI, also called classes, can be defined as the ROI separability to achieve the high

82

accuracy land cover and forest species map. We have chosen six classes including grassland, larch,
cedar, water, bare soil and broadleaf (Table 7.1) based on unsupervised classification result and
forest inventory data. We did not choose water for validation (reference) due to the limited number
of water pixel. ROI separability reports for single-date imageries and multi-temporal imagery are
given in Appx 20-25.

Table 7.1 Training pixels and reference pixels used for image classification and validation
Number of training pixels Number of reference pixels
Grassland 19426 19113
Larch 2991 3281
Cedar 6396 6788
Water 147
Bare soil 2371 529
broadleaf 554 379

Scholz et al. (1979) and Hixson et al. (1980) discovered that selection of training data may
be as important as or even more important than choice of classification algorithm in determining
classification accuracies of agricultural areas in the central United States. They concluded that
differences in the selection of training data were more important influence upon accuracy than were
differences among some (Scholz et al. 1979, Hixson et al. 1980) different classification procedures.
The results of their studies show little difference in the classification accuracies achived by
the five classification algorithms they considered if the same training statistics were used. However,
a classification algorithm which gave two alternative training methods for the same data produced
significantly different results. This finding suggests that the choice of training method, at least in
some instances, is as important aspect of training fields is that all cover types in the scene must be
adequetly represented by a sufficient number of samples in each spectral subclass.
The ROI separability function computes the spectral separability between selected ROI pairs
for a given input file. Both the J effries-Matusita and Transformed Divergence separability measures
are reported. These values range from 0 to 2.0 and indicate how well the selected ROI pairs are
statistically separate. Values which are greater than 1.9 indicate that the ROI pairs have good
separability. For ROI pairs with lower separability values, we should attempt to improve the
separability by editing the ROIs or by selecting new ROIs. For ROI pairs with very low separability
values (less than 1), we might need to combine them into a single ROI (Richards 1999).

83

7.3.6 Classification accuracy assessment
As widespread used overall accuracy (OA) can give misleading results, other metrics we are
used hereinafter to assess classification results:
- Kappa coefficient expresses whether correctly assigned pixels may have been assigned by chance
or not based on the classification decision rule (Cohen 1960). A value of 1 indicates perfect
agreement and 0 indicates a pattern arising by chance;
- producers accuracy (PA) represents the number of correctly assigned pixels for a class divided by
the actual number of ground truth pixels for that class;
- and users accuracy (UA) is the number of correctly assigned pixels for a class divided by the total
pixels assigned to that class (Congalton and Green 2009).

7.3.7 Ancillary biophysical information for post-classification
Mountain forests often display a characteristic discontinuity in their distribution. This
boundary represents the upper limit of forest canopies and is associated with temperature decrease
along elevational gradients. The elevation of forests can be viewed as the outcome of a
confrontation of abiotic and biotic factors in which the climate of higher elevations becomes
increasingly less favourable and interacts with the tolerance of tree species, determining their upper
limit. Studies of mountain forest distributions at regional and global scales identified a strong
association between temperature and forest elevation (Troll 1973, Cogbill and White 1991,
Malyshev 1993). Air temperature is one of the most important factors controlling vegetation
zonation and key processes such as evapotranspiration, carbon fixation and decomposition, plant
productivity and mortality in mountain ecosystems (Chen et al. 1999, Nagy et al. 2003). Differences
in temperature due to landscape position can have a prominent effect on water balance, and as a
result, hydrologic and ecologic processes (Dobrowski et al. 2009).
The landscape studies of natural ecosystems are successfully categorized, classified and
mapped based on a concept of territorial distribution characteristics of nature (Vostokova et al. 1995,
Gunin et al. 2005). Based on natural ecosystem distributions, various levels of categorization are
available for an organization: I level elementary of monoecosystems, II level mesoecosystems
and III level macroecosystems (Vostokova et al. 1995).
Altitude-zonal forest types which were introduced by Smagin (1965) relate to
mesoecosystems. An altitude-zonal forest complex is an association of forest types in the system of
ecogenetic line, reflecting simultaneously zonal-provincial and altitudinal features of climate and
soil of the certain altitudinal zone. Every altitude-zonal forest complex characterized by the set of
forest types, soil type, life form of forest stand and their succession trend.

84

First, Korotkov (1976, 1978) divided Mongolian forest-vegetation regions into three regions
and eight provinces for Northern Mongolia where forests grow. Korotkov and Tsedendash (1983)
divided forest-vegetaiton regions into three regions and nine provinces. He used group of zonal
characteristics to divide forest-vegetation regions and for forest province division a zonal
characteristic used. Later, Tsedendash (1993) divided forest-vegetation regions into district level for
mountainous region of Khentey and Dugarjav (1996, 2006) divided forest-vegetation into district
level in whole Mongolian case.
In this study, we used ecosystem categorization of level II, mesoecosystem (altitudinal belt
zone). Altitudinal belt zones of Khentey provice are shown in Table 7.2. In the East Khentii,
altitudinalbelt structure of forest ecosystem comprises of subtaiga larch forests (42%), taiga
Siberian pine and larch forests (48 %), subgoltsy Siberian pine forest and Siberian pine-larch
woodland (10 %).

Table 7.2 Forests altitudinal belt zones in Eastern Khentey province, Mongolia
Subtaiga-forest steppe Taiga Subgoltsy*
Elevation 960-1700 m 1700-2000 m 1900-2000 m (2200 m)
General forest type Larch forest with pine Larch forest with cedar Cedar forest with larch
Forest types Pure larch
Pure pine
Pure birch
Mixed pine-larch
Mixed cedar-larch
Mixed larch-cedar
Pure cedar
Mixed larch-cedar
*goltsy local Russian name for Bald Mountains

We adopt the biophysical distribution of general characteristics of forest types (Table 7.2) for
improving image classification result. In order to apply biophysical information, in Table 1, both
cedar and larch forests are limited to distributing by their general biophysical distribution
characteristics of elevation (Fig. 7.2 and Table 7.3). The assumption is that cedar trees do not grow
lower than 1700 m and pure larch forests is not distributed in higher elevation 2000 m in this
province. This idea supported by forests distribution characteristics of belt zones in elevation of two
major dominant tree species which are light green coniferous forest (larch) dark green coniferous
forest (cedar) (Dugarjav 2006).

85

Fig. 7.2 Biophysical information rules of forests distributions along elevation for dominant forest
types post classification.

The pine and the fir forests, as well as any mixed forests were not considered in this
classification. Broadleaved forests represent birch, aspen, poplar, willow and shrubs. This method is
applicable to other regions, since belt zones of elevation are well developed.

86

Table 7.3 Conditional distributions of forest types by elevation
Classified classes
P
o
s
t

c
l
a
s
s
i
f
i
c
a
t
i
o
n

c
l
a
s
s
e
s

Rules Larch Cedar Broadleaf
DEM<960 Broadleaf Broadleaf Broadleaf
DEM<1700 m Larch Larch Broadleaf
DEM>2000 m Cedar Cedar Broadleaf

We have used national forest inventory data-map for alternative accuracy assessment (Fig.
7.3). But only for dominant forest type distribution were assessed due to the data-maps topographic
and geographic distortion because of the most of land cover classes were obviously off the edge of
natural boundary. However, large numbers of dominant classes were partly covered on exact land
classes; so, this frequent number of classes somehow could assess the improvement of post
classification result.

Fig. 7.3 Alternative, accuracy assessment for image post-classification.

87

7.4 Results
7.4.1 Land cover and forest type classification result and accuracy assessment for multi-temporal
imagery
Land cover type and species separabiltiy within Landsat spectral sensitivity region and
classification results of single-date imagies are given in Appx 26-27. Image classification of the
best accuracy yielded on multi-temporal Landsat image composites if it is compared with single-
data images (Tables 7.4-7.6 and Appx. 27). Based on independent ROI data, the overall accuracy of
the classified land cover and forest type was 95% by SVM and was 97% by Maximum likelihood
and the kappa coefficient was 0.92 and 0.95, respectively. The classification accuracy assessment of
SVM and Maximum likelihood are given in Tables 7.4-7.6. According to reference image pixels
and their percentage in classes, both SVM and Maximum likelihood classifier results show that land
classes are separated fairly well. UA of SVM classifier for grassland, cedar and bare soil is 98-99%,
broadleaf is 80% and larch is 72%; UA Maximum likelihood classifier for grassland, larch, cedar
and bare soil is 97-99% and broadleaf is 69%. PA of SVM for six classes approximately is 85-100%;
PA of Maximum likelihood for bare soil is 47% and other five classes are 96-100%.
The error reported in confusion table were caused between the two coniferous tree species
larch and cedar and also the two deciduous tree species broadleaved and larch tree species in SVM
case. In Maximum likelihood case, the error reported in confusion table was caused between the
two deciduous tree species broadleaved and larch tree species. Only limited numbers of water pixels
were available and all pixels were used for image classification and therefore the accuracy
assessment was not possible.

