Вы находитесь на странице: 1из 10

International Journal of Applied Earth Observation and Geoinformation 13 (2011) 884893

Contents lists available at ScienceDirect

International Journal of Applied Earth Observation and


Geoinformation
journal homepage: www.elsevier.com/locate/jag

Comparing object-based and pixel-based classications for mapping savannas


Timothy G. Whiteside a, , Guy S. Boggs b , Stefan W. Maier b
a
b

Faculty of Health, Business and Science, Bachelor Institute of Indigenous Tertiary Education, Batchelor, NT 0845, Australia
School of Environmental and Life Sciences, Charles Darwin University, Darwin, NT 0909, Australia

a r t i c l e

i n f o

Article history:
Received 29 November 2010
Accepted 28 June 2011
Keywords:
Object-based image analysis
Accuracy assessment
Tropical savanna
Northern Australia

a b s t r a c t
The development of robust object-based classication methods suitable for medium to high resolution
satellite imagery provides a valid alternative to traditional pixel-based methods. This paper compares
the results of an object-based classication to a supervised per-pixel classication for mapping land
cover in the tropical north of the Northern Territory of Australia. The object-based approach involved
segmentation of image data into objects at multiple scale levels. Objects were assigned classes using
training objects and the Nearest Neighbour supervised and fuzzy classication algorithm. The supervised
pixel-based classication involved the selection of training areas and a classication using the maximum
likelihood classier algorithm. Site-specic accuracy assessment using confusion matrices of both classications were undertaken based on 256 reference sites. A comparison of the results shows a statistically
signicant higher overall accuracy of the object-based classication over the pixel-based classication.
The incorporation of a digital elevation model (DEM) layer and associated class rules into the objectbased classication produced slightly higher accuracies overall and for certain classes; however this was
not statistically signicant over the object-based using spectral information solely. The results indicate
object-based analysis has good potential for extracting land cover information from satellite imagery
captured over spatially heterogeneous land covers of tropical Australia.
2011 Elsevier B.V. All rights reserved.

1. Introduction
The Northern Territory of Australia is characterised by large
areas of land and a very small population. This situation is suitable for remote sensing data and analysis to map natural resource
information. A range of remotely sensed data has previously been
used to map land cover in northern Australia, however prior to this
study, most studies of land cover classication were based on traditional pixel-based methods (Ahmad et al., 1998; Hayder et al., 1999;
Menges et al., 2000), although there has been recent interest within
the Northern Territory in using object-based approaches to map
land cover (Crase and Hempel, 2005; Whiteside and Ahmad, 2004).
The nature of land cover in tropical Australia creates some issues
relating to pixel-based methods of classication (Whiteside, 2000).
Most of the native vegetation in northern Australia is relatively
intact and has undergone little modication apart from grazing and
re (Wilson et al., 1990). Thus, there are large areas of spectrally
similar but compositionally different vegetation cover or continuums of cover of varying densities of tree cover (Hayder, 2001) that
may be difcult to differentiate, potentially increasing classication
uncertainty. This discontinuous nature of tree cover within savan-

Corresponding author. Tel.: +61 8 89463831; fax: +61 8 89463833.


E-mail address: tim.whiteside@batchelor.edu.au (T.G. Whiteside).
0303-2434/$ see front matter 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.jag.2011.06.008

nas also contributes to the heterogeneous appearance of savannas


within imagery (Hutley and Settereld, 2008; Pearson, 2002). Similarly, when undertaking a per-pixel classication for the purposes
of mapping vegetation cover at a community scale, the spectral
heterogeneity within a particular land cover can lead to spurious
(misclassied at that scale) pixels appearing within classes creating
a salt and pepper effect (Blaschke et al., 2000). In addition to this,
the increased application of higher resolution imagery (pixel size
less than 5 m) is problematic as it is difcult to map community
scale classes accurately using traditional pixel-based methods. The
increased spectral heterogeneity within land cover classes often
leads to an inconsistent classication of pixels.
The development of robust object-based image analysis (OBIA)
methods suitable for the classication of medium (pixel size
1030 m) to high (pixel size 210 m) spatial resolution satellite
imagery provides a valid alternative to the traditional pixel-based
(PB) methods of analysing and categorising remotely sensed data
(Baatz et al., 2004; Benz et al., 2004). Pixels are grouped together
into objects or segments based on some criterion of homogeneity (either spectral or spatial). One of the benets of using objects
is that in addition to spectral information (for example, the mean
band values for each object), objects have a number of geographical/geometrical features attributed to them such as shape and
length, and topological entities, such as adjacency and, found
within (Baatz et al., 2004). These attributes create a knowledge base

T.G. Whiteside et al. / International Journal of Applied Earth Observation and Geoinformation 13 (2011) 884893

