Вы находитесь на странице: 1из 49

Classification

and Information
Extraction

Natural Resources Ressources naturelles


CanadaCanada
Centre for Remote Sensing, Natural Resources Canada
Canada
Classification and Information
Extraction (Image Exploitation)
- Outline -
™ Classification Techniques
Supervised and Unsupervised Classification
Classification Algorithms
Accuracy Assessment
Maximum Likelihood Classification Example
New Classification Approaches

™ Change Detection
Difference Image
Ratio Image
Classification Comparison
Change Vector Analysis

Canada Centre for Remote Sensing, Natural Resources Canada


Classification and Information
Extraction (Image Exploitation)
- Outline -

™ Data Integration

RGB Colour Space

IHS Colour Space

Principal Component Analysis

Canada Centre for Remote Sensing, Natural Resources Canada


Introduction

™ Currently the majority of operational classification and


information extraction is performed using manual
interpretation approaches.
™ Manual approaches tend to be very time consuming
and expensive.
™ Several successful automated approaches are
operational, such as flood mapping.
™ Emerging techniques will increase the use of
automated approaches in the future.
™ This section reports on the automated quantitative
approaches using calibrated data.

Canada Centre for Remote Sensing, Natural Resources Canada


Introduction - Classification
™ Image classification categorizes image pixels into
classes producing a thematic representation.
™ Classification performed on single or multiple image
channels to separate areas according to their different
scattering or spectral characteristics.
™ Classified data can be used in thematic maps, imported
into a GIS or can be further incorporated into digital
analysis.
™ Thematic maps provide an interpretable summary of
classes enabling analysts to associate detection
capabilities of SAR imagery with terrain features.
™ Digital image classification procedures are differentiated
as being either supervised or unsupervised (clustering).

Canada Centre for Remote Sensing, Natural Resources Canada


Supervised Classification

™ Requires image analyst to “train” the computer to


recognize a set of pixels with similar signatures.
‹ Encompasses three components:
- training area selection
- classification
- post-classification analysis and accuracy assessment

™ Analyst determines the best classification scheme to


meet objectives and applies knowledge of the site
during the training process.
™ Figure 7.1 illustrates examples of supervised
classification.

Canada Centre for Remote Sensing, Natural Resources Canada


Figure 7.1 - Examples of Classifications

Canada Centre for Remote Sensing, Natural Resources Canada


Training Area Selection

™ Training areas are small samples of homogeneous


areas selected by the image analyst prior to
classification.
™ Appropriate training areas are determined from maps,
ground data, interpreted stereo airphoto or other
information.
™ Training areas should be:
‹ Free of anomalies
‹ Large enough to provide good statistical class
representation
‹ Sufficient in number to account for small local variations

Canada Centre for Remote Sensing, Natural Resources Canada


Training Area Selection
(cont'd)
™ Training areas should avoid:
‹ Edge pixels containing the combined backscatter of multiple
targets

‹ Inconsistencies within the area such as roadways, powerlines,


intermittent cover, etc.

™ Once defined, training areas are used to generate


signature statistics for each defined class.

™ Class signatures include class means and a class


covariance matrix.

Canada Centre for Remote Sensing, Natural Resources Canada


Supervised Classification

™ Figure 7.2 illustrates the problem with supervised


classification using linear boundaries for classes.
™ The separation of the major classes with a minimum
of error is possible with an n-dimensional decision
boundary.
™ The graph uses 1 band and 2 classes to illustrate how
the overlapping areas of both class 1 and class 2
distributions have erroneously classified pixels.

Canada Centre for Remote Sensing, Natural Resources Canada


Figure 7.2 - Supervised Classification

One dimensional
decision boundary
Number of pixels

Class 1 Class 2

Pixels in class 2 erroneously Pixels in class 1 erroneously


assigned to class 1 Assigned to class 2

Source: Jensen, 1996


Canada Centre for Remote Sensing, Natural Resources Canada
Classification Strategies and Algorithms
™ During classification, each pixel is compared to
each of the class signatures.
™ Comparison performed by computer using a
predetermined classification algorithm.
™ Most commonly used classifiers in remote
sensing are:
‹ Minimum Distance (to Means) Classifier
‹ Parallelepiped Classifier
‹ Maximum Likelihood Classifier (MLC)

™ Once a pixel has been assigned to a class, it is


given the class value in the corresponding cell
of the "classified" image.