88

Table 7.4 Confusion matrix of multi-temporal image classification result by SVM
C
l
a
s
s
i
f
i
e
d

d
a
t
a

Ground reference data
Class grassland larch cedar bare soil broadleaf Total
Pixels
grassland 19022 0 0 5 0 19027
larch 0 2370 135 0 49 2554
cedar 0 878 6644 0 25 7547
bare soil 91 0 0 524 0 615
broadleaf 0 33 9 0 305 347
Total 19113 3281 6788 529 379 30090
Percent
grassland 99.52 0 0 0.95 0 63.23
larch 0 72.23 1.99 0 12.93 8.49
cedar 0 26.76 97.88 0 6.6 25.08
bare soil 0.48 0 0 99.05 0 2.04
broadleaf 0 1.01 0.13 0 80.47 1.15
Total 100 100 100 100 100 100

89

Table 7.5 Confusion matrix of multi-temporal image classification result by Maximum likelihood
C
l
a
s
s
i
f
i
e
d

d
a
t
a

Pixels
grassland 18516 0 0 8 0 18524
larch 0 3171 58 0 81 3310
cedar 0 110 6725 0 35 6870
bare soil 597 0 1 521 0 1119
broadleaf 0 0 4 0 263 267
Total 19113 3281 6788 529 379 30090
Percent
grassland 96.88 0 0 1.51 0 61.56
larch 0 96.65 0.85 0 21.37 11
cedar 0 3.35 99.07 0 9.23 22.83
bare soil 3.12 0 0.01 98.49 0 3.72
broadleaf 0 0 0.06 0 69.39 0.89
Total 100 100 100 100 100 100

Table 7.6 Classification accuracy assessment of Maximum likelihood and SVM
Class Maximum Likelihood SVM
Commission Omission Commission Omission Commission Omission Commission Omission
Percent Pixels Percent Pixels
grassland 0.04 3.12 8/18524 597/19113 0.03 0.48 5/19027 91/19113
larch 4.2 3.35 139/3310 110/3281 7.2 27.77 184/2554 911/3281
cedar 2.11 0.93 145/6870 63/6788 11.97 2.12 903/7547 144/6788
bare soil 53.44 1.51 598/1119 8/529 14.8 0.95 91/615 5/529
broadleaf 1.5 30.61 4/267 116/379 12.1 19.53 42/347 74/379
PA UA PA UA PA UA PA UA
grassland 96.88 99.96 18516/19113 18516/18524 99.52 99.97 19022/19113 19022/19027
larch 96.65 95.8 3171/3281 3171/3310 72.23 92.8 2370/3281 2370/2554
cedar 99.07 97.89 6725/6788 6725/6870 97.88 88.03 6644/6788 6644/7547
bare soil 98.49 46.56 521/529 521/1119 99.05 85.2 524/529 524/615
broadleaf 69.39 98.5 263/379 263/267 80.47 87.9 305/379 305/347

90

7.4.2 Land cover and biophysically adjusted forest type image classification result and accuracy
assessment
After adjustment of biophysical information of larch and cedar classes, we tested their
accuracy assessment. Based on independent ROI data, the overall accuracy of the classified land
cover and forest type was 94% by SVM and was 93% by Maximum likelihood and the kappa
coefficients were 0.89 and 0.88, respectively. The classification accuracy assessment of SVM and
Maximum likelihood are shown in Tables 7.7-7.9. The changes were done between the two
coniferous tree species larch and cedar in confusion table. UA of SVM for larch decreased
approximately by 24% and for cedar, it increased by 7%; UA of Maximum likelihood for larch
forests decreased approximately by 25% and for cedar, generally, the changes were small which is
2%. PA of SVM for larch increased approximately by 19% and for cedar, it decreased
approximately 17%; PA of Maximum likelihood for larch increased by 3% and for cedar, it
decreased by 17%.

Table 7.7 Confusion matrix of post-classification result by SVM.
C
l
a
s
s
i
f
i
e
d

d
a
t
a

Pixels
grassland 19022 0 0 5 0 19027
larch 0 2976 1255 0 74 4305
cedar 0 272 5524 0 0 5796
bare soil 91 0 0 524 0 615
broadleaf 0 33 9 0 305 347
Total 19113 3281 6788 529 379 30090
Percent
grassland 99.52 0 0 0.95 0 63.23
larch 0 90.7 18.49 0 19.53 14.31
cedar 0 8.29 81.38 0 0 19.26
bare soil 0.48 0 0 99.05 0 2.04
broadleaf 0 1.01 0.13 0 80.47 1.15
Total 100 100 100 100 100 100

91

Table 7.8 Confusion matrix post-classification result by Maximum likelihood.
C
l
a
s
s
i
f
i
e
d

d
a
t
a

Pixels
grassland 18516 0 0 8 0 18524
larch 0 3266 1203 0 116 4585
cedar 0 15 5580 0 0 5595
bare soil 597 0 1 521 0 1119
broadleaf 0 0 4 0 263 267
Total 19113 3281 6788 529 379 30090
Percent
grassland 96.88 0 0 1.51 0 61.56
larch 0 99.54 17.72 0 30.61 15.24
cedar 0 0.46 82.2 0 0 18.59
bare soil 3.12 0 0.01 98.49 0 3.72
broadleaf 0 0 0.06 0 69.39 0.89
Total 100 100 100 100 100 100

Table 7.9 Post-classification accuracy assessment of Maximum likelihood and SVM
Class Maximum Likelihood SVM
Commission Omission Commission Omission Commission Omission Commission Omission
grassland 0.04 3.12 8/18524 597/19113 0.03 0.485/19027 91/19113
larch 28.77 0.46 1319/4585 15/3281 30.87 9.31329/4305 305/3281
cedar 0.27 17.8 15/5595 1208/6788 4.69 18.62272/5796 1264/6788
bare soil 53.44 1.51 598/1119 8/529 14.8 0.9591/615 5/529
broadleaf 1.5 30.61 4/267 116/379 12.1 19.5342/347 74/379
PA UA PA UA PA UA PA UA
grassland 96.88 99.96 18516/19113 18516/18524 99.52 99.97 19022/19113 19022/19027
larch 99.54 71.23 3266/3281 3266/4585 90.7 69.13 2976/3281 2976/4305
cedar 82.2 99.73 5580/6788 5580/5595 81.38 95.31 5524/6788 5524/5796
bare soil 98.49 46.56 521/529 521/1119 99.05 85.2 524/529 524/615
broadleaf 69.39 98.5 263/379 263/267 80.47 87.9 305/379 305/347

92

7.4.3 Dominant forest type, larch and cedar mapping accuracy assessment
Larch and cedar forest covers of national forest inventory data are comparable with the study
results (overall accuracy of 68%, Kappa of 0.31) in the case of SVM and (overall accuracy of 63%,
Kappa of 0.24) in the case of Maximum likelihood (Table 7.10). These results are further improved
by ancillary information of forest-vegetational biophysical characteristics of altitudinal-belt zones
(overall accuracy of 78%, Kappa of 0.55) in the case of SVM and (overall accuracy of 74%, Kappa
of 0.48) in the case of Maximum likelihood (Table 7.11) (Fig. 7.4).

Table 7.10 Accuracy assessment for the larch and cedar cover type map of the study area with
comparison of inventory map-data
Inventory reference data
C
l
a
s
s
i
f
i
e
d

d
a
t
a

Class larch cedar Row Total User Acc.
SVM
larch 6136 3404 9540 0.64
cedar 6799 15929 22728 0.70
Col. Total 12935 19333 32268

Prod.Acc. 0.47 0.82 Diag. Total 22065

Overall Accuracy 0.68

Kappa Coefficient 0.31
Maximum likelihood
larch 7153 5619 12772 0.56
cedar 5579 11912 17491 0.68
Col. Total 12732 17531 30263




93

Table 7.11 Accuracy assessment for the larch and cedar cover type (that was adjusted based on
forests distribution characteristics along elevation) map of the study area with comparison of
inventory map-data
Inventory reference data
C
l
a
s
s
i
f
i
e
d

d
a
t
a

Class larch cedar Row Total User Acc.
SVM
larch 10351 4512 14866 0.70
cedar 2584 14818 17402 0.85
Col. Total 12935 19333 32268



Maximum likelihood
larch 11039 6368 17407 0.63
cedar 1693 11163 12856 0.87
Col. Total 12732 17531 30263




94

a b c d
Fig. 7.4 Multi-temporal image composite differences between SVM and Maximum likelihood
classification results before and after applying ancillary biophysical information in dominant fores
type, larch and cedar classes. a and c - SVM and Maximum likelihood classification results; b and d
- SVM and Maximum likelihood classification results adjusted by forests biophysical distribution
characteristics along altitudinal-belt zones. Grassland coral, Larch green, Cedar sea green,
Water blue, Bare soil white, Broadleaf - yellow

7.5 Discussion
7.5.1 Classification accuracy improvement in multi-temporal imagery
The level of discrimination between forest types which can be achieved by using single-date
multispectral satellite data, such as Landsat TM imagery, is generally quite low. As a result, many
land cover surveys which use satellite remote sensing limit themselves to a small number of forest
classes. This low level of discrimination is due to a combination of the limited spectral resolution of
moderate satellite sensors, and the intrinsically small variations that exist between different forest
types spectral characteristics. Researchers overcome this problem by compositing multiple single-
date imageries into single multi-temporal imagery (Potapov et al. 2011). Specifically, imageries
acquired in different seasons have been given better results (Prishchepov et al. 2012, Rodriguez-
Galiano et al. 2012, Wolter and Townsend 2011).
According to our study result, seasonal composite imagery gave better result than single-date
imageries. Summer imagery was enhances the detail of forest distribution areas specifically
coniferous species, larch and cedar. September imageries were enhances the different reflectance of

95

autumn senescence for decisious tree species which are larch and broadleaf. Winter imagery was
enhances the distribution reflectance of evergreen tree species which is cedar.

7.5.2 Land cover mapping and accuracy
The spatial distribution and abundance of forest species within the study area highlight some
obvious and interesting patterns (Fig. 7.4). The results from this comparison suggest that both
classifiers fairly well classify the five classes (grassland, larch, cedar, bare soil and broadleaves).
However, differences were found between two classifiers. That is, SVM was sensitive to spectral
purity of pixels but was insensitive to heterogeneous pixels spectra which is in the case of
deciduous tree species which are larch and broadleaves because the image composites were
consisted by different seasons imageries which use three fall one winter and one summer imageries;
while maximum likelihood classifier is good at classifying general abundant land cover or forest
types which are larch and cedar forests and is poor at classifying less abundant forest type which is
broadleaf forest types. Therefore, we suggest the radial based kernel SVM is better for classification
of forest species type in heterogeneous mountainous area of North-Easthern Mongolia.

7.5.3 Dominant forest type mapping and accuracy
While ancillary information of forest distribution boundary along altitude is very usefull as
separate classes, it is commonplace to attempt to produce a single class that represents dominant
cover types (Moore and Bauer 1990, Schriever and Congalton 1995, Wolter et al. 1995, Reese et al.
2002, Bauer et al. 2009). Many possible strategies may be used to accomplish this (Wolter and
Townsend 2011), but we chose to compile simple dominant forest species by using a method that
ignores species with slightly lower abundance. National forest inventory data is used for assessing
the accuracy of the dominant species maps (Table 7.11). Differences between classified imagery
and inventory map-data could have resulted in disparities when determining all classified land types
or forest species. Due to airborne based polygons, geographic and topographic distortion errors
possibly exist in national forest inventory data and positional accuracy of data locations were not
tested and/or corrected. Also, there were data arrangement concerns over study area which is
complex and multivariate. However, generally, dominant tree species are comparable regardless of
differences (for instance: structure, density and mixture composition etc.) due to their relative
abundance and thus, the 78% of overall accuracy (Kappa 0.55) in the case of SVM and the 74% of
overall accuracy (Kappa 0.48) in the case of Maximum likelihood could be substantially
understated.