885

for the sample objects that is far richer in information than that for
individual pixels, and which can be called upon in the development
of a rule-based classication process.
This paper describes a comparison between a supervised
object-based classication process and a supervised per-pixel
classication for mapping tropical savanna land cover at a community level in the monsoonal north of the Northern Territory,
Australia using medium spatial resolution ASTER data. The comparison uses the VNIR bands of the ASTER image. Accuracies for
the classications are produced and compared using statistical
tests. A comparison between the object-based process based on
the spectral ASTER data and another incorporating an ASTER digital
elevation model (DEM) is also conducted.
1.1. Object-based image classication
Object-based (OB) image classication involves the segmentation of an image into homogeneous objects followed by the analysis
and classication of these objects. One of the advantages of segmentation is that it creates objects representing land cover types that
may be spectrally variable at the pixel level and thus eliminates
the salt and pepper effect associated with per-pixel classication.
Another advantage is that OB classication uses non-arbitrary units
for analysis as opposed to pixels; objects can approximate real
world features better than pixels. In addition, the use of shape features, the hierarchical structures of objects and classes, and the
topological features relating to the objects are other benets of OB
approaches. Object-based analysis enables the construction of rule
sets that can be used across a variety of scenes producing a repeatable methodology. One of the disadvantages of OBIA is a requisite
a priori knowledge of the area and the types of land cover under
investigation, which may not necessarily be available. Another
disadvantage is that the segmentation process and subsequent calculation of the topological relationships described between objects
can utilise a large amount of computer memory. Further, there is no
denitive algorithm or parameters for the creation of image objects.
Most assessment of the suitability of segmentation is undertaken by
visual assessment, although recently the use of local variance has
been applied to determine a suitable segmentation scale (Dragut
et al., 2010).
1.2. Comparison of OBIA and pixel-based image analysis
While there have been some studies comparing object-based
and pixel-based classication techniques little has been conducted
in northern Australia. Many publications claim that object-based
classication has greater potential for classifying higher resolution
imagery than pixel-based methods (Mansor et al., 2002; Oruc et al.,
2004; Willhauck et al., 2000). However, Dingle Robertson and King
(2011) found no statistically signicant difference between the two
methods in their study using McNemars non-parametric test for
proportional difference (de Leeuw et al., 2006), although visual
inspection indicated the OB approach incurred less signicant error
in the larger regions of homogeneous cover and performed better
in temporal analysis of land cover change (Dingle Robertson and
King, 2011). Niemeyer and Canty (2003) claim that object-based
classication has greater potential for detecting change in higher
resolution imagery. Castillejo-Gonzalez et al. (2009) found that
an object-based method out-performed ve pixel-based supervised classication algorithms (parallelepiped, minimum distance,
Mahanalobis Distance Classier, Spectral Angle Mapper, and MLC)
for mapping crops and agro-environmental associated measures.
Based on their reference data, Gao et al. (2006) achieved far greater
accuracy in mapping 12 land cover classes using an object-based
versus a pixel-based approach (83.25% and 46.48%, respectively),
while the object-based methodology used by Gao and Mas (2008)

Fig. 1. Location of the study site.

outperformed both MLC and NN pixel-based methods in mapping cover using SPOT 5 (10 m spatial resolution) data. However
the authors noted that after smoothing lters were applied to the
imagery, the accuracy of the pixel-based methods increased while
object-based accuracies decreased.
Jobin et al. (2008) noted that one of the advantages of OBIA is the
utility of a knowledge base that is beyond purely spectral information and includes object-related features such as shape, texture and
context/relationship, along with the capability to include ancillary
data. Manakos et al. (2000) found that ancillary data utilised within
object-based classication improved classication accuracy. Similarly, Devhari and Heck (2009) found the overall accuracy of the
object-based classication was greater than pixel-based and better
when ancillary surface data (DEM and contours) were introduced
into the segmentation. Myint et al. (2011) found that including principal component images and a Normalized Difference Vegetation
Index within an object-based rule-set classication produced signicantly higher accuracy than MLC. Yu et al. (2006) incorporated
ancillary and derivative data (IHS transformations, elevation slope
and aspect, and a water channel GIS layer) with 4 band airborne
digital camera imagery to map vegetation in Douglas-r, California
Bay and Coast live oak communities to 60% overall accuracy.
2. Methods
2.1. Study area
The study area is part of the Florence Creek region of Litcheld National Park, in the northwest of the Northern Territory of
Australia (Fig. 1). The area covers 1373 ha and is located near two
of the parks major features, Florence Falls and Buley Rockhole. The
regions climate is characteristic of the wet/dry tropics; consisting
of a long dry season (MaySeptember) with little to no rainfall, and
over 75% of the annual rainfall (1500 mm) occurring in the period
between November and March. Maximum daily temperatures vary
from just under 32 C in June and July to over 36 C in October and
November. The vegetation within the study area is predominantly
open forest and savanna woodland with a discontinuous Eucalyptus spp. (mostly E. tetradonta and E. miniata) dominated canopy and
continuous annual grass (Sarga spp.) understorey (Grifths et al.,
1997). Patches of monsoon rain forest are located on springs near
the base of the escarpment and other areas of permanent water.
Melaleuca spp. forests occur along creek lines and share overlapping
species with the monsoon rain forest (i.e. Xanthostemon eucalyptoides and Lophostemon lactiuus) (Lynch and Manning, 1988). The
southern portion of the study area is generally plateau surfaces

886

T.G. Whiteside et al. / International Journal of Applied Earth Observation and Geoinformation 13 (2011) 884893

intersected by drainage lines, while low lying areas subject to inundation are located to the north. At the northern end of the scene,
where drainage lines intersect with the plateau edge, a number of
gorges and associated waterfalls occur. This terrain will produce
areas of shading and slope effects which will inuence irradiance
although the area affected is small.
2.2. Remote sensing data
ASTER data for the area were captured on 28 July 2000 providing 15 spectral bands; four in the visible and near infrared
(0.520.86 m) (VNIR) at 15 m nominal pixel size, six in the shortwave infrared (1.602.43 m) at 30 m and ve in the thermal
infrared at 90 m (8.1211.65 m) (Yamaguchi et al., 1998). The data
were delivered processed to level surface reectance with topographic corrections. The near infrared (NIR) band 3 is captured at
nadir (3N) and backwards looking (3B), producing a stereo pair
of images that can be used to create a DEM (Hirano et al., 2003).
The relative DEM image derived from ASTER bands 3N and 3B was
requested from NASAs EROS Data Center and acquired in October 2002. A slope layer (%) was derived from the DEM image for
inclusion.
Pre-processing included geometric correction and the creation
of subsets for the study area. Geometric rectication of the imagery
was undertaken by the authors using a rst order polynomial with
a nearest neighbour interpolation, incorporating the DEM with 25
ground control points taken from a 1:100,000 topographic map
producing a RMSE of less than 0.5 pixels (7.5 m). Further topographic adjustments of the spectral imagery were not undertaken
due to the relatively small area affected by slope and the unsuitability of the ASTER DEM in accurately reecting such small variations.
Subsets for the study area were created from the ASTER VNIR bands
1, 2 and 3N (green, red and NIR nadir, respectively) and the DEM
(Fig. 2). Pixel dimensions of the subset are 249 (columns) 245
(rows) for the 15 m ASTER layer.
2.3. Object-based classication
Two object-based classications were undertaken using
Deniens Developer version 7 software. The rst classication was
conducted using only the VNIR bands from the ASTER imagery for
comparison with the per-pixel classication, while the second classication incorporated information from the ASTER DEM. Both OB
classications involved two sub-processes: (i) segmentation and
(ii) classication (Fig. 3).
2.3.1. Multi-scale segmentation
The object-based approach rst involved the segmentation of
image data into objects on two scale levels. The subset images were
initially segmented into object primitives or segments using the
multiresolution segmentation algorithm that follows the fractal net
evolution approach (Baatz and Schpe, 2000). The segmentation of
the images into object primitives is inuenced by three parameters: scale, colour and form (Willhauck et al., 2000). The algorithm
is primarily an iterative bottom-up segmentation method starting
with individual pixels and merging these pixels based upon pixel
heterogeneity and object shape and colour. These are determined
within the algorithm by two parameters; (i) scale and (ii) colour
versus form. The scale parameter within the algorithm is set by
the operator and is inuenced by the heterogeneity of the pixels. The colour parameter balances the homogeneity of an objects
colour with the homogeneity of its shape. The form parameter is
a balance between the smoothness of a segments border and its
compactness. The weighting of these parameters establishes the
homogeneity criterion for the object primitives. Visual inspection of
the objects resulting from a number of segmentations using varia-