Canada Centre for Remote Sensing, Natural Resources Canada


Minimum Distance (to Means) Classifier

™ Simplest algorithm and thus low computational


time.
™ Determines each pixel's "distance" from class
means, and assigns them to the closest class, see
Figure 7.3.
™ If pixel is further than the analyst defined distance
from any category, it remains unclassified or
"unknown”.
™ Classifier does not evaluate differing degrees of
within class variance, therefore has lower overall
accuracy than MLC.

Canada Centre for Remote Sensing, Natural Resources Canada


Figure 7.3 - Classification Algorithms

Channel B
Channel B

Channel A Channel A
Minimum distance Parallelepiped
classifier Classifier
Channel B

Channel A
Maximum Likelihood
Classifier

Source: PCI, Chapter 10, 1997

Canada Centre for Remote Sensing, Natural Resources Canada


Parallelepiped Classifier
™ Used when multi-band imagery is available.
™ Parallelepiped classifier is more sensitive to within
class variance.
™ Algorithm considers range of values within each
category of the training set, denoted as minimum and
maximum value for each image band (appears as a
rectangle in Figure 7.3).
™ Range limits define small decision region with clear
class segmentation compared to minimum distance
classifier.

Canada Centre for Remote Sensing, Natural Resources Canada


Parallelepiped Classifier
(cont’d)
™ Outliers can increase the decision region
inappropriately causing errors of commission.
™ Problems occur when classes overlap, as in Figure
7.3.
™ These pixels are labelled as overlap and are caused
from class distributions exhibiting correlations poorly
described by the rectangular decision regions.
™ Low computational requirement with adequate
classification accuracies.

Canada Centre for Remote Sensing, Natural Resources Canada


Maximum Likelihood Classifiers

™ Assumes that the training statistics for each class


have a normal or "Gaussian" distribution
‹ NOTE: Radar statistics are often non-”Gaussian”

™ Uses training statistics to compute a probability


value of whether it belongs to a particular land cover
category class
™ Training statistics with bi- or tri- modal histograms
are not suitable as they indicate non-homogeneity
within classes and are non-”Gaussian”

Canada Centre for Remote Sensing, Natural Resources Canada


Maximum Likelihood Classifiers
(cont'd)

™ Examines the probability function of a pixel for each of


the classes, and assigns the pixel to the class with the
highest probability.

™ Usually provides the highest classification accuracies.

™ Larger number of computations required to classify


each pixel, resulting in a high computational
requirement.

™ Can use a-priori knowledge to weight probability


function.

Canada Centre for Remote Sensing, Natural Resources Canada


Unsupervised Classification
(Clustering)
™ Unsupervised classification does not require training areas
or analyst's knowledge of area
™ Creates natural groupings present in the image values
™ Values with similar grey levels are assumed to belong to
the same cover type
™ Analyst must determine the identity of the computer
derived spectral clusters
™ Principal clustering algorithms include
‹ K- means clustering
‹ ISODATA clustering
‹ Narendra-Goldberg clustering

™ See figure 7.1 for an example of unsupervised


classification
Canada Centre for Remote Sensing, Natural Resources Canada
Post-Classification Filtering

™ Resulting classification image map may be difficult to


interpret.
™ Classified data have a salt-and-pepper appearance
due to inherent variability of the per-pixel classifier.
™ Post-classification filtering removes pixels and pixel
groups not satisfying a minimum requirement.
™ Figure 7.4 is an example of a mode post-classification
filter, where the pixel is reassigned to the
surrounding class majority pixels.
™ Post-classification filtering usually enhances
interpretability and increases classification accuracy.