96

It should also be noted that, based on biophysical distribution characteristics of forests
altitude-belt, adjustment performed on individual species was an excellent way to reliably
determine presence or absensce of species for a given pixel in a certain elevation. It does provide
information on species co-occurrence or dominance. Thus, alternative approaches to accuracy
assessment such as using fuzzy sets (Gopal and Woodcock 1994, Foody 1995, Townsend 2000)
may be warranted (Table 7.11).

7.6 Conclusion
This research evaluated the potential of seasonal composites of multi-temporal imageries
together with use of ancillary biophysical information for improve the land cover and forest type of
mountainous forest type classification algorithm in North-Eastern Khentii, Mongolia. The result of
this research provides an insight into the performance of SVM in the context of forest type mapping
and forests altitudinal-belt zone restriction from ancillary biophysical information.
The accuracy of classification produced from multi-temporal spectral SVM model
significantly increased with the adjustment of biophysical information. This increase was observed
in comparison of classifications performed by using single and multi-temporal imageries and using
with/without adjustment of ancillary biophysical information. The result further tested with
comparison on national forest inventory data which is selected only for larch and cedar forest type
distribution classes in the study area, because of distorted geographic locations and different
topographies effects. However, relative abundance of larch and cedar species were comparable with
consideration of the problems mentioned above.
Adjusting forest distributions by their characteristical information of altitudinal-belt zone led
to more accurately classify the mixed species in most confused areas, where species mixture
percentage unknown, and species distribution boundary merged together (e.g. increases in accuracy
of larch 33% exchange of reducing accuracy of cedar 6% in case of SVM; increases in accuracy of
larch 31% exchange of reducing accuracy of cedar 4% in case of Maximum likelihood classifier).
Moreover, considering altitudinal-belt zone did not affect the classification of the other classes
which are grassland, bare soil and broadleaf classes.
At last, the comparison result of SVM and Maximum likelihood classifiers shows that the
SVM gives better performance than Maximum likelihood classifier.

97

Part II. Summary
Forest biomass assessment is important for forest management and understanding the role of
forest as source of carbon. The approaches used to quantify the biomass can be laborious and time
consuming. With limited field data, the remote sensing techniques have gotten the reputation in
forest resource assessment over large area in a cost effective manner. This study explored the use of
remote sensing in the estimation of biomass and carbon in forest area of Mongon morit soum,
Mongolia. The overall objective of this study is to combine the limited number of field data and
satellite images to find out the most promotable approach of the forest biomass estimation. The
species-specific equation was applied to calculate stand biomass. The forest biomass in each plot
was then linked with the remotely sensed data derived from Landsat-5 TM and was applied to
generate the general (mixed species) regression equation of forest biomass. This equation was used
to estimate the forest biomass in the satellite image.
Our study revealed that the model which is NDGVI integrated with factor of shortwave-
infrared correction show promising relationship between vegetation index (VI) NDGVIc with
aboveground biomass (AGB) with R
2
=0.70. In addition, there is a significant relationship between
brightness (B) and surface area (SA) VIs and leaf area index (LAI), R
2
=0.83 and 0.81, relatively.
These correlations are found in fall images when vegetation starts to senescence. Due to bud
dormancy period, green light absorption increased and, red reflectance elevated. This seasonal
change could be monitored during fall season with the help of SWIR band water content
information. In another way, this relationship will be found in case of uncorrelated near infrared
(NIR) reflectance with AGB. Otherwise, SWIR corrected normalized difference vegetation index
(NDVIc) is better in significant correlation between NIR and AGB.
This study concludes that NDGVIc can be used as a predictor of the AGB on fall images in
the Taiga forests, as well as B and SA can be used as predictor of LAI. The equations developed in
this study are area-specific and only larch and birch species were used, so at other locations only
after verification should be applied.
Due to limited number of study plots, the reflectance range could not be fully determined in
case of NDGVIc for SWIR correction, min and max values, therefore total AGB of study area
would not calculated. Plus, cedar forests estimates were missing in our regression model for AGB
and LAI.
Appropriate image classification methods for forest province and/or regional level are always
problematic due to different kind of terrain effect and heterogeneous land cover and forest type
mixtures. For a regional level study moderate resolution imageries are suitable for mapping and
detection and the studies often use seasonal composites, multi-temporal imageries. A classification

98

result of multi-temporal image composites brings better accuracy. However, there are always
confusion of certain level of mixed pixel which exist in heterogeneous-mountainous area. In this
case, ancillary biophysical information of forest distribution can be used to separate confused pixels.
Any biophysical information that relate to distribution characteristics of vegetation may have
potential to improve separation of mixed or misclassified pixels. In this study, we particularily used
altitudinal-belt zone distribution characteristics of forests in province level to improve forest type
separation after classifying land cover. This method is applicable in other regions of forested area in
Mongolian mountainous forest areas which are inaccessible for doing forest inventory works.
Therefore, the method is potential to accelerate timeframe of national forest inventory information
update in a shorter span by using seasonal composites of imagery and relative biophysical
information of forests altitudinal-belt zone distribution.
The main contribution of our study is that we demonstrate how to estimate forests volume,
biomass and carbon based on the forests structural estimates measured in ground study. Then the
AGB and the LAI estimates later used to build regression model for finding relationships with VIs.
AGB of Khentii region could be determined using these indexes which are corrected by SWIR
bands. But, due to limited number of study plots we could not calculate total-AGB of study area.
However, the relationships between AGB/LAI and VIs were still existed in other test imageries.
Further we need to collect more ground plots to estimate volumes/AGBs in order to calculate total
AGB for Khentii region. Our developed method is capable to predict regional AGB, once we assess
the accuracy of AGB/LAI regression model and with the forest type map of our image classification
result it is possible to measure species specific AGB/LAI.

99

References
General
Bailey RL, Dell TR (1973) Quantifying diameter distribution with the Weibull-function. For Sci
19(2): 97-104
D'Agostino RB, Stephens MA (1986) Goodness-of-fit Techniques. Marcel Dekker. New York
Danilin IM (1995) Structure and biomass of larch stands regernating naturally after clearcut logging.
Water, Air and Soil Pollution 82: 125-131
Gove J H, Ducey MJ , Leak WB, Zhang L (2008) Rotated sigmoid structures in managed uneven-
aged northern hardwood stands: a look at the Burr Type III distribution. Forestry 81:161-176
Hafley WL, Schreuder HT (1977) Statistical distributions for fitting diameter and height data in
even-aged stands. Can J For Res 7:481-487
J aworski A, Podlaski R (2012) Modelling irregular and multimodal tree diameter distributions by
finite mixture models: an approach to stand structure characteristion. J For Res 17(1):79-88
J ohnson NL (1949) Systems of frequency curves generated by methods of translation. Biometrika
36:149176
Lin C, Chan MH, Chen FS, Wang YN (2007) Age structure and growth pattern of an endangered
species, Amentotaxus formosana Li. J Integr Plant Biol 49(2):157-167
Lindsay SR, Wood GR, Woollons RC (1996a) Stand table modelling through the Weibull
distribution and usage of skewness information. For Ecol Manag 81:1923
Lindsay SR, Wood GR, Woollons RC (1996b) Modelling the diameter distribution of forest stands
using the Burr distribution. J Appl Stat 23:609-619
Tadikamalla PR (1980) A look at the Burr and related distributions. Int Stat Rev 48:337-344
Tsogt K, Tsogt Z, Marnaan B, Dugarjav D, Lin C. (2011a). To describe Birch forest diameter
distribution study by Weibull and Lognormal model in Eastern Khentii. NUM memoir,
biology. Ulaanbaatar, 15:346. p 93-102 (in Mongolian with English summary)
Tsogt K, Narangarav D, Lin C, Tsogt Z (2011b) Larch Stand Structure Analysis of Boreal Forest in
Mongolia. ICEST2011 Singapore. IACSIT Press, 6: 2123-2127
Tsogt K, Lin C, Tsogt Z (2011c) Diameter structure analysis of forest stand and selection of suitable
model. Mongolian J Bio Sci 9(1-2):19-22
Usoltsev VA (2002) Forest biomass of Northern Eurasia: standards and elements of geography.
Botanical garden, Ural branch, RAS, and Ural State Forest Engineering University, Ministry
of Education of Russian Federation, Ekatrinburg (in Russian)

100

Tsogt T (1993) Central Khangai and Eastern Khentii young larch forests dynamic formation,
structure, productivity and forest thinning. Ph.D. dissertation, Mongolian State University of
Agriculture, Ulaanbaatar (in Mongolian)
Zhang L, Liu C (2006) Fitting irregular diameter distributions of forest stands by Weibull, modified
Weibull, and mixture Weibull models. J For Res 11(5):369-372
Zimmer WJ , Burr IW (1963) Variables sampling plans based on non-normal populations. Industrial
Quality Control pp 1836

References - Chapter 1
DeLiocourt F (1898) De lamenagement des Sapiniers. Bull Soc For Franche-Compt et Belfort, pp.
396409
Rodriguez RN (1977) A guide to the Burr type XII distributions. Biometrika 64:129-134
Westphal C, Tremer N von Oheimb G, Hansen J , von Gadow K, Hardtle W (2006) Is the reverse J -
shaped diameter distribution universally applicable in European virgin beech forests? For
Ecol Manag 223:75-83

Assmann E (1961) Waldestragskunde. Munchen, Bonn, Wienn: BLV Verlagsges
Eversole KR (1955) Spacing test in a Douglas-fir plantation. For Sci 1:14-8
Tsogt K, Lin C. (2012) A Flexible Modeling of Irregular Diameter Structure for the Volume
Estimation of Forest Stands. J For Res DOI:10.1007/s10310-012-0380-z
Wang M, Upadhyay A, Zhang L (2010) Trivariate distribution modeling of tree diameter, height,
and volume. For Sci 56:290-300