Fig. 2. Subsets of ASTER VNIR bands with RGB displaying NIR, Red and Green, respectively, (a) and the ASTER DEM (b). (For interpretation of the references to colour in
this gure legend, the reader is referred to the web version of the article.)

tions in the weightings was used to determine the overall values for
the parameter weighting at each scale level (Table 1). Scale parameters greater than 10 tended to undersegment the image with
noticeable mixes of land covers. Scale parameters smaller than 5
tended to oversegment the image with many adjacent objects of the
same land cover observed. Due to irregular patterns and boundaries
of land covers in the natural landscape within the image, emphasis was placed on the colour (lack of spectral variability) criterion

T.G. Whiteside et al. / International Journal of Applied Earth Observation and Geoinformation 13 (2011) 884893

887

Table 2
Classes used in this paper. Note: The last two classes WF, CFl, EW1 and WS
are interim classes not used in the nal classication and thus have no colour
assigned to them. (For interpretation of the references to colour in this table,
the reader is referred to the web version of the article.)

Colour code

Fig. 3. Data ow for object-based classication. Rectangles represent image layers


used in the procedure; circles represent processing and analysis steps; and clouds
represent objects.

of an object over its shape (60% to 40%, respectively). An increase


in the ratio of colour to shape in the algorithm resulted in objects
with extremely convoluted boundaries, whereas an increase in the
shape parameter saw regular shaped segments not representing
identiable land cover boundaries. Two levels were chosen as the
rst level provided objects at a suitable scale for mapping however
some of the objects still contained noticeable spectral variability.
Thus, a ner scale was introduced to produce smaller objects of
less variability that could then be combined into suitable classes
later.
2.3.2. Classication
A total of ten land cover classes for the study area were identied
based on the structural formation of the vegetation and characteristic taxon (Table 2). Two of these classes were introduced to include
areas of the study site that were identied as recently burnt. The
OB classication involved supervised classication using objects
selected for training data based on their class as determined by
a combination of eld observation and aerial photo interpretation. The mean spectral values of the training objects were used
as input into the Nearest Neighbour algorithm which uses feature
space to classify objects based on the closest training examples
with an object being assigned the class of the majority of its neighbours.
The OB classication procedure incorporating the DEM involved
a combination of a supervised classication and the development
of class rules based upon the mean spectral signatures, as well as
the DEM and slope values of the objects. Samples for each class
were selected from the image objects to act as training areas for
the classication.

Class code

Class name

EOFB

Burnt Eucalypt open forest

EOF

Eucalypt open forest

CFo

Mixed closed forest

MelF

Melaleuca riparian forest

EW

Eucalypt woodland

EWB

Burnt Eucalypt woodland

EWRO

Eucalypt woodland with rocky


outcrops

OW

Open woodland

MW

Mixed woodland

GL

Grassland

WF

Woodland flats

CFl

Creek flats

EW1

Eucalypt woodland 1

WS

Woodland slope

Details of the multilevel OB classication are depicted in Fig. 5.


Initial classication of the broad scale (Level 2) segmentation was
conducted using the Nearest Neighbour (NN) classication algorithm with the relevant training samples shown in Fig. 4. The NN
classication used the object means of all three spectral bands to
create ve broad classes (Riparian, EOF, EW, GL, and MF). The ner
scale Level 1 objects were initially categorised based on the class
assigned to their Level 2 super-object. Further division of classes
was then undertaken using NN supervised classication based on
the sample objects (Fig. 4). The samples were then used as a basis
for fuzzy classication of the data in Deniens (Baatz et al., 2004).
Fuzzy classications use fuzzy logic to account for the heterogeneous nature (consisting of multiple land covers) of pixels (or in this
case objects) in medium spatial resolution imagery as well as the
lack of hard, sharp boundaries in nature where classes grade into
one another (Foody, 1992). Where there exists m number of classes,
pixels are assigned m class memberships providing the degree of
truth to which pixels (or objects) belong to each class (Jensen,
2005). The class with the highest membership (or truth) value is
assigned as the best and nal class for each object (Baatz et al.,
2004). After the Level 1 classication step, the interim WF (Woodland ats) and CFl (Creek ats) objects were then merged into MW
(Mixed woodland) class producing 10 classes for the purpose of
comparison with the pixel-based classication and reference data
(Fig. 5).
2.4. Object-based classication with ancillary DEM layer