Canada Centre for Remote Sensing, Natural Resources Canada


Figure 7.4
Post-Classification Mode Filtering

Pixel (7,4) Mode


and Filtered
Classified Image 3x3 window Result

Source: PCI, Chapter 10, 1997

Canada Centre for Remote Sensing, Natural Resources Canada


Classification Accuracy Assessment
™ Evaluate accuracy of the classification procedure by checking against
known homogeneous areas.
™ Overall accuracy vs. accuracy by class.
™ Results of the accuracy can be displayed in a confusion matrix, such as
Figure 7.5.
™ Confusion matrix plots known pixels against classified pixels.
™ Errors of Commission
‹ Pixels incorrectly assigned to a particular class that actually belong in other
classes, see the lower left half of confusion matrix in Figure 7.5.

™ Errors of Omission
‹ Pixels incorrectly excluded from a particular class, see the upper right half of
confusion matrix in Figure 7.5.

Canada Centre for Remote Sensing, Natural Resources Canada


Figure 7.5
Maximum Likelihood Confusion Matrix
Classification of 3 bands: C-HH, C-HV, C-VV
Bare Soy
Grains Corn Water Forest Urban
Soil beans
Grains 60.6 13.0 26.2 0.1 0.0 0.1 0.0

Corn 1.8 70.7 25.9 0.2 0.0 1.4 0.0

Bare Soil 2.3 17.1 80.1 0.3 0.0 0.2 0.0

Soy beans 0.0 0.0 0.3 96.0 0.0 0.2 3.4

Water 0.0 0.0 0.0 0.0 98.3 0.0 1.7

Forest 17.8 15.9 14.7 2.6 0.0 47.8 1.2

Urban 0.1 0.0 1.1 5.0 0.0 1.8 91.9

Average Accuracy: 77.92% Kappa Coefficient: 0.61889


Overall Accuracy: 70.09% Standard Deviation: 0.00391

Canada Centre for Remote Sensing, Natural Resources Canada


New Classification Approaches
Introduction

™ Supervised and unsupervised classifications generally


use "per-pixel" approach.

™ Due to radar speckle, SAR classification is often done


on a per field or polygon basis using either the
thresholding technique or polygon averages.

™ Newer classification methods associate pixels with


their surrounding neighbours similar to classification
performed by human visual interpretation.

Canada Centre for Remote Sensing, Natural Resources Canada


New Classification Approaches

™ NEURAL NETWORKS
‹ Building block design algorithm imitating a "human"
decision-making process to classification.
‹ Do not make assumptions about the underlying distribution
of the data
‹ Uses both spectral and textural patterns in the classification
process.
‹ Major advantage is that it can identify subtle and non-linear
patterns that traditional classifiers do not detect.
‹ Problem of neural networks is that it can be very difficult to
train.

Canada Centre for Remote Sensing, Natural Resources Canada


New Classification Approaches

™ FUZZY LOGIC
‹ Simulates vagueness or uncertainty encountered in nature

‹ Categorizes data according to non-discrete class structure

™ CONTEXTUAL CLASSIFIERS
‹ Classification of a pixel is influenced by the class(es)
assigned to its neighbours
‹ Pixel is examined in "context" to surrounding pixels

™ Numerous other specialized classifiers available.

Canada Centre for Remote Sensing, Natural Resources Canada


Change Detection
™ Change detection methodologies useful for
‹ Urbanization
‹ Agricultural Development
‹ Forest Land Management
‹ Ice Forecasting, etc.

™ Utilizes two or more scenes covering same geographic


area acquired over a temporal period
‹ Channels of data from one pass or one instrument
‹ Two different passes, same radar, same scene, same mode
‹ Two different passes, different radars, same scene

Canada Centre for Remote Sensing, Natural Resources Canada


Change Detection
(cont'd)
™ Spaceborne SAR is ideal due to high revisit
capability.
™ Must consider SAR properties:
‹ Imaging geometry
‹ Relief displacement
‹ Image to image registration
‹ Calibration requirements.

™ Must also consider:


‹ Environmental conditions (precipitation, moisture
conditions of vegetation and soil, snow cover, etc.)
‹ Time of year.