Dilis NV (1981) The Larch. Lesnaya Promishlenost, Moscow (in Russian)
Dugarjav C (2006) Mongolian Larch Forests. Bembi San, Ulaanbaatar (in Mongolian)
Hafley WL, Schreuder HT (1977) Statistical distributions for fitting diameter and height data in
even-aged stands. Can J For Res 7:481-487
Hoyer GE (1980) Height growth of dominant western hemlock trees that had been released from
understory suppression. Washington state department of natural resources note No. 33, 5 pp.
Kellomaki S (1987) Metsaekologia. Silva Carelica 7. University of J oensuu (in Finnish)

101

Kelty MJ (1992) Comparative productivity of monocultures and mixed-species stands. in: Kelty
MJ , Larson BC, Oliver CD (Eds), The Ecology and Silviculture of Mixed-Species Forests: A
Festchrift for David M. Smith. Kluwer Acadmic Publisher, Boston
Kimmins J P (2004) Forest ecology: A Foundation for sustainable forest management and
environmental ethics in forestry, third ed. Upper Saddle River, New J ersey
Kujala V (1924) Tervaleppa Suomessa - Kasvimaantieteellinen tutkimus. (Alnus glurinosa L.)
Gaertn. in Finland - a plant geographical study. Communicationes Forestalia Fennica Vol. 7
(in Finnish)
Larson BC (1992) Pathways of development in mixed-species stands. in: Kelty MJ , Larson BC,
Oliver CD (Eds), The Ecology and Silviculture of Mixed-Species Forests: A Festschrift for M.
Smith. Kluwer Academic Publishers, Boston
Liu C, Zhang L, Davis CJ , Solomon DS, Gove J H (2002) A finite mixture model for characterizing
the diameter distributions of mixed-species forest stands. For Sci 48:653-661
Lopatin E (2007) Long-term dynamics in growth of Scots pine and Siberian spruce in Komi
Republic (European part of Russia). Dissertation, University of J oensuu, J oensuu
Nelson TC (1964) Diameter distribution and growth of loblolly pine. For Sci 10:105-115
Oliver CD (1976) Growth response of suppressed hemlocks after release, in Atkinson W, Zasoski
RJ (Eds), Western hemlock management, University of Washington Institute of Forest
Products Contribution, No. 34, pp. 266-272
Osawa A, Zyrvanova OA, Matsuura Y, Kajimoto T, Wein RS Permafrost Ecosystems: Siberian
Larch Forests, first ed. Springer, New York
Popov PP (2003) Structure and Differentiation of Spruce Populations in Eastern Europe and
Western Siberia. Russian J Ecol 34:27-33
Sarvas R (1964) Havupuut. Werner Sderstrom Osakeyhti, Helsinki (in Finnish)
Shunzhong W, Limin D, Guohua L, J ianqiong Y, Hengmin Z, Qingli W, (2006) Modeling diameter
distribution of the broadleaved-Korean pine mixed forest on Changbai Mountains of China.
Sci in China 49:177-188
Tsedendash G (2007) Mongolian Spruce Forests. Bull. Institute Botany, MAS 17:125-138 (in
Mongolian)
Tsogt Z (1993) Process of evolving to form young larch forests in Central Khangai and Eastern
Khentii, and to drive this process by science principle. Dissertation, Mongolian State
University of Agriculture, Ulaanbaatar. (in Mongolian)
Veblen TT (1992) Regeneration dynamics. in: Glenn-Lewin DC, Peet RK, Veblen TT (Eds), Plant
succession: theory and prediction. Chapman and Hall, London, pp. 152-187

102

Yoda K, Kira T, Ogawa H, Hozumi K (1963) Self-thinning in overcrowded pure stands under
cultivated and natural conditions. J Biol Osaka City Univ 14:107-129
Zasada M, Cieszewski CJ (2003) A finite mixture distribution approach for characterizing tree
diameter distributions by natural social class in pure even-aged Scots pine stands in Poland.
For Ecol Manag 204:145-158
Zhang L, Gove J H, Liu C, Leak WB (2001) A finite mixture of two Weibull distributions for
modeling the diameter distributions of rotated-sigmoid, uneven-aged stands. Can J For Res
31:1654-1659

Burr IW (1942) Cumulative frequency functions. Ann Math Stat 13: 215-232
Curtis RO (1967) Height-diameter and height-diameter-age equations for second-growth Douglas-
r. For Sci 13:365375
Dorado FC, Diguez-Aranda U, Barrio MA, Rodr guez MS, von Gadow K (2006) A generalized
height and diameter model including random components for radiata pine plantations in
northwestern Spain. For Ecol Manag 229:202213
Fang Z, Bailey RL (1998) Heightdiameter models for tropical forests on Hainan Island in Southern
China. For Ecol Manag 110:315327
Hamilton F, Brack CL (1999) Stand volume estimates from modelling inventory data. Aust For
62:360-367
Huang DS, Price, Titus SJ (2000) Development of ecoregion-based heightdiameter models for
white spruce in boreal forests. For Ecol Manag 129:125141
Larsen DR, Hann DW (1987) Height-diameter equations for seventeen tree species in southwest
Oregon. Research paper 49, Forest research lab, Oregon State University, Corvallis
Podlaski R (2010a) Two-component mixture models for diameter distributions in mixed-species,
two-age cohort stands. For Sci 56:379-390
Podlaski R (2010b) Diversity of patch structure in Central European forests: are tree diameter
distributions in near-natural multilayered Abies-Fagus stands heterogeneous? Ecol Res
25:599-608
Rennolls K. Wang M (2005) A new parameterization of J ohnson's SB distribution with application
to fitting forest tree diameter data. Can J For Res 35:575-579
Rodriguez RN (1977) A guide to the Burr type XII distributions. Biometrika 64:129-134
Semechkin IW, Tetenkin AE (1980) Northern Mongolian forest structure characteristic. Mongolian
For Bull V 12, Science, Moskov, pp 66-116 (in Russian)

103

Tsogt Z, Danilin IM, Naurzbaev MM (2009) Course of the growth of larch forests in Mongolia.
Ecological-geographical aspects of the forest process: All Russian conference with the
participation of foreign scientists. Krasnoyarsk, 23-25 September 2009. V.N.Sukachev
Institute of Forest of SB RAS. p173-178 (in Russian)
Wiant HV, J r, Wood GB, Miles J A (1989) Estimating the volume of a radiata pine stand using
importance sampling. Aust For 52:286-292
Wykoff WR, Crookston NL (1982) Stage AR. Users guide to the stand prognosis model. USDA
For Serv Gen Tech Rep INT-133

Dorjsuren C, Tsogt Z, Tsogt K (2003) Growth dynamic of young larch forests Central Khangai. In
Forest Utilization and Rehabilitation, Scientific Bulletin of Mongolian University of Science
and Technology, Ulaanbaatar, ~1/52 pp 147-155 (in Mongolian)
Dorjsuren C, Tsogt Z (2005) Biomass dynamic growth of larch forests in Central Khangai.
Scientific Bulletin of Mongolian University of Science and Technology, Ulaanbaatar ~3/74
pp 116-119 (in Mongolian)
Dorjsuren C, Tsogt Z, Tsogt K (2007) Eastern Khentii larch forests biomass table. Scientific
Bulletin of Institute of Botany, MAS, Ulaanbaatar ~17 pp 94-96 (in Mongolian)
Dugarjav C (2006) Larch forests in Mongolia: Origin, current state and regeneration. Institute of
Botany, MAS, Bembi San, Ulaanbaatar (in Mongolian)
Garzuglia M, Saket M (2003) Wood volume and woody biomass: review of FRA 2000 estimates.
Forestry Department, Food and Agriculture Organization of the United Nations, Forestry
Resources Assessment Programme, Rome, July 2003. Working paper 68
Krasnoschekov YN (1982) The detention of rainfall canopy larch stands in the Central Khangai. In
the ecological role of forest ecosystems in Siberia. Institute of Forest and Wood, Siberian
branch, Academy of Sciences, USSR, Krasnoyarsk, pp 40-46 (in Russian)
Savin EN, Dugarjav C (1983) Silvicultural activities in clearings in Larch. In Forests of the
Mongolian People's Republic: Larch forests of the Central Khangai. Nauka, Novosibirsk, pp
81-102 (in Russian)
Tsolmon R (2003) Methodology to estimate coverage and biomass of Boreal forests using satellite
data. Ph.D. dissertation, Graduate School of Science and Technology Chiba University
Usoltsev VA (2001) Forest biomass of Northern Eurasia: database and geography. Botanical garden,
Ural branch, RAS, and Ural State Forest Engineering University, Ministry of Education of
Russian Federation, Ekatrinburg (in Russian)

104

Tsogt Z, Dorjsuren C, Tsogt K (2006) Young larch forests growth table of Eastern Khentii.
Scientific Bulletin of Institute of Botany, MAS ~16 pp 114-122 (in Mongolian)

Aldred AH, Alemdag IS (1988) Guidelines for forest biomass inventory. Petawawa National
Forestry Inst. Rep. PI-X-77. Canadian Forest Service, National Resources Canada, Ottawa
Ariunzul YY (2008)
(Using satellite data and by remote sensing
technique to assess forests of Mongolian forest-steppe zone). Ph.D. dissertation, National
University of Mongolia, Ulaanbaatar (in Mongolian)
Birth GS, McVey G (1968) Measuring the color of growing turf with a reflectance
spectroradiometer. Agron J 60:640-643
Campbell J B (2006) Introduction to remote sensing 4
th
ed. Taylor and Francis, London
Chen J M, Pavlic G, Brown L, Cihlar J , Leblanc SG, White HP, Hall RJ , Peddle DR, King DJ ,
Trofymow J A, Swift E, Van der Sanden J , Pellikka PK (2002) Derivation and validation of
Canada-wide coarse-resolution leaf area index maps using high-resolution satellite imagery
and ground measurements. Remote Sens Environ 80:165-184
Cochrane MA (2000) Using Vegetation Reflectance Variability for Species Level Classification of
Hyperspectral Data. Int J Remote Sens 21:2075-2087
Cohen WB, Maiersperger TK, Gower ST and Turnr DP (2003) An improved strategy for regression
of biophysical variables and Landsat ETM+data. Remote Sens Environ 80:561-571
Crist EP, Kauth RJ (1986) The tasseled cap de-mystified. Photogramm Eng Remote Sensing
52(1):81-86
Dahlberg U (2001) Quantification and classification of Scandinavian mountain vegetation based on
field data optical satellite images. Lientiate thesis, Swedish University of Agricultural
Sciences, Deperatment of Forest Resource Management and Geomatics, Report 12
Draper NR, Smith H (1966) Applied regression analysis. Wiley, New York
Gausman HW, Allen WA, Cardenas R (1969) Reflectanceof Cotton Leaves and their Structure.
Remote Sens Environ 1:110-122
Gausman HW, Escobar DE, Everitt J H, Richardson AJ , Rodriguez RR (1978) Distinguishing
succulent plants from crop and woody plants. Photogramm Eng Remote Sens 44:487-491
Gitelson AA, Kaufman YJ , Strak R, Rundquist D (2002) Novel algorithms for remote estimation of
vegetation fraction. Remote Sens Environ 80:76-87