Table 1
Segmentation parameters used at both scale levels.
Scale level
2
1

Scale parameter
10
5

Shape factor

Compactness

Smoothness

0.4
0.2

0.7
0.7

0.3
0.3

A second object-based classication was conducted this time


incorporating the DEM data. The procedure using training sample objects and the NN algorithm was the same as for the initial
object-based classication; however two DEM based class rules
were introduced for the purpose of classifying the areas of the

888

T.G. Whiteside et al. / International Journal of Applied Earth Observation and Geoinformation 13 (2011) 884893
Table 3
Reference data details.
Year

No. sites

2000
2002
2003
2004
2006
2006

1
11
29
20
4
100

Air photo
Total

91
256

2.5. Pixel-based supervised classication


The pixel-based classication used a supervised maximum likelihood (MLC) algorithm (Jensen, 2005; Lillesand and Kiefer, 2008).
50 training areas representative of the ten land cover classes (Fig. 4)
were selected to develop class signature les similar to those used
in the object-based classication to ensure consistency. The MLC
then assigned pixels to the class of highest probability producing a
land cover map for these 10 classes.
2.6. Accuracy assessment

Fig. 4. Sample objects used for the object-based NN classication at levels 1 and
2 draped over the ASTER VNIR imagery (RGB = bands 1, 2, 3N). Colours and land
cover codes correspond to those in Table 2. Points are locations of sites used for
accuracy assessment. Circles are eld sites, triangles are derived from aerial photo
interpretation. (For interpretation of the references to colour in this gure legend,
the reader is referred to the web version of the article.)

subset where irradiance was affected by slope. Level 2 Eucalypt


woodland objects were classied at level 1 as Open woodland
objects if their DEM value was less than 140 m AMSL and as Woodland slope if their slope was greater than 8%. The threshold values
for these class rules distinguished the affected areas. Woodland
slope objects were then reclassied as Eucalypt woodland for the
nal output.

The accuracy of the fuzzy classication was estimated using the


mean membership value of the best classication (class with highest membership value) and the mean stability for each class (Baatz
et al., 2004). The mean stability is the difference between an objects
rst and second class memberships and is expressed as a percentage the higher the value the greater the stability and the less
expected mixing of the classes.
The accuracy of the classied image was also assessed using a
range of reference data including eld data collected in the study
area between 2000 and 2006 (Table 3) and interpretation of aerial
photography of the area. Each campaign has collected vegetation
data that have enabled the determination of the structural classication of the communities. Data included the dominant taxon
(Genus or species), canopy height, stem diameter (at 130 cm above
the ground), basal area, foliage projective cover, canopy cover and
canopy density. Vegetation structural classes were determined

Fig. 5. Classication processes for the object-based land cover classications based on spectral bands.

T.G. Whiteside et al. / International Journal of Applied Earth Observation and Geoinformation 13 (2011) 884893

889

based on the standard systems used in Australia (Hnatiuk et al.,


2009; Specht, 1981). In areas that were inaccessible (due to cultural restrictions or terrain), aerial photograph interpretation was
undertaken to delineate the vegetation community. A stereo pair
of 1:43,000 colour positive aerial photographs taken at 1500 h on
5 May 2000 covering the study area were observed under a stereoscope with 6 magnication. Cover calculations after Fensham
et al. (2002) and dominant taxon were identied and structural
classes determined over 90 points derived from a stratied random
sampling over the area. Locations of all reference sites are shown
in Fig. 4.
For ease of comparison between classication methods, thematic accuracy was undertaken using only the point-based
reference data mentioned above. There was no consideration of
any object-based accuracy assessment or accuracy measures relating to the geometric accuracy of the objects (such as location and
shape).
Accuracy assessments of all three classications were undertaken using confusion matrices and Kappa statistics (Congalton,
1991). Producer and User accuracies (Story and Congalton, 1986)
for each class were calculated along with the overall accuracies and Kappa coefcient statistics (Congalton and Green, 1999).
Conditional Kappa was calculated for each class. The statistical
signicance of the difference between the two overall Kappa coefcients for the pixel-based classication and the object-based
classication was assessed using a Z-test (Congalton and Green,
2009).
As one set of reference data was used to assess the accuracy of
both classications, the two confusion matrices may show dependence and thus a further test for statistical signicance of the
difference between the two classications was undertaken using
McNemars test (de Leeuw et al., 2006). The test is non-parametric
(Eq. (1)) assuming that the number of correctly and incorrectly
identied points is equal for both classications (de Leeuw et al.,
2006; Dingle Robertson and King, 2011; Gao et al., 2006).
2 =

(f12 f21 )
f12 + f21

(1)

where f12 and f21 , respectively, are the number of points correctly
identied by one classier and not the other. The statistic follows a chi-squared distribution with one degree of freedom (Dingle
Robertson and King, 2011). The rst object-based classication was
then compared to the object-based classication incorporating the
DEM data using the same statistical tests.
3. Results
Within the object-based analysis, the segmentation provided
615 level 2 objects and 1495 level 1 objects for classication
(Fig. 6a). At Level 2 it is visually apparent that objects of this size still
can contain more than one spectrally distinct land cover (Fig. 6b).
The objects within the segmentation at level 1 tend to follow the
spectral boundaries more closely. The image resulting from the
object-based classication is shown in Fig. 7a. For comparison, the
area (ha) assigned to each class is presented in Table 4.
According to the object-based classication the land cover class
occupying the largest area is Eucalypt woodland (376 ha) with the
number of objects identied as belonging in that class being 675.
Within the per-pixel classication only 256 ha were assigned to this
class. The class with the largest area in the per-pixel classication is
Eucalypt open forest with 302 ha (140 ha in the object-based classication). The land cover with the smallest area within the study
site is Grassland occupying only 15 ha and consisting of just 21
image objects in the OB classication and 16 ha in the PB classication. A visual comparison of the resultant land cover images shows