Canada Centre for Remote Sensing, Natural Resources Canada


Change Detection
(cont'd)

™ Change Detection Methodologies:


‹ Difference Image
‹ Ratio Image
‹ Classification Comparison
‹ Change Vector Analysis

™ Figure 7.6 is an example of change detection using


different colors for each radar image

Canada Centre for Remote Sensing, Natural Resources Canada


Figure 7.6 - Change Detection

Canada Centre for Remote Sensing, Natural Resources Canada


Difference Image
™ This task should be done in the power domain.
™ This task should only be carried out on "sets of pixels"
or a "per field basis”.
™ Difference image is created by subtracting the mean
value of parcels of pixels in two different images of
the same area.
™ Results in either a positive or negative value where
change has occurred.
™ Zero values indicate parcels of no change.
™ Must consider threshold boundaries between change
and no-change.
™ Must consider calibration issues.

Canada Centre for Remote Sensing, Natural Resources Canada


Ratio Image
™ Ratio image is created by dividing the mean value of parcels of
pixels in two different images of the same area.
™ For SAR imagery, ratios can only be constructed from multi-look
images and should be in the power domain.
™ Band ratios deliver the combined information content of two
image bands.
™ Ratios can help minimize unwanted information and/or noise.
™ System and processing effects must be considered when
producing ratios.
™ Ratio images must be scaled to produce an acceptable product
for visual interpretation.
™ Interpreting ratio images requires a knowledge of target
reflectance illumination and ground conditions.

Canada Centre for Remote Sensing, Natural Resources Canada


Classification (Thematic) Comparison

™ Compare the thematic output from two or more


sources.
™ Identifies areas of change and the nature of change
(e.g. from agriculture to urban).
™ Accuracy depends upon initial classification
accuracies of input imagery; any errors are
compounded.

Canada Centre for Remote Sensing, Natural Resources Canada


Change Vector Analysis

™ Uses spectral or spatial differences to detect a change


or disturbance.
™ SAR requires data sets acquired at separate times.
™ Plotted against each other on a graph, the two spectral
variables will show the magnitude and direction of
change from the 1st to the 2nd date, see “A” in Figure
7.7.
™ The vector describing direction and magnitude of
change from the 1st to the 2nd date is the spectral
change vector.
™ Decision that a “change” has occurred is made if a
threshold is exceeded, as in “C” and “D” in Figure 7.7.
Source: Jensen, 1996

Canada Centre for Remote Sensing, Natural Resources Canada


Figure 7.7: Change Vector Analysis
A
SPECTRAL CHANGE
VECTOR

SPECTRAL VARIABLE Y
MAGNITUDE
OF CHANGE

ANGLE OF CHANGE

SPECTRAL VARIABLE X

C D
B
LITTLE OR CHANGE CHANGE
NOT CHANGE (e.g. CLEARED FOR (e.g. REGROWTH OF
SUBDIVISION) NATURAL VEGETATION)
YEAR 2

YEAR 2
YEAR 2
YEAR 1 YEAR 1
YEAR 1

THRESHOLD THRESHOLD DECISION


THRESHOLD

Source: Jensen,
Source: 1996 1996
Jensen,

Canada Centre for Remote Sensing, Natural Resources Canada


Data Integration

™ Remotely sensed data can contribute in a variety of


ways to resource management activities involving
multiple data sets.
™ Multiple data sets provide
‹ Spatial continuity and geometric flexibility
‹ Multi-temporal coverage
‹ Complete coverage regardless of site location and access
‹ Digital data facilitates custom image analysis and output
‹ Synergism between data sets

Canada Centre for Remote Sensing, Natural Resources Canada


Data Integration
(cont'd)
™ Remotely sensed data can contribute in several
forms
‹ Raster image providing continuous detail and an
accurate base

‹ Polygon data extracted by classification or visual


interpretation

‹ Vector data extracted by enhancement or visual


interpretation

™ Figure 7.8 is an example of multiple data sets


integration.

Canada Centre for Remote Sensing, Natural Resources Canada


Figure 7.8 GIS Data Integration

Canada Centre for Remote Sensing, Natural Resources Canada


Integrating Remotely Sensed Data

™ Synergism between remotely sensed data acquired


at different times, wavelength, resolutions, etc. can
increase the useful information content.

™ Multi-temporal data are often used to take advantage


of seasonal or phenological changes in vegetation
and for change detection.

™ Multi-sensor data can make use of the different


information as a function of wavelength, resolution
and/or scale.