105

Gower ST, Norman J M (1991) Rapid estimation of leaf area index in conifer and broad-leaf
plantations. Ecology 72(5):1896-1900
Heiskanen J (2006) Estimating tree biomass and leaf area index in a mountain birch forest using
ASTER satellite data. Int J Remote Sens 27(6):1135-1158
J ensen J R (2005) Introductory digital image processing: A remote sensing perspective. Pearson
Education, USA
Koriba K (1958) On the periodicity of tree-growth in the tropics, with special reference to the mode
of branching, the leaf fall, and the formation of the resting bud. Gardens Bulletin, Singapore,
Seeries 3, 17:11-81
Nemani RR, Pierce LL, Running SW (1993) Forest ecosystem processes at the watershed scale:
sensitivity to remotely-sensed leaf area index estimates. Int J Remote Sens 14(13):2519-2534
Qi J , Cabot F, Moran MS, Dedieu G (1995) Biophysical parameter estimations using
multidirectional spectral measurements. Remote Sens Environ 54:71-83
Richter R (2000a) Value adding products derived from the ATCOR Models: DLR-IB 564-01/00,
DLR
Richter R (2000b) Atmospheric correction algorithm for flat terrain: Model ATCOR2, DLR-IB
564-02/00, DLR
Rock BN, Williams DL, Vogelmann J E (1985) In machine processing of remotely sensed data
symposium. Purdue Universit, West Lafayette IN pp.71-81
Satterwhite MB (1984) Discriminating vegetation and soils using Landsat MSS and thematic
mapper bands and band ratios. Proceedings, American Society for Photogrammetry & Remote
Sensing (ASPRTS) 2:479-485
Schmidt KS, Skidmore AK (2001) Exploring spectral discrimination of grass species in African
rangelands. Int J Remote Sens 22:3421-3434
Schmidt KS, Skidmore AK (2003) Spectral discrimination of vegetation types in a coastal wetland.
Thenkabail PS (2004) Inter-sensor relationships between Ikonos and Landsat 7 ETM+NDVI data
in three ecoregions of Africa. Center for Earth Observation, Yale University, School of
Forestry and Environmental Studies, Yale University, New Haven, Connecticut
Thomas J R, Namken LN, Oerther GF, Brown RG (1971) Estimating leaf water content by
reflectance measurements. Agron J 63:845-847
Tsogt K, Yi-Chun C, Lin C (2008) Hyperspectral characteristics and seasonal changes of red
cypress (Chamaecyparis formosensis). In Proceedings of National Forestry Conference-2008.
National Chiayi University, Chiayi, October 28-29. V2 pp 275-285

106

Tucker CJ (1979a) Red and photographic infrared linear combinations for monitoring vegetation.
Tucker CJ (1979b) Remote sensing of leaf water content in the near infrared. NASA Technichal
memorandum 80291 Greenbelt MD: Goddard Space Flight Center, 17 pp
Tucker CJ (1980) Remote Sensing of leaf water content in the near infrared. Remote Sens Environ
10:23-32
Tucker CJ and Seller PJ (1986) Satellite remote sensing of primary production. Int J Remote Sens
7:1395-1416

Bauer M, Loeffelholz B, Shinneman D. (2009) Border Lakes land-cover classicationMetadata,
description of land-cover classication methods, foldout map, and digital versions of
classication. U.S.D.A Forest Service, Northern Research Station, RMAP NRS-1 14 pp.
Bradshaw CJ A, Warkentin IG, Sodhi NS (2009) Urgent preservation of boreal carbon stocks and
biodiversity. Trends Ecol Evol 24:541-546
Burges CJ C (1998) A tutorial on support vector machines for pattern recognition. Data Min Knowl
Disc 2:121167
Campbell J B (2006) Introduction to remote sensing, 4
th
ed.Taylor & Francis, London
Camps-Valls G, Gomez-Chova L, Munos-Mar J , Vila-Francs J , Calpe-Maravilla J (2006)
Composite kernels for hyperspectral image classification. Geosci Rem Sens Lett IEEE 3:93
97
Chen J , Saunders SC, Crow TR, Naiman RJ , Brosofske KD, Mroz GD, Brookshire BL, Franklin J F
(1999) Microclimate in forest ecosystem and landscape ecology. Bioscience 49:288297
Cogbill CV, White PS (1991) The latitude-elevation relationship for spruce-r forest line along the
Appalachian mountain chain. Vegetatio 94:153175
Cohen J (1960) A coefficient of agreement for nominal scales. Educational Psychological
Measurement 20:3746
Congalton RG, Green K (2009) Assessing the Accuracy of Remotely Sensed Data: Principles and
Practices. CRC Press, Boca Raton, FL
Dobrowski SZ, Abatzoglou J T, Greenberg J A, Schladow SG (2009) How much influence does
landscape-scale physiography have on air temperature in a mountain environment?
Agricultural and Forest Meteorology 149:1751-1758
Drake J M, Randin C, Guisan A (2006) Modelling ecological niches with support vector machines. J
Appl Ecol 43:424432

107

FAO Food and Agriculture Organization (2006) Global forest resources assessment 2005. UNFAO,
Rome
Foody GM (1995) Cross-entropy for the evaluation of the accuracy of a fuzzy land cover
classication with fuzzy ground data. ISPRS Journal of Remote Sensing 17:212
Foody GM, Mathur A (2004) A relative evaluation of multiclass image classification by support
vector machines. T Geosci Remote IEEE 42:13351343
Foody GM, Mathur A (2006) The use of small training sets containing mixed pixels for accurate
hard image classification: training on mixed spectral responses for classification by a SVM.
Remote Sens Environ 103:179189
Forest law (2012) Mongolian law about forest; in State information, Mongolian state congressional
official statement. Imprint department of state congress, Ulaanbaatar 22
Gopal S, Woodcock CE (1994) Theory and methods for accuracy assessment of Thematic Maps
using fuzzy sets. Photogrammetric Engineering and Remote Sensing 60:181188
Gunin PD, Vostokova EA, Baja SN, Bayasgalan A, Drobishev YI, Dorofeyuk NI, Dugarjav C,
Prishchepa AV (2005) Ecosystems of Selenge River Basin. Nauka (in Russian)
Guo Q, Kelly M, Graham CH (2005) Support vector machines for predicting distribution of Sudden
Oak Death in California. Ecol Model 182:7590
Hill MJ, Ticehurst CJ , Lee J-S, Grunes MR, Donald GE, Henry D (2005) Integration of optical and
radar classifications for mapping pasture type in Western Australia. T Geosci Remote IEEE
45:16651681
Hsu CW, Chang CC, Lin CJ (2009) A practical guide to support vector classification. Technical
Note, Department of Computer Science and Information Engineering, National Taiwan
University, Taiwan
IPCC The Intergovermental Panel on Climate Change (2000) Land use, land-use change, and
forestry. Cambridge University Press, UK
Kasischke ES, Stock BJ (2000) Fire, climate change and carbon cycling in the boreal forest.
Springer-Verlag, New York
Kurz WA, Beukema SJ , Apps MJ (1998) Carbon budget implication of the transition from natural
to managed disturbance regimes in forest landscapes. Mitig Adapt Strat Gl 2:241-405
Lardeux C, Frison P.-L, Tison C, Souyris J .-C, Stoll B, Fruneau B, Rudant J .-P (2009) Support
vector machine for multifrequency SAR polarimetric data classification. T Geosci Remote
IEEE 47:41434152

108

Lennon M, Mercier G, Hubert-Moy L (2002) Classification of hyperspectral images with nonlinear
filtering and support vector machines. IEEE International Geoscience and Remote Sensing
Symposium, 16701672
Luyssaert S, Schulze ED, Borner A, Knohl A, Hessenmoller D, Law BE, Ciais P, Grace G (2008)
Old-growth forests as global carbon sinks. Nature 455:213-215
Malyshev L (1993) Levels of the upper forest boundary in Northern Asia. Vegetatio 109:175186
Moore MM, Bauer ME (1990) Classication of forest vegetation in north-central Minnesota using
Landsat Multispectral Scanner and Thematic Mapper data. Forest Science 36(2):330342
Nagy L, Grabherr G, Krner C, Thompson DBA (Eds.) (2003) Alpine Biodiversity in Europe.
Springer Verlag, Berlin
Nemmour H, Chibani Y (2006) Multiple support vector machines for land cover change detection:
an application for mapping urban extensions. ISPRS J Photogramm 61:125133
Otgonsuren B, Khuvisgalt B, Amgalan B, Ser-Od B, Erdenekhuu E (2007) Mongon-morit soum,
Tov aimag; Forest management project. National forest and water inventory center,
Ulaanbaatar
Pal M, Mather PM (2004) Assessment of the effectiveness of support vector machines for
hyperspectral data. Future Generation Computer Systems 20:12151225
Pimm S Roulet N, Weaver A (2009) Boreal forests carbon stores need better management. Nature
462:267
Potapove P, Hansen MC, Stehman SV, Loveland TR, Pittman K (2008) Combining MODIS and
Landsat imagery to estimate and map boreal forest cover loss. Remote Sens Environ
112(9):3708-3719
Potapov P, Turubanova S, Hansen M (2011) Regional-scale boreal forest cover and change
mapping using Landsat data composites for European Russia. Remote Sens Environ 115:548-
561
Prishchepov AV, Radeloff VC, Dubinin M, Alcantara C (2012) The effect of Landsat ETM/TM+
image acquisition dates on the detection of agricultural land abandonment in Eastern Europe.
Reese HM, Lillesand T, Nagel DE, Stewart J S, Goldmann RA, Simmons TE, et al. (2002)
Statewide land cover derived from multiseasonal Landstat TM dataA retrospective of the
WISCLAND project. Remote Sensing of Environment 82:224237
Richards J A (1999) Remote sensing digital image analysis. Springer-Verlag, Berlin
Richter R (2000a) Atmospheric correction algorithm for flat terrain: Model ATCOR2, DLR-IB 564-
02/00, DLR