Fig. 6. A section of the study area showing hierarchical segmentation at scale level
2 (a) and level 1 (b).

noticeable differences between the classications (Fig. 7). While


both methods produce aggregations of pixels based on land cover
classes, the object-based classication yields multi-pixel features
whereas the pixel-based classication contains many small groups
of pixels or individual pixels. This produces classes with mixed clusters of pixels as displayed by the heterogeneous nature of the PB
classied image.
The areas of classes with cover that is more spectrally homogeneous such as Mixed closed forest, Melaleuca riparian forest
and Grassland classes are relatively similar in both OB and PB
classications (Fig. 7 and Table 4). The area classied as Eucalypt woodland is noticeably lower in the pixel-based classication
compared to the two object-based classications. This is supported
by visual comparison with the original image showing Eucalypt

890

T.G. Whiteside et al. / International Journal of Applied Earth Observation and Geoinformation 13 (2011) 884893
Table 4
Areas of classes determined by object-based and pixel-based classications.
Class name

Burnt Eucalypt open forest


Eucalypt open forest
Mixed closed forest
Melaleuca riparian forest
Eucalypt woodland
Burnt Eucalypt woodland
Eucalypt woodland with
rocky outcrops
Open woodland
Mixed woodland
Grassland

Fig. 7. Resultant images of the object-based (a) and pixel-based (b) classications.

woodland classes apparently under-represented in the pixel-based


classication, while there is an apparent over-representation of
Eucalypt open forest, Eucalypt woodland with rocky outcrops and
Grassland classes.
A summary of the accuracy assessment of the fuzzy classication for both object-based classications is presented in Table 5.

Area (ha)
Object-based

OB with DEM

Pixel-based

228.38
140.54
19.58
168.09
376.02
114.39
42.59

228.38
140.54
19.58
168.09
368.24
114.39
42.59

188.53
302.24
21.94
173.68
256.10
77.38
118.19

57.55
208.84
15.18

65.33
208.84
15.18

37.59
180.86
16.11

The Mean p1st column displays the mean best classication of


objects belonging to that class. Nearly all classes have a mean
p1st value over 0.8 suggesting that the degree to which image
objects were being assigned to the best class was high. Within the
rst object-based classication, the class with the lowest mean
best classication value is Open woodland with 0.68. Most of
the classes display stability in their classication; only the Eucalypt open forest, Melaleuca riparian forest and Mixed woodland
classes have a mean stability value under 0.1. This suggests uncertainty in some much as a number of objects within these classes
may not belong to their assigned class or potentially contain more
than one class.
From the results of the confusion matrices based on the reference data set, the proportions of pixels classied due to chance
agreement (Congalton and Green, 2009) were 13% for the OB classication and 12.5% for the PB classication. The overall accuracy
of the object-based classication was higher than for the pixelbased classication, 78.5% versus 69.5%, respectively (Table 6).
This was also the case for the overall Kappa statistic: the objectbased classication had an overall Kappa of 0.74 with a standard
error of 0.03 while the pixel-based classications overall Kappa
statistic was 0.65 with a standard error of 0.03. The resultant Z-value (Table 7) calculated from the comparison of Kappa
coefcients from the two classication methods was 2.29, which
is greater than the critical value for Z at the 95% condence
level (1.96). This indicates a statistically signicant difference
between the results of the two classications. The 2 value from
the McNemars test (Table 7) was 8.97 with a p-value of 0.01,
strengthening the argument for a statistically signicant difference between the two classication methods. In addition, the
Producer and User accuracies were greater for the majority of the
classes in the object-based classication. Five classes in the OB
classication had User and Producer accuracies over 70%, while
only three did for the PB classication. Six of the classes from
the OB classication had conditional Kappa coefcients of 0.75
or greater while the PB classication had four. The land cover
classes that were more accurately classied using the pixel-based
method were the heavily treed classes of homogeneous cover,
Melaleuca riparian forest and Mixed closed forest. The classes
that had poor Users accuracy in both classications were Mixed
woodland and Grassland and may, in part, be attributable to
the small number of reference data points for those classes (19
and 4, respectively), as any reference site erroneously included
in these classes will have an effect. Object-based classication
appears to be able to differentiate more accurately the relatively
heterogeneous Eucalypt open forest and the several woodland
classes.
Comparison of the OB and OBDEM classications shows a
slightly higher accuracy for the OBDEM classication. The two
classes (Eucalypt woodland and Open woodland) inuenced by

T.G. Whiteside et al. / International Journal of Applied Earth Observation and Geoinformation 13 (2011) 884893

891

Table 5
Fuzzy classication accuracy assessment of the initial object-based image analysis and with DEM.
Class

EOFB
EOF
MFo
MelF
EW
EWB
EWRO
OW
MW
GL

Class name

Object-based

Burnt Eucalypt open forest


Eucalypt open forest
Mixed closed forest
Melaleuca riparian forest
Eucalypt woodland
Burnt Eucalypt woodland
Eucalypt woodland with rocky outcrops
Open woodland
Mixed woodland
Grassland

Object-based with DEM

Mean p1st

Mean stability

Mean p1st

Mean stability

0.880
0.827
0.810
0.796
0.832
0.856
0.839
0.676
0.874
0.840

0.136
0.092
0.362
0.048
0.145
0.373
0.362
0.373
0.048
0.138

0.880
0.827
0.810
0.796
0.858
0.856
0.839
0.756
0.874
0.840

0.136
0.092
0.362
0.048
0.215
0.373
0.362
0.385
0.048
0.138

Table 6
Summary of confusion matrices for the accuracy of object-based and pixel-based classications including the Producers accuracy (PA), Users accuracy (UA) and conditional
Kappa (CK) for each class as well as overall accuracy, probability of chance results and Kappa statistic.
Class name

Object-based classication

Burnt Eucalypt open forest


Eucalypt open forest
Mixed closed forest
Melaleuca riparian forest
Eucalypt woodland
Burnt Eucalypt woodland
Eucalypt woodland with rocky outcrops
Open woodland
Mixed woodland
Grassland

