Remote Sensing and

GIS

A Ravi Raja M.Tech, (Ph.D)

[email protected]
☎ +91 9493149772

https://www.youtube.com/watch?reload=9&v=KCy6o5gqaNA
Unit-III
 Unit-III
Topics Chapter/ Section Page Number
No.
Introduction 7.1 211
Basic Character of Digital Image 7.2 211
Preprocessing 7.3 216
Geometric Correction Methods 7.3.1 216
Radiometric Correction Methods 7.3.2 219
Atmospheric Correction Methods 7.3.3 224
Image Registration 7.4 225
Conversion of Geographical Coordinates to Conical Orthomorphic 7.4.1 226
Coordinates
Transformation of Conical Orthomorphic Coordinates to Digital 7.4.2 227
Imagery Coordinates
Image Enhancement Techniques 7.5 229
Contrast Enhancement 7.5.1 229
Spatial Filtering Techniques 7.6 236
Low Pass Filters 7.6.1 236
High Pass Filters 7.6.2 240
Filtering for Edge Enhancement 7.6.3 241
Image Transforms 7.7 244
NDVI Transformation 7.7.1 245
PCA Transformation 7.7.2 248
 Unit-III

Topics Chapter/ Section Page Number

No.
Image Classification 7.8 250
Supervised Classification 7.8.1 251
Training Dataset 7.8.2 254
Unsupervised Classification 7.8.3 256
Performance Analysis of IRS-bands for land use/land cover 7.9 256
classification system using Maximum Likelihood Classifier
Classification Methodology 7.9.1 257
The Land Use and Land Cover Classification System 7.9.2 258
Data Analysis 7.9.3 259
Classification Accuracy Approach 7.9.4 259
Image Classification and GIS 7.10 261
 CHAPTER-VII

DIGITAL IMAGE PROCESSING

Source: World Wide Web

 6.1 INTRODUCTION

 The remote sensing data can be analysed using visual image interpretation techniques if
the data are in the hardcopy or pictorial form.

 It is used extensively to locate specific features and conditions, which are then
geocoded for inclusion in GIS . Visual image interpretation techniques have certain
disadvantages and may require extensive training and are labour intensive.

 In this technique, the spectral characteristics are not always fully evaluated because of
the limited ability of the eye to discern tonal values and analyse the spectral changes.

 If the data are in digital mode, the remote sensing data can be analysed using digital
image processing techniques and such a database can be used in raster GIS.

 In applications where spectral patterns are more informative, it is preferable to analyse

digital data rather than pictorial data.
Pixel 6.2 BASIC CHARACTER OF DIGITAL IMAGE

Pixel values typically represent gray levels,

colors, heights, opacities etc…

7
Gray Level Image

8
 The following figure is graphical representation of the A to D conversion process. The
original electrical signal from the sensor is a continuous analog signal shown by the
continuous line plotted in the figure.

 This continuous signal is sampled at a set time interval (~T) and recorded numerically
at each sample point (a, b, ..... , i, k). The sampling rate for a particular signal is
determined by the least to twice the highest frequency presents in the original signal in
order to adequately represent the variation in the signal.
9
 A digital image is defined as a matrix of digital numbers (DNs).

 Each digital number is the output of the process of analog to digital conversion.

 The image display subsystem carries out the conversion operation, that of taking a
digital quantity of an individual pixel value from an image.

 Fig. 6.3 illustrates the surface of the ground divided into a number of parcels. Each
parcel of land can be represented as a pixel (picture element) on the image and each
pixel is occupied by a digital number and is called pixel value. This pixel value or
digital number shows the radiometric resolution of remote sensing data.

10
11
 Image Formation

f(x,y) = reflectance(x,y) * illumination(x,y)

Reflectance in [0,1], illumination in [0,inf]

12
SAMPLING AND QUANTIZATION

13
 6.3 PREPROCESSING

 Remotely sensed raw data, received from imaging sensor mounted on satellite
platforms generally contain flaws and deficiences.
 The correction of deficiences and removal of flaws present in the data through
some methods are termed as pre-processing methods.
 This correction model involves the initial processing of raw image data to correct
geometric distortions, to calibrate the data radiometrically and to eliminate the
noise present in the data.
 All pre-processing methods are considered under three heads, namely,
 (i) geometric correction methods,
 (ii) radiometric correction methods, and
 (iii) atmospheric correction methods.