109

Rodriguez-Galiano VF, Chica-Olmo M, Abarca-Hernandez F, Atkinson PM, J eganathan C (2012)
Random Forest classification of Mediterranean land cover using multi-seasonal imagery and
multi-seasonal texture. Remote Sens Environ 121:93-107
Schlkopf B, Smola A (2002) Learning with Kernels. Cambridge, MA, MIT Press
Schriever J R, Congalton RG (1995) Evaluating seasonal variability as an aid to cover-type mapping
from Landsat Thematic Mapper in the Northeast. Photogrammetric Engineering and Remote
Sensing 61(3):321327
Smagin VN (1965) Lesa basseina r Ussuri (Forests of Ussuri river basin). Nauka, Moscow
Townsend PA (2000) A quantitative fuzzy approach to assess mapped vegetation classications for
ecological applications. Remote Sensing of Environment 72:253267
Troll C (1973) The upper timberlines in different climatic zones. Arctic and Alpine Research 5:3-18
Vapnik V, Chervonenkis A (1998) Statistical learning theory. In: Support Vector Machines for
Pattern Recognition. J ohn Wiley & Sons, New York
Vostokova et al. (1995) Ecosystems of Mongolia: Distribution and current state. In Biological
resources and natural conditions of Peoples Republic of Mongolia #39 Nauka Publishing
House (in Russian)
Yaroshenko AY, Komarova AF, Kreindlin ML (2009) The second anniversary of the new forest
code of Russian Federation. Greenpeace Russia, Moscow
Waske B, Benediktsson J A (2007) Fusion of support vector machines for classification of
multisensor data. T Geosci Remote IEEE 45:38583866
Wolter PT, Mladenoff DJ , Host GE, Crow TR (1995) Improved forestclassication in the northern
Lake State using multi-temporal Landsat imagery. Photogrammetric Engineering and Remote
Sensing 61(9):11291143
Wolter PT, Townsend PA (2011) Multi-sensor data fusion for estimating forest species composition
and abundance in northern Minnesota. Remote Sensing of Environment 115:671-691

110

Appendixes
Appendix 1
Burr distribution
The Burr distribution was introduced to the forestry research by Lindsay et al. (1996). This
distribution is inherently more flexible, because it covers a much larger area of the skewness-kurtosis
plane than the Weibull distribution (Lindsay et al., 1996; Rodriguez, 1977; Tadikamalla, 1980).
The Burr (Zimmer & Burr, 1963) distribution has a flexible shape, controllable scale and location,
which makes it appealing to fit to data. It is sometimes considered as an alternative to a Normal
distribution when data show slight positive skewness. With the support random variable x: + < s x ,
the pdf of Burr 4 parameter distribution is given as:
1
1
1
) (
+
|
|
.
|
\
|
|
|
.
|
\
|
+
|
|
.
|
\
|
=
k
x
x
k
x f
o
o
|
|
|
o

where, k, > 0 - two shape parameters, > 0 - scale parameter, - location parameter if =0, then the
distribution is Burr 3 parameter.

Dagum distribution
Dagumdistribution (Dagum 1977) is inverse Burr distribution (Klugman et al 1998). It is also
known kappa distribution (Mielke 1973; Mielke and Johnson 1973), and extended Burr-III distribution
(Shao et al 2008). The Dagum distribution has a wider region of applicability than either the Burr XII or
the Weibull distribution. With the support random variable x: + < s x , the pdf (probability density
function) of Dagum 4 parameter distribution is given as:
1
1
1
) (
+
|
|
.
|
\
|
|
|
.
|
\
|
+
|
|
.
|
\
|
=
k
k
x
x
k
x f
o
o
|
|
|
o

where, k, > 0 are the two shape parameters, > 0 is the scale parameter, is the location parameter if
=0, then the distribution is 3 parameter.

J ohnson S
B
distribution
The Johnson S
B
(1949) have been much commonly used in forest distributional studies (Hafley &
Schreuder, 1977), because of its flexibility of distributional form and its ability to represent equally well
positive and negative skewed distributions. The pdf of S
B
distribution transforms a bounded random
variable by subtracting the minimum and dividing by the range. The logit of this transformation is then

111

distributed as a standard normal variable. Following J ohnson, consider this transform z on the random
variable x:

=
x
z
where, - minimum value of x, + - maximum value of x.
+ s s x
Within our context x is a D measurement. Then the function of D is defined as
( )
|
|
.
|
\
|
|
|
.
|
\
|
|
.
|
\
|
=
2
1
ln
2
1
exp
1 2
) (
z
z
z z
x f o
t
o

where, and shape ( > 0), scale ( > 0) and location parameter.

Lognormal distribution
Lognormal distribution is right skewed distribution and it is easy to work with D distributions
which exhibit positive skew or symmetry. With the support random variable x: + < x ,
the PDF of Lognormal 3 parameter distribution is given as:
t o
o

2 ) (
ln
2
1
exp
) (
2
|
|
.
|
\
|
|
.
|
\
|
=
x
x
x f
where, and continuous parameters (>0), and -continuous location parameter if =0, then the
distribution is 2 parameter
.

Weibull distribution
One of the most popular models is the Weibull distribution, first introduced to the forestry
research field by Bailey and Dell (1973). The popularity of the Weibull distribution depends largely on
its simplicity and yet relatively good flexibility. It describes the inverse J shape for <1 and the
exponential distribution for =1. For 1< <3.6 the density function is mound shaped and positively
skewed and for =3.6 the density function becomes approximately normal. If >3.6 the density
function becomes increasingly negatively skewed. With the support randomvariable x: + < s x , the
pdf of Weibull 3 parameter distribution is given as:
|
|
.
|
\
|
|
|
.
|
\
|
|
|
.
|
\
|
=
o o
|

|

|
o x x
x f exp ) (
1

where, > 0 - shape parameter, > 0 - scale parameter, - location parameter if =0, then the
distribution is 2 parameter.

112

Appendix 2
Allometric equations coefficients/characteristics for larch Larix sibirica, birch - Betula platyphylla, aspen - Populus laurifolia and willow Salix tree
species
Fraction a0 ax lnA (lnA)
2
lnD lnH lnN PF/m lnPS lnM R
2
SE
larch Larix sibirica
Ln(P
S
) -0.8029 0.1861

-0.0133

0.9963 0.995 0.089
Ln(P
F
/M) -0.3909 -0.0741 -0.1982

-0.9551

-0.1245

0.716 0.484
Ln(PB/M) 2.7099 0.1313 -1.4962 0.1526 0.5155 -0.7234 -0.0638 0.3001

0.810 0.286
birch - Betula platyphylla
Ln(P
S
) -0.6852 0.0106

0.0401

0.9764 0.983 0.128
Ln(P
F
/M) -0.5701 -0.3585 -0.9632 0.1155 0.249

0.792 0.325
Ln(PB/M) 0.9469 0.4192 -0.673 0.0817

-0.7234 -0.1531 0.417

0.711 0.212
aspen - Populus laurifolia
Ln(P
S
) -1.1685 0.1187 0.1715

0.0448

0.9389 0.989 0.124
Ln(P
F
/M) -0.1184 -0.4939 -0.2337 0.1155

-0.9939

0.904 0.271
Ln(P
B
/M) -0.2685 -0.0984 0.3958 0.0817 0.2625 -0.5132 -0.1565 0.8042

0.829 0.275
willow Salix
Ln(P
S
) -0.6874 0.0092

0.9688 0.977 0.156
Ln(P
F
/M) -1.7252 -0.0254

-1.1109

0.798 0.391
Ln(P
B
/M) -1.3775 0.2852

0.9199 -1.2149

0.2105

0.830 0.219

113

Appendix 3
PVI perpendicular vegetation index (VI),
B brightness,
G greenness,
W wetness,
TVI triangular VI,
NDVI normalized difference VI,
II infrared index,
GRABS greenness above bare soil,
MSI moisture stress index,
MidIR shortwave-infrared index,
ARVI atmospherically resistant VI,
NDBI normalized difference built up index,
EVI enhanced VI,
SAVI soil adjusted VI,
SARVI soil and atmospherically resistant VI,
AFRI aerosol free VI,
OSAVI optimized soil adjusted VI,
ND
SWIR
I (SWIR-R)/(SWIR+R), normalized difference SWIR index
GNDVI - green normalized difference VI,
NLI nonlinear index,
SR simple ratio,
TCARI transformed chlorophyll absorption in reflectance index,
VNIRV Near infrared,
IPVI infrared percentage VI,
NDWI normalized difference water index,
WDRVI wide dynamic range VI,
MTVI1 modified triangular VI,
MSR modified simple ratio,
RDVI renormalized difference VI,
NDGVIc SWIR corrected green and red normalized difference VI,
MND modified normalized difference,
NDVIc SWIR corrected normalized difference VI,
DVI difference VI,
RSR reduced simple ratio,
SA surface albedo

114

Appendix 4
Correlations among the Landsat TM bands, Sep 07 2009
b1 b2 b3 b4 b5
b2 0.202
0.488
b3 0.489 0.877
0.076 0.000
b4 -0.130 0.404 0.318
0.657 0.152 0.267
b5 0.647 0.643 0.898 0.091
0.012 0.013 0.000 0.758
b7 0.663 0.626 0.875 -0.069 0.973
0.010 0.017 0.000 0.816 0.000
P-Value

115

Appendix 5

Relationship between stem biomass and Landsat TM bands 3, 5 and 7. Correlation coefficients (r)
are given for bands.