Object-based classication with DEM

Pixel-based classication

PA (%)

UA (%)

CK

PA (%)

UA (%)

CK

PA (%)

UA (%)

CK

91.00
81.00
100.0
67.74
77.42
85.71
66.67
91.67
73.68
50.00

75.00
76.32
66.67
80.77
90.57
80.00
100.0
91.67
50.00
50.00

0.68
0.73
0.66
0.75
0.88
0.78
1.00
0.87
0.43
0.51

91.00
81.00
100.0
67.74
77.77
85.71
66.67
92.31
73.68
50.00

75.00
76.32
66.67
80.77
92.45
80.00
100.0
92.31
50.00
50.00

0.68
0.73
0.66
0.75
0.89
0.78
1.00
0.88
0.43
0.51

63.64
75.00
91.67
87.10
50.00
84.62
63.63
83.33
80.00
75.00

65.63
57.45
100.0
90.00
75.61
55.55
87.5
76.92
51.61
42.86

0.61
0.50
1.00
0.89
0.68
0.53
0.85
0.76
0.48
0.39

Overall accuracy
Chance results
Kappa

78.51%
13.15%
0.7526

79.30%
13.24%
0.7611

the rule set have higher Producer and User accuracies. The overall accuracy and Kappa were also higher although the difference
in accuracies (based on Kappa) was not statistically signicant
(Table 8).
4. Discussion
The object-based image analysis method applied in this paper
provided results with statistically signicant higher accuracies
than the pixel-based classication. This is consistent with ndings within the literature (Castillejo-Gonzalez et al., 2009; Gao and
Mas, 2008; Gao et al., 2006). This result suggests that object-based
analysis has potential as an alternative method (over per-pixel
approaches) for extracting land cover information from medium to
high resolution satellite imagery captured over tropical savannas
in Australia.
The improved classication using OBIA can be attributed primarily to its use of objects to reduce the spectral variability in
land cover types that are heterogeneous such as savanna. Based
on a community level classication such as used here, pixel-based
Table 7
Comparison statistics Z-test and McNemars 2 for pixel-based and object-based
classications.
Z-value for Kappas
McNemars 2

2.285
8.966

p-value 0.05 (p = 0.01)


p-value 0.01 (p = 0.0028)

Table 8
Comparison statistics Z-test and McNemars 2 for the Object-based classication
and Object-based with DEM classication.
Z-value for Kappas
McNemars 2

0.2175
0.995

p = 0.41
p = 0.35

69.53%
12.5%
0.6516

classications do misclassify pixels, particularly in land covers that


are spectrally heterogeneous in medium (pixels of 30 m or less)
and high (pixels of 5 m or less) spatial resolution imagery, such
as tropical savanna. Classifying savanna into discrete community
level classes (such as forest or woodland) is difcult with perpixel approaches due to their sensitivity to the discontinuous and
variable nature of the woody cover within such landscapes. For
example, within the pixel-based classication many of the pixels
classied as Eucalypt woodland with rocky outcrops are actually
grass understorey or bare ground between tree crowns where as
tree crown canopy itself has been classied as Open forest as the
co-dominants within savanna woodland are trees and grass. The
combination of the two classes should produce a Woodland classication. This would account for the greater confusion found in
the Eucalypt dominant classes as mentioned in the results where
the Eucalypt woodland classes are under-represented while there
is an over-representation of the Eucalypt open forest, Eucalypt
woodland with rocky outcrops and Grassland classes. Objectbased classication appears to overcome some of the problems
encountered using pixel-based methods to classify community
level Eucalypt land cover types and their characteristic spatial heterogeneity (Pearson, 2002), while it is evident that pixel-based
classication is still quite successful in classifying land cover of a
spectrally homogenous nature (i.e. Mixed closed forest).
One of the advantages of object-based classication is the ability
to use ancillary data (such as derivative data sets, data from other
sensors and existing GIS layers) as additional information layers to
assist with land cover mapping is being investigated. Another feature is the utility of contextual information related to objects. One
such data set as used in this study is the incorporation of height
information in the form of DEM and derivative slope layer. Such
information enables better delineation of land cover types that
occupy certain locations within the landscape such as on slopes

892

T.G. Whiteside et al. / International Journal of Applied Earth Observation and Geoinformation 13 (2011) 884893

and drainage ats. This is consistent with ndings of other studies


(Devhari and Heck, 2009; Manakos et al., 2000).
Comparing object-based and pixel-based classications in this
manner presents a number of issues. While, both methods produce
a thematic map, the object-based approach appears to eliminate
the salt and pepper or noise effect by considering mean pixel values
within objects as opposed to individual pixel values. However, in
using the traditional point-based method of accuracy assessment,
the comparison between point-based reference data and the classication is based on individual or site-specic cases (i.e. pixels in
a pixel-based classication and specic points within objects). As
such, information relating to the spatial accuracy of the classied
objects is not considered (Lucieer, 2004). For example, is that class
consistent across the entire object? Is the object located in a similar
location to the real-world object it is representing? To what degree
(area, length, etc.) is the shape of the object similar to the real-world
object? This information is important in determining the quality of
an object-based classication.
The tendency to produce spurious or misclassied pixels within
classes (the so-called salt & pepper effect) means that heterogeneous land covers will nearly invariably have slightly lower
accuracies for pixel-based classications than object-based using
classes such as used here. Part of this may attributed to misregistration between the imagery and eld data. Methods to
improve accuracies for pixel-based classications include some
post-classication editing such as ltering and manual removal
(Johansen and Phinn, 2006). Potential under-evaluation will occur
within certain classes that are heterogeneous in cover such as
savanna in which cover is co-dominated by grass and discontinuous and variable woody cover. Trees will be assigned to a forest
class and understorey gaps between trees will be assigned to grassland class. Thus it may be necessary to redene classes away from
the traditional land cover or vegetation classes into more contextual classes, e.g. canopy versus non-canopy or using quantitative
measures (e.g. % canopy cover). This is where the hierarchical structure of OB classication has potential in enabling the use of these
types of classes at a particular level and then the proportions of
these classes in super-objects at higher level to determine level of
canopy cover in those objects.