14
 6.3.1 GEOMETRIC CORRECTION METHODS

 Remotely sensed images are not maps (Mather, 2000).

 Frequently information extracted from remotely sensed images is Integrated with
map data in a geographical information system.
 The transformation of a remotely sensed image into a map with a scale and
projection properties is called “geometric correction”

15
 6.3.1 GEOMETRIC CORRECTION METHODS

 Geometric correction of remotely sensed images is required when the image or

product derived from the image such as a vegetation index or a classified image, is
to be used in one of the following circumstances:
 * to transform an image to match a map projection
 * to locate points of interest on map and image
 * to bring adjacent images into registration
 * to overlay temporal sequences of images of the same area, perhaps
acquired by different sensors
 * to overlay images and maps within GIS, and
 * to integrate remote sensing data with GIS.

16
 6.3.1 GEOMETRIC CORRECTION METHODS
 To correct sensor data, both internal and external errors must be determined and
be either predictable or measurable.
 Internal errors are due to sensor effects, being systematic or stationary, or,
constant for all practical purposes.
 External errors are due to platform perturbations and scene characteristics, which
are variable in nature and can be determined from ground control and tracking
data.
 The sources of image geometric errors are listed in Table 6.1

17
 6.3.2 RADIOMETRIC CORRECTION METHODS
 The primary function of remote sensing data quality evaluation is to monitor the
performance of the sensors.

 The performance of the sensors is continuously monitored by applying

“radiometric correction models” on digital Image data sets.

 The radiance measured by any given system over a given object is influenced by
factors, such as, changes in scene illumination, atmospheric conditions, viewing
geometry and instrument response characteristics.

 One of the most important radiometric data processing activity involved in many
quantitative applications of digital image data is conversion of digital numbers to
absolute physical values; namely, radiance and reflectance.

18
 6.3.2 RADIOMETRIC CORRECTION METHODS
 Computation of Radiance (L)
 Radiance is a measure of the radiant energy given out by an object and picked up
by a remote sensor Fig. (6.4).
 Spectral radiance (L) is defined as the energy within a wavelength band radiated
by a unit area per unit solid angle of measurement.

19
 6.3.2 RADIOMETRIC CORRECTION METHODS
 Radiance in a single band is calculated by using the following formula

20
 6.3.2 RADIOMETRIC CORRECTION METHODS
 Computation of Reflectance
 Unlike radiance which is a measure of radiant energy, reflectance is an energy
ratio.
 It is a function of radiance and is defined by the following formula:

21
 6.3.2 RADIOMETRIC CORRECTION METHODS
 Cosmetic Operations
 The first is the correction of digital images containing either partially or entirely missing
scan lines.
 The second is the correction of images because of destripping of the imagery.
 For Example, LANDSAT MSS has six detectors for each band. When it scans, it scans six
lines at a time per band.
 It is expected that all the six detectors should record similar signatures from the same
object on the earth surface. But, in practice, it was noticed that the sensitivity of all
the detectors is not uniform. This means sometimes detector recorded irradiance
(reflected) for the same object may differ.
 This causes the stripping in remote sensing imagery. The second phenomenon is called line
drop.
 The line dropout or missing scan line is usually overcome by replacing the zero value by
the mean values of the pixels of the previous and the following line.
 For example, if the10th line is the dropout line, then all the pixels in the 10 th line will be
replaced by the corresponding mean of the 9th and the 11 th line.

22
6.3.2 RADIOMETRIC CORRECTION METHODS

23
 6.3.2 RADIOMETRIC CORRECTION METHODS
 Random Noise Removal

24
 6.3.2 RADIOMETRIC CORRECTION METHODS
 Random Noise Removal

25
 6.3.3 ATMOSPHERIC CORRECTION METHODS
 according to Rayleigh scattering , the effect of scattering is inversely proportional
to the fourth power of wavelength of energy,
 That is, scattering is more in the lower wavelength (visible) than in the higher
wavelength (infrared band).