-5
0
5
10
15
20
25
30
0 5 10 15 20 25
S
t
e
m

b
i
o
m
a
s
s

(
t
/
h
a
)
Reflectance (%)
Band 3, r =-0.649 Band 5, r =-0.648 Band 7, r =-0.549

116

Appendix 6
Correlations of the biomass components of forest stand and leaf area index (LAI)
LAI AGB N M Stem Bark Foliage Branch Root
AGB 0.754

0.002

N -0.333 -0.252

0.245 0.385

M 0.779 0.996 -0.255

0.001 0.000 0.379

Stem 0.773 0.998 -0.264 1.000

0.001 0.000 0.362 0.000

Bark 0.805 0.984 -0.202 0.993 0.990

0.001 0.000 0.489 0.000 0.000

Foliage -0.067 0.222 -0.142 0.141 0.157 0.088

0.820 0.446 0.628 0.632 0.591 0.765

Branch 0.605 0.936 -0.016 0.910 0.915 0.900 0.377

0.022 0.000 0.956 0.000 0.000 0.000 0.184

Root -0.322 -0.051 -0.166 -0.134 -0.115 -0.201 0.946 0.108

0.262 0.861 0.571 0.648 0.696 0.492 0.000 0.714

Grass -0.347 -0.170 0.871 -0.175 -0.179 -0.148 -0.224 0.077 -0.205

0.225 0.560 0.000 0.549 0.541 0.613 0.442 0.792 0.481

P-Value

117

Appendix 7
Correlations between band ratios and AGB and/or LAI
band ratio LAI AGB band ratio LAI AGB band ratio LAI AGB
b1/b2 -0.055 -0.322 b3/5 -0.741 -0.344 b1/2/5 0.152 -0.056
0.851 0.261 0.002 0.228 0.603 0.85
b1/b3 0.12 -0.116 b3/6 -0.777 -0.655 b1/2/6 0.074 -0.243
0.683 0.693 0.001 0.011 0.801 0.403
b1/4 -0.104 -0.337 b3/7 -0.679 -0.329 b1/2/7 0.145 -0.043
0.724 0.238 0.008 0.251 0.62 0.883
b1/5 -0.077 -0.19 b4/5 0.158 0.349 b2/3/4 0.837 0.756
0.795 0.516 0.589 0.221 0 0.002
b1/6 -0.152 -0.371 b4/6 -0.172 -0.083 b2/3/5 0.703 0.797
0.605 0.192 0.556 0.779 0.005 0.001
b1/7 -0.076 -0.174 b4/7 0.145 0.278 b2/3/6 0.73 0.714
0.798 0.552 0.62 0.336 0.003 0.004
b2/3 0.684 0.774 b5/6 -0.418 -0.676 b2/3/7 0.643 0.709
0.007 0.001 0.137 0.008 0.013 0.005
b2/4 -0.195 -0.074 b5/7 0.072 0.016 b3/4/5 -0.026 0.116
0.504 0.802 0.807 0.958 0.928 0.694
b2/5 0.001 0.419 b6/7 0.385 0.534 b3/4/6 -0.275 -0.355
0.997 0.136 0.174 0.049 0.342 0.213
b2/6 -0.451 -0.192 b3/2 -0.654 -0.744 b3/4/7 0.003 0.159
0.105 0.512 0.011 0.002 0.993 0.587
b2/7 0.038 0.359 b1/b2/3 0.289 -0.011 b4/5/6 0.316 0.421
0.898 0.208 0.317 0.971 0.272 0.134
b3/4 -0.491 -0.458 b1/2/4 0.099 -0.238 b4/5/7 0.387 0.512
0.075 0.1 0.736 0.413 0.172 0.061
P-Value

118

Appendix 8
Correlations of the aboveground biomass of forest stand and leaf area index against the VIs.

VI LAI AGB VI LAI AGB VI LAI AGB
PVI -0.543 -0.354 EVI 0.499 0.409 IPVI 0.489 0.461
0.045 0.215 0.070 0.147 0.076 0.097
B -0.837 -0.710 SAVI 0.485 0.458 NDWI 0.142 0.338
0.000 0.004 0.079 0.100 0.629 0.237
G -0.234 -0.092 SARVI 0.513 0.445 WDRVI 0.485 0.463
0.421 0.753 0.060 0.111 0.079 0.095
W 0.276 0.426 AFRI5 0.139 0.337 MTVI1 -0.078 0.068
0.339 0.128 0.635 0.239 0.791 0.817
TVI 0.096 0.230 AFRI7 0.126 0.278 MSR 0.483 0.462
0.745 0.428 0.668 0.336 0.080 0.096
NDVI 0.489 0.461 OSAVI 0.488 0.460 RDVI 0.484 0.462
0.076 0.097 0.077 0.098 0.080 0.096
II 0.142 0.338 SWIRV2 0.675 0.316 NDGVIc 0.694 0.803
0.629 0.237 0.008 0.271 0.006 0.001
GRABS -0.116 0.009 SWIRV1 0.741 0.337 MND -0.590 -0.376
0.694 0.975 0.002 0.238 0.026 0.185
MSI -0.134 -0.334 GNDVI 0.191 0.075 TCARI/OSAVI -0.335 -0.371
0.648 0.244 0.512 0.798 0.242 0.192
MidIR 0.072 0.016 NLI 0.093 0.190 NDVIc 0.609 0.649
0.807 0.958 0.751 0.516 0.021 0.012
ARVI 0.517 0.441 SR 0.476 0.458 DVI -0.265 -0.115
0.058 0.114 0.086 0.100 0.360 0.696
VARI 0.643 0.750 TCARI -0.358 -0.381 RSR 0.589 0.621
0.013 0.002 0.209 0.178 0.027 0.018
NDBI -0.142 -0.338 VNIR1 0.125 0.307 SA -0.747 -0.640
0.629 0.237 0.671 0.286 0.002 0.014
built up area -0.290 -0.399 VNIR2 0.191 0.075
0.315 0.158 0.512 0.798
Cell Contents: Pearson correlation and P-Value

119

Appendix 9
Regression equation of AGB by Brightness

AGB =y0+a*log(B)

R Rsqr Adj Rsqr Standard Error of Estimate

0.7348 0.5399 0.5015 6.2256

Coefficient Std. Error t P VIF

y0 159.1833 39.6887 4.0108 0.0017 568.9747<
a -111.3617 29.6789 -3.7522 0.0028 568.9747<

Analysis of Variance:

Uncorrected for the mean of the observations:
DF SS MS
Regression 2 2058.0084 1029.0042
Residual 12 465.1041 38.7587
Total 14 2523.1125 180.2223

Corrected for the mean of the observations:
DF SS MS F P
Regression 1 545.6897 545.6897 14.0792 0.0028
Residual 12 465.1041 38.7587
Total 13 1010.7938 77.7534

Statistical Tests:

PRESS 687.5729

Durbin-Watson Statistic 1.9417 Passed
Normality Test Passed (P =0.8608)
K-S Statistic =0.1549 Significance Level =0.8608
Constant Variance Test Passed (P =0.2006)
Power of performed test with alpha =0.0500: 0.8758

120

Appendix 10
Regression equation of AGB by Surface albedo

AGB =y0+a*log(SA)


0.6909 0.4773 0.4338 6.6353


y0 199.4698 57.1439 3.4907 0.0045 1038.3640<
a -91.7518 27.7165 -3.3104 0.0062 1038.3640<


DF SS MS
Regression 2 1994.7903 997.3951
Residual 12 528.3222 44.0269
Total 14 2523.1125 180.2223

DF SS MS F P
Regression 1 482.4716 482.4716 10.9586 0.0062
Residual 12 528.3222 44.0269
Total 13 1010.7938 77.7534

Statistical Tests:

PRESS 1092.8504


121

Appendix 11
Regression equation of AGB by VARI

AGB =y0+a*VARI


0.7502 0.5628 0.5264 6.0682


y0 10.3865 1.6218 6.4042 <0.0001 1.0000
a 84.0871 21.3929 3.9306 0.0020 1.0000


DF SS MS
Regression 2 2081.2260 1040.6130
Residual 12 441.8833 36.8236
Total 14 2523.1093 180.2221

DF SS MS F P
Regression 1 568.9109 568.9109 15.4496 0.0020
Residual 12 441.8833 36.8236
Total 13 1010.7942 77.7534

Statistical Tests:

PRESS 617.9288

Durbin-Watson Statistic 1.3851 Failed

122

Appendix 12
Regression equation of AGB by NDVIc

AGB=exp(a*NDVIc)


0.6989 0.4885 0.4885 6.3063


a 4.3349 0.2205 19.6569 <0.0001 1.0000


DF SS MS
Regression 1 2006.1148 2006.1148
Residual 13 516.9977 39.7691
Total 14 2523.1125 180.2223

DF SS MS F P
Regression 0 493.7961 (+inf) (+inf) (NAN)
Residual 13 516.9977 39.7691
Total 13 1010.7938 77.7534

Statistical Tests:

PRESS 692.4377


123

Appendix 13
Regression equation of AGB by NDGVIc

AGB =a/(1+b*NDGVIc)


0.8412 0.7077 0.6833 4.9621


a 3.6267 0.8872 4.0879 0.0015 5.4139<
b -3.0981 0.1525 -20.3131 <0.0001 5.4139<


DF SS MS
Regression 2 2227.6440 1113.8220
Residual 12 295.4685 24.6224
Total 14 2523.1125 180.2223

DF SS MS F P
Regression 1 715.3253 715.3253 29.0518 0.0002
Residual 12 295.4685 24.6224
Total 13 1010.7938 77.7534

Statistical Tests:

PRESS 461.5185


124

Appendix 14
Regression equation of LAI by Brightness

Exp(LAI)&=a/(1+b*log(B))


0.9112 0.8302 0.8161 5.4892


a -1.3711 0.2224 -6.1659 <0.0001 6.1540<
b -0.8183 0.0047 -173.9356 <0.0001 6.1540<


DF SS MS
Regression 2 7049.7500 3524.8750
Residual 12 361.5776 30.1315
Total 14 7411.3275 529.3805

DF SS MS F P
Regression 1 1768.2789 1768.2789 58.6855 <0.0001
Residual 12 361.5776 30.1315
Total 13 2129.8565 163.8351

Statistical Tests:

PRESS 653.7767


125

Appendix 15
Regression equation of LAI by Surface albedo

exp(LAI)=a/(1+b*log(SA))


0.9097 0.8275 0.8132 5.5327


a -0.9390 0.1455 -6.4530 <0.0001 5.5261<
b -0.5157 0.0019 -266.4323 <0.0001 5.5261<


DF SS MS
Regression 2 7043.9993 3521.9997
Residual 12 367.3282 30.6107
Total 14 7411.3275 529.3805