5. Conclusions
The results of this study show a signicant difference in the
accuracy between a pixel-based maximum likelihood classier
and object-based hybrid nearest-neighbour/rule-based classier
for mapping land cover from tropical savanna using ASTER data.
Resultant noise in the pixel-based classication suggests that thematic mapping using high spatial resolution satellite data requires
a new methodology in land cover classication forgoing the traditional community level classications for the initial stages of
classication and perhaps focussing on the smaller spatial elements
such as tree/canopy, grass, and bare ground. These objects could
then be the basis to develop the community/structural level classication grouping based on the proportional values of the various
components. This is where OBIA has great advantage over per-pixel
classication methods.

References
Ahmad, W., Hill, G.J.E., OGrady, A.P., Menges, C., 1998. Use of multispectral scanner
data for the identication and mapping of tropical forests of northern Australia.
Asian-Pacic Remote Sensing and GIS Journal 11, 1422.
Baatz, M., Schpe, A., 2000. Multiresolution segmentation an optimization
approach for high quality multi-scale image segmentation. In: Strobl, J.,
Blaschke, T., Griesebner, G. (Eds.), Angewandte Geographische Informationsverarbeitung XII. Wichmann-Verlag, Heidelberg, pp. 1223.

Baatz, M., Benz, U., Dehghani, S., Heynen, M., Hltje, A., Hofmann, P., Lingenfelder,
I., Mimler, M., Sohlbach, M., Weber, M., Willhauck, G., 2004. eCognition Professional: User Guide 4. Deniens-Imaging, Munich.
Benz, U., Hofmann, P., Willhauck, G., Lingenfelder, I., Heynen, M., 2004. Multiresolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready
information. ISPRS Journal of Photogrammetry and Remote Sensing 58,
239258.
Blaschke, T., Lang, S., Lorup, E., Strobl, J., Zeil, P., 2000. Object-oriented image processing in an integrated GIS/remote sensing environment and perspectives for
environmental applications. In: Cremers, A., Greve, K. (Eds.), Environmental
Information for Planning, Politics and the Public. Metropolis Verlag, Marburg,
pp. 555570.

Castillejo-Gonzalez, I.L., Lpez-Granados, F., Garcia-Ferrer, A., Pena-Barragan,


J.M., Jurado-Expsito, M., Sanchez-de la Orden, M., Gonzalez-Audicana, M.,
2009. Object- and pixel-based analysis for mapping crops and their agroenvironmental associated measures using QuickBird imagery. Computers and
Electronics in Agriculture 68, 207215.
Congalton, R.G., 1991. A review of assessing the accuracy of classications of
remotely sensed data. Remote Sensing of Environment 37, 3546.
Congalton, R.G., Green, K., 1999. Assessing the Accuracy of Remotely Sensed Data:
Principles and Practice. Lewis Publishers, New York.
Congalton, R.G., Green, K., 2009. Assessing the Accuracy of Remotely Sensed Data:
Principles and Practices, 2nd ed. CRC Press, Boca Raton, FL.
Crase, B., Hempel, C., 2005. Object based land cover mapping for Groote Eylandt: a
tool for reconnaissance and land-based surveys. In: Proceedings of NARGIS 05,
the North Australian Remote Sensing and GIS Conference, Darwin, July 47.
de Leeuw, J., Jia, H., Yang, L., Liu, X., Schmidt, K., Skidmore, A.K., 2006. Comparing accuracy assessments to infer superiority of image classication methods.
International Journal of Remote Sensing 27, 223232.
Devhari, A., Heck, R.J., 2009. Comparison of object-based and pixel based infrared
airborne image classication methods using DEM thematic layer. Journal of
Geography and Regional Planning 2, 8696.
Dingle Robertson, L., King, D.J., 2011. Comparison of pixel- and object-based classication in land cover mapping. International Journal of Remote Sensing 32,
15051529.
Dragut, L., Tiede, D., Levick, S.R., 2010. ESP: a tool to estimate scale parameter for multiresolution image segmentation of remotely sensed data. International Journal
of Geographical Information Science 24, 859871.
Fensham, R.J., Fairfax, R.J., Holman, J.E., Whitehead, P.J., 2002. Quantitative assessment of vegetation structural attributes from aerial photography. International
Journal of Remote Sensing 23, 22932317.
Foody, G.M., 1992. A fuzzy sets approach to the representation of vegetation continua
from remotely sensed data: an example from lowland health. Photogrammetric
Engineering and Remote Sensing 58, 221225.
Gao, Y., Mas, J.F., 2008. A comparison of the performance of pixel-based and
object-based classications over images with various spatial resolutions. In:
Proceedings of GEOBIA 2008 Pixels, Objects, Intelligence: Geographic ObjectBased Image Analysis for the 21st Century, Calgary, Alberta, August 67.
Gao, Y., Mas, J.F., Maathius, B.H.P., Xiangmin, Z., van Dijk, P.M., 2006. Comparison of
pixel-based and object oriented image classication approaches a case study
of a coal re area, Wuda, inner Mongolia, China. International Journal of Remote
Sensing 27, 40394055.
Grifths, A.D., Woinarski, J.C.Z., Armstrong, M.D., Cowie, I.D., Dunlop, C.R., Horner,
P.G., 1997. Biological Survey of Litcheld National Park, Report No. 62. PWCNT,
Darwin.
Hayder, K., 2001. Study of remote sensing and GIS for the assessment of their capabilities in mapping the vegetation form and structure of tropical savannas in
Northern Australia. Unpublished Ph.D. Thesis. Northern Territory University,
Darwin, 315 p.
Hayder, K., Ahmad, W., Williams, R.J., 1999. Use of varied resolution imagery for
land cover mapping in a semi-arid tropical savanna at Kidman Springs, Northern
Territory, Australia. In: Proceedings of NARGIS 99, 4th North Australian Remote
Sensing and GIS Conference, Darwin, NT, June 2830.
Hirano, A., Welch, R., Lang, H., 2003. Mapping from ASTER stereo image data:
DEM validation and accuracy assessment. ISPRS Journal of Photogrammetry and
Remote Sensing 57, 356370.
Hnatiuk, R.J., Thackway, R., Walker, J., 2009. Vegetation. Australian Soil and Land
Survey Field Handbook. CSIRO, Melbourne.
Hutley, L.B., Settereld, S.A., 2008. Savanna. In: Jorgensen, S.E., Fath, B.D. (Eds.),
Encyclopedia of Ecology, vol. 4: Ecosystems. Elsevier, Oxford.
Jensen, J.R., 2005. Introductory Digital Image Processing: A Remote Sensing Perspective, 3rd ed. Prentice Hall, Upper Saddle River, NJ.
Jobin, B., Labrecque, S., Grenier, M., Falardeau, G., 2008. Object-based classication
as an alternative approach to the traditional pixel-based classication to identify
potential habitat of the Grasshopper Sparrow. Environmental Management 41,
2031.
Johansen, K., Phinn, S., 2006. Mapping structural parameters and species composition of riparian vegetation using IKONOS and landsat ETM+ data in Australian
Tropical Savannahs. Photogrammetric Engineering & Remote Sensing 72,
7180.
Lillesand, T.M., Kiefer, R.W., 2008. Remote Sensing and Image Interpretation, 6th ed.
Wiley, New York/Hoboken.
Lucieer, A., 2004. Uncertainties in segmentation and their visualisation. Unpublished Ph.D. Thesis. International Institute for Geo-Information Science
and Earth Observation (ITC) and the University of Utrecht, Netherlands,
177 p.