Canada Centre for Remote Sensing, Natural Resources Canada


Integrating Remotely Sensed Data
(cont'd)
™ Multi-channel data can make use of information within
different regions of the EM spectrum.

™ Multi-polarization data can make use of different information


in the microwave band related to target interaction with the
radar waves.

™ Polarimetric SAR data can make use of phase as well as


magnitude.

™ Multi-feature data can make use of different information


from the same scene, for example, tone and texture.

™ See Figure 7.9 for an example of SAR and TM integration

Canada Centre for Remote Sensing, Natural Resources Canada


Figure 7.9 - Multi-Sensor Combinations

Canada Centre for Remote Sensing, Natural Resources Canada


RGB Colour Spaces
™Red, green, blue (RGB) colour space is based
upon the additive properties of primary colours.
™System is optimised for computer video screens
but not for human vision.
™Multi-channel data is displayed using RGB
technique where each channel is assigned a colour
with the intensity related to the magnitude of the
spectral data.
™The RGB colour cube is shown in Figure 7.10.
™The colour cube shows the interrelationships
between the colours and is defined by the
brightness levels of each of red, green and blue.

Canada Centre for Remote Sensing, Natural Resources Canada


Figure 7.10 - RGB Colour Spaces

Blue Cyan

Magenta

White

e
Sc al
y
Gra

Black Green

Yellow
Red

Adapted from: Schowengerdt, 1983 RGB


Canada Centre for Remote Sensing, Natural Resources Canada
IHS Colour Space

™ Intensity-Hue-Saturation (IHS) colour space is an


alternative way to describe colours by their RGB
components, see Figure 7.11.
™ IHS converts three bands into an alternative colour
space closer to what the human eye perceives them.
™ IHS is more adapted to human vision than the more
standard RGB colour space.
™ The 3 bands can be from different sensors, such as
RADARSAT and Landsat TM.

Canada Centre for Remote Sensing, Natural Resources Canada


Figure 7.11 - IHS Color Space
HUE
Green Yellow
120º 60º Red

White
Cyan
180º White
Blue Magenta
300º

INTENSITY
INTENSITY

Black
Black

SATURATION

No colour Full colour

IHS
Intensity - Hue - Saturation

Adapted from: Drury, S.A. Image Interpretation in Geology, Second Edition, 1993.
Chapman & Hall, p.135.
Canada Centre for Remote Sensing, Natural Resources Canada
IHS Colour Space
(cont’d)
™ Intensity is the colour brightness, hue is the actual
colour and saturation defines the purity or "greyness"
of the colour.
™ A common approach is to modulate the intensity
channel using a SAR image, with other data
(geophysics, geochemistry, visible/infra-red image)
modulating hue and a flat image replacing saturation.
™ IHS can improve image sharpness and edge
extraction.
™ The IHS image in Figure 7.12 modulates the intensity
and the hue channels using a SAR image and a
DEM.

Canada Centre for Remote Sensing, Natural Resources Canada


Figure 7.12 - IHS Transform

Canada Centre for Remote Sensing, Natural Resources Canada


Principal Component Analysis
™ Requires multi-channel data set (e.g. multi-date, multi-
polarization, multi-sensor).
™ Can be used to identify new axes that maximize variance
in the data set (see Figure 7.13).
™ Reduces the dimensionality of the multi-channel input to
the dimension of the information content.
™ Any features or patterns identified on a PCA should be
confirmed through interpretation of supporting image
products and other more conventional data sets.
™ Eigen Vectors (or principal component channels) usually
do not transfer well between data sets.

Canada Centre for Remote Sensing, Natural Resources Canada


Figure 7.13 - Axes Rotation Along the
Principal Component Vectors
SCATTER
PLOT FREQUENCY
B1 VS B2 HISTROGRAMS

Band 1 BAND 1

BAND 2

Band 2

FREQUENCY
Component 1
HISTROGRAMS
IDENTIFY NEW AXES WHICH
MAXIMIZE VARIANCE IN THE
DATA SET. E´´

B1 E´´
Composante 2

B2 E´
Rotation of axes

Source: CCRS

Canada Centre for Remote Sensing, Natural Resources Canada

Вам также может понравиться