DF SS MS F P
Regression 1 1762.5282 1762.5282 57.5789 <0.0001
Residual 12 367.3282 30.6107
Total 13 2129.8565 163.8351

Statistical Tests:

PRESS 430.4392


126

Appendix 16
Regression equation of LAI by Visible atmospheric resistant index (VARI)

Log(LAI)=a/(1+b*VARI)


0.6907 0.4770 0.4334 0.0827


a 0.4157 0.0241 17.2199 <0.0001 1.3100
b -2.4598 0.7520 -3.2709 0.0067 1.3100


DF SS MS
Regression 2 2.6599 1.3299
Residual 12 0.0822 0.0068
Total 14 2.7420 0.1959

DF SS MS F P
Regression 1 0.0749 0.0749 10.9454 0.0062
Residual 12 0.0822 0.0068
Total 13 0.1571 0.0121

Statistical Tests:

PRESS 0.1177


127

Appendix 17
Regression equation of LAI by NDVIc

Log(LAI) =y0+a*exp(NDVIc)


0.6357 0.4041 0.3545 0.0883


y0 0.0400 0.1386 0.2883 0.7780 34.4889<
a 0.2394 0.0839 2.8529 0.0145 34.4889<


DF SS MS
Regression 2 2.6484 1.3242
Residual 12 0.0936 0.0078
Total 14 2.7420 0.1959

DF SS MS F P
Regression 1 0.0635 0.0635 8.1391 0.0145
Residual 12 0.0936 0.0078
Total 13 0.1571 0.0121

Statistical Tests:

PRESS 0.1228


128

Appendix 18
Regression equation of LAI by NDGVIc

Log(LAI) =y0+a*exp(NDGVIc)


0.7076 0.5007 0.4591 0.0809


y0 -0.7597 0.3436 -2.2113 0.0472 252.7891<
a 1.0489 0.3024 3.4688 0.0046 252.7891<


DF SS MS
Regression 2 2.6636 1.3318
Residual 12 0.0784 0.0065
Total 14 2.7420 0.1959

DF SS MS F P
Regression 1 0.0787 0.0787 12.0327 0.0046
Residual 12 0.0784 0.0065
Total 13 0.1571 0.0121

Statistical Tests:

PRESS 0.1139


129

Appendix 19
Correlations among AGB and LAI and the Landsat TM bands, Sep 15 2006

AGB LAI b1 b2 b3 b4 b5
LAI 0. 754
0. 002
b1 - 0. 233 - 0. 548
0. 422 0. 043
b2 - 0. 038 - 0. 423 0. 264
0. 897 0. 132 0. 363
b3 - 0. 356 - 0. 659 0. 660 0. 783
0. 212 0. 010 0. 010 0. 001
b4 - 0. 262 - 0. 552 0. 251 0. 776 0. 579
0. 366 0. 041 0. 387 0. 001 0. 030
b5 - 0. 512 - 0. 613 0. 752 0. 236 0. 676 0. 243
0. 061 0. 020 0. 002 0. 417 0. 008 0. 403
b7 - 0. 485 - 0. 536 0. 703 0. 039 0. 530 0. 057 0. 964
0. 079 0. 048 0. 005 0. 895 0. 051 0. 846 0. 000

Correlations among AGB and LAI and the Landsat TM bands, Sep 02 2007

AGB LAI b1 b2 b3 b4 b5
LAI 0. 754
0. 002
b1 - 0. 407 - 0. 156
0. 149 0. 593
b2 - 0. 373 - 0. 578 0. 511
0. 189 0. 031 0. 062
b3 - 0. 331 - 0. 400 0. 759 0. 798
0. 248 0. 157 0. 002 0. 001
b4 - 0. 569 - 0. 752 - 0. 081 0. 352 - 0. 015
0. 034 0. 002 0. 783 0. 217 0. 960
b5 - 0. 665 - 0. 601 0. 684 0. 805 0. 772 0. 442
0. 009 0. 023 0. 007 0. 001 0. 001 0. 114
b7 - 0. 481 - 0. 431 0. 748 0. 791 0. 877 0. 183 0. 947
0. 082 0. 124 0. 002 0. 001 0. 000 0. 531 0. 000

Cel l Cont ent s: Pear son cor r el at i on
P- Val ue

130

Appendix 20
Input File: 20020726 summer imagery
ROI Name: (J effries-Matusita, Transformed Divergence)

grassland-1 [Coral] 19426 points:
grassland-2 [Coral] 19113 points: (0.26330438 0.27679991)
larch-1 [Green] 2991 points: (2.00000000 2.00000000)
cedar-1 [Sea Green] 6396 points: (1.99999704 2.00000000)
waters [Blue] 147 points: (2.00000000 2.00000000)
bare soil-1 [White] 2371 points: (1.86551117 1.98250890)
broadleaf-1 [Yellow] 554 points: (1.99999998 2.00000000)


larch-1 [Green] 2991 points:


cedar-1 [Sea Green] 6396 points:

131



waters [Blue] 147 points:

bare soil-1 [White] 2371 points:


broadleaf-1 [Yellow] 554 points:

Pair Separation (least to most);

cedar-1 [Sea Green] 6396 points and cedar-2 [Sea Green] 6788 points - 0.23545208
grassland-1 [Coral] 19426 points and grassland-2 [Coral] 19113 points - 0.26330438
larch-1 [Green] 2991 points and larch-2 [Green] 3281 points - 0.71692593
larch-2 [Green] 3281 points and cedar-2 [Sea Green] 6788 points - 1.17925325
broadleaf-1 [Yellow] 554 points and broadleaf-2 [Yellow] 379 points - 1.18373520
larch-1 [Green] 2991 points and broadleaf-1 [Yellow] 554 points - 1.56513601
bare soil-1 [White] 2371 points and bare soil-2 [White] 529 points - 1.65035064
cedar-1 [Sea Green] 6396 points and broadleaf-2 [Yellow] 379 points - 1.66358911
grassland-1 [Coral] 19426 points and bare soil-1 [White] 2371 points - 1.86551117

132

cedar-1 [Sea Green] 6396 points and bare soil-1 [White] 2371 points - 1.99999648
grassland-1 [Coral] 19426 points and cedar-1 [Sea Green] 6396 points - 1.99999704
larch-2 [Green] 3281 points and waters [Blue] 147 points - 1.99999750
waters [Blue] 147 points and broadleaf-1 [Yellow] 554 points - 1.99999948
grassland-1 [Coral] 19426 points and broadleaf-2 [Yellow] 379 points - 1.99999983
bare soil-1 [White] 2371 points and broadleaf-1 [Yellow] 554 points - 1.99999999
grassland-1 [Coral] 19426 points and larch-2 [Green] 3281 points - 1.99999999
larch-2 [Green] 3281 points and bare soil-1 [White] 2371 points - 2.00000000
cedar-1 [Sea Green] 6396 points and waters [Blue] 147 points - 2.00000000
waters [Blue] 147 points and bare soil-1 [White] 2371 points - 2.00000000
grassland-2 [Coral] 19113 points and waters [Blue] 147 points - 2.00000000

133

Appendix 21
Input File: 20060915 fall imagery






134









135


136

Appendix 22
Input File: 20070902 fall imagery






137









138


139

Appendix 23
Input File: 20090907






140









141


142

Appendix 24
Input File: 20091212 winter imagery






143









144


145

Appendix 25
Input File: multi-temporal imageries






146









147


148

Appendix 26

Summer 2002 Fall 2006

Fall 2007 Fall 2009

Winter 2009
Land cover type and species separabiltiy within Landsat spectral sensitivity region.
Grassland coral, Larch green, Cedar sea green, Water blue, Bare soil white, Broadleaf -
yellow

149

Appendix 27
Single-date Landsat imageries of classification accuracy assessment for the land cover type map of the study area
Class Grassland Larch Cedar Bare soil Broadleaf Row Total User Acc.
Jul 06, 2002
C
l
a
s
s
i
f
i
e
d

D
a
t
a

Grassland 19050

19076 1.00
Larch 1842 259 26 49 2150 0.86
Cedar 1334 6464

79 7877 0.82
Bare soil 63

503 566 0.89
Broadleaf 105 65

250 420 0.60
Col. Total 19113 3281 6788 529 378 30089
Prod.Acc. 1.00 0.56 0.95 0.95 0.66 Diag. Total 28109

Overall accuracy 0.93

Kappa coefficient 0.88
Sep 15, 2006
C
l
a
s
s
i
f
i
e
d

D
a
t
a

Grassland 19041

20 19061 1.00
Larch 1945 586

78 2609 0.75
Cedar 1174 6186

71 7431 0.83
Bare soil 72

509 581 0.88
Broadleaf 162 16

230 408 0.56
Col. Total 19113 3281 6788 529 379 30090


Sep 02, 2007
C
l
a
s
s
i
f
i
e
d

D
a
t
a

Grassland 18896 6 17 4 18923 1.00
Larch 2047 996

25 3068 0.67
Cedar 1226 5754

63 7043 0.82
Bare soil 217

512 729 0.70
Broadleaf 8 32

287 327 0.88
Col. Total 19113 3281 6788 529 379 30090



150

Appendix 27
Single-date Landsat imageries of classification accuracy assessment for the land cover type map of the study area
Sep 07, 2009
C
l
a
s
s
i
f
i
e
d

D
a
t
a

Grassland 18671

118 18789 0.99
Larch 2306 836

30 3172 0.73
Cedar 532 5884

274 6690 0.88
Bare soil 442

411 853 0.48
Broadleaf 443 68

75 586 0.13
Col. Total 19113 3281 6788 529 379 30090


Dec 07, 2009
C
l
a
s
s
i
f
i
e
d

D
a
t
a

Grassland 18151 8 69 259 5 18492 0.98
Larch 1903 412

170 2485 0.77
Cedar 28 1150 6176

43 7397 0.84
Bare soil 34 3 12 270 15 334 0.81
Broadleaf 13 217 119

146 495 0.30
Col. Total 18226 3281 6788 529 379 29203



Khongor Dissertation

Загружено:

Сведения о документе

Исходное описание:

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Khongor Dissertation

Загружено:

Авторское право:

Доступные форматы

i

Ph.D. Program of Agriculture Science

, monthly average air temperature is -27.7

Вам также может понравиться