T.G. Whiteside et al. / International Journal of Applied Earth Observation and Geoinformation 13 (2011) 884893
Lynch, B.T., Manning, K.M., 1988. Land Resources of Litcheld Park, Report No. 36.
Conservation Commission of the Northern Territory, Darwin.
Manakos, I., Schneider, T., Ammer, U., 2000. A comparison between the ISODATA
and the eCognition classication on basis of eld data. In: Proceedings of XIX
ISPRS Congress, Amsterdam, July 1622.
Mansor, S., Hong, W.T., Shariff, A.R.M., 2002. Object oriented classication for land
cover mapping. In: Proceedings of Map Asia 2002, Bangkok, GIS Development,
August 79.
Menges, C.H., Bell, D., van Zyl, J.J., Ahmad, W., Hill, G.J.E., 2000. Image classication
of AIRSAR data to delineate vegetation communities in the tropical savannas
of northern Australia. In: Proceedings of 10th Australian Remote Sensing &
Photogrammetry Conference, Adelaide, August 2125, pp. 12861298.
Myint, S.W., Gober, P., Brazel, A., Grossman-Clarke, S., Weng, Q., 2011. Per-pixel
vs. object-based classication of urban land cover extraction using high spatial
resolution imagery. Remote Sensing of Environment 115, 11451161.
Niemeyer, I., Canty, M.J., 2003. Pixel-based and object-oriented change detection
analysis using high-resolution imagery. In: Proceedings of 25th Symposium on
Safeguards and Nuclear Material Management, Stockholm, May 1315.
Oruc, M., Marangoz, A.M., Buyuksalih, G., 2004. Comparison of pixel-based and
object-oriented classication approaches using Landsat-7 ETM spectral bands.
In: Proceedings of ISPRS Conference, Istanbul, July 1923.
Pearson, D.M., 2002. The application of local measures of spatial autocorrelation
for describing pattern in north Australian landscapes. Journal of Environmental
Management 64, 8595.
Specht, R.L., 1981. Foliage projective cover and standing biomass. In: Gillison, A.N.,
Anderson, D.J. (Eds.), Vegetation Classication in Australia: Proceedings of a

893

Workshop Sponsored by CSIRO Division of Land Use Research, October 1978.


CSIRO and ANU Press, Canberra.
Story, M., Congalton, R.G., 1986. Accuracy assessment: a users perspective. Photogrammetric Engineering and Remote Sensing 52 (3), 397399.
Whiteside, T., 2000. Multi temporal land cover change in the Humpty Doo region
NT, 19901997. Unpublished Master of Natural Resources Management Thesis.
University of Adelaide, Adelaide, 90 p.
Whiteside, T., Ahmad, W., 2004. Object-oriented classication of ASTER imagery
for landcover mapping in monsoonal northern Australia. In: Proceedings of
12th Australasian Remote Sensing and Photogrammetry Conference, Fremantle,
Spatial Sciences Institute, October 1822.
Willhauck, G., Schneider, T., De Kok, R., Ammer, U., 2000. Comparison of objectoriented classication techniques and standard image analysis for the use of
change detection between SPOT multispectral satellite images and aerial photos.
In: Proceedings of XIX ISPRS Congress, Amsterdam, July 1622.
Wilson, B.A., Brocklehurst, P.S., Clark, M., Dickinson, K.J.M., 1990. Vegetation Survey
of the Northern Territory, Australia, Report No. 49. Conservation Commission of
the Northern Territory, Palmerston.
Yamaguchi, Y., Kahle, A.B., Tsu, H., Kawakami, T., Pniel, M., 1998. Overview
of Advanced Spaceborne Thermal Emission and Reection Radiometer
(ASTER). IEEE Transactions on Geoscience and Remote Sensing 36, 1062
1071.
Yu, Q., Gong, P., Clinton, N., Biging, G.S., Kelly, M., Schirokauer, D., 2006. Objectbased detailed vegetation classication with airborne high spatial resolution
remote sensing imagery. Photogrammetric Engineering and Remote Sensing 72,
799811.