26
 6.4 IMAGE REGISTRATION
 Image registration is the translation and rotation alignment process by which two images/maps of like
geometries and of the same set of objects are positioned coincident with respect to one another so that
corresponding elements of the same ground area appear in the same place on the registered images.
 This is often called image-to-image registration.
 One more important concept with respect to geometry of satellite image is rectification.
 Rectification is the process by which the geometry of an image area is made planimetric.
 It may not remove the distortion caused by topographic replacement In images.
 This process almost always involves relating Ground Control Points (GCP) pixel coordinates (row and
column) with map coordinate counterparts.
 This is the most precise geometric correction since each pixel can be referenced not only by its row
and column in a digital image matrix after rectification is completed, but it is also rigorously
referenced in degrees, feet or meters In a standard map projection.
 Whenever accurate data, direction and distance measurements are required, geometric rectification is
required. This is often called as an image-to-map rectification.

27
6.4 IMAGE REGISTRATION

28
 6.5 IMAGE ENHANCEMENT TECHNIQUES
 Low sensitivity of the detectors, weak signal of the objects present on the earth
surface, similar reflectance of different objects and environmental conditions at the
time of recording are the major causes of low contrast of the image.
 The main aim of digital enhancement is to amplify these slight differences for better
clarity of the image scene.
 This means digital enhancement increases the separability (contrast) between the
interested classes or features.
 The digital image enhancement may be defined as some mathematical operations that
are to be applied to digital remote sensing input data to improve the visual
appearance of an image for better interpretability or subsequent digital analysis

29
 6.5 IMAGE ENHANCEMENT TECHNIQUES
6.5.1 CONTRAST ENHANCEMENT
 In remote sensing literature, many digital enhancement algorithms are available.
 They are contrast stretching enhancement, ratioing , linear combinations, principal
component analysis, and spatial filtering
 Broadly, the enhancement techniques are categorised as point operations and local
operations.
 Point operations modify the values of each pixel in an image data set independently,
 whereas local operations modify the values of each pixel in the context of the pixel
values surrounding it.

30
 6.5 IMAGE ENHANCEMENT TECHNIQUES
6.5.1 CONTRAST ENHANCEMENT
 In linear contrast stretch, the digital number (DN) value in the lower end of the
original histogram is assigned to digital number zero (DN = 0) that is, extremely
black, and a value at the higher end is assigned to extremely white (that is ON =
127).
 The intermediate values are interpolated between 0 and 255 by following a linear
relationship,
 Y = a + bX
 Where X and Y are the input gray value of any pixel and output gray value of the
same pixel.
 a and b are intercept and slope respectively.

 The linear contrast stretch is widely used to improve the contrast of most of the
original brightness values, but there is a loss of contrast of the extreme high and low
ends at the tail of the histogram

31
 6.5 IMAGE ENHANCEMENT TECHNIQUES
6.5.1 CONTRAST ENHANCEMENT
 The logarithmic contrast enhancement is very much useful for non-linear contrast
enhancement. Here the output pixel grey values (Yij) will be generated from input
pixel grey values (Xij) following some logarithmic expressions, as follows :
 Yij = a log (Xij) + b
 The coefficients 'a' and 'b' are determined by taking the maximum and minimum grey
values of the input image and the corresponding maximum and minimum values in
the output image. This transformation only highlights the features lying in the darker
region of the histogram
 The advantages and important characteristics of logarithmic contrast enhancement
are;
 (i) It makes the low contrast details more visible by enhancing low contrast
edges;
 (ii) It provides a contrast signal to noise ratio;
 (iii) It some what matches the responses of the human visual systems;
 (iv) It usually provides a more equal distribution of gray values and
 (v) It transforms multiplicative noise into additive noise.

32
 6.5 IMAGE ENHANCEMENT TECHNIQUES
6.5.1 CONTRAST ENHANCEMENT
  Exponential contrast enhancement is also considered as a non-linear contrast
enhancement. The grey values (Xij) in the input image transform to the gray values
(Yij) in the output image is as follows:

 Where a, band c are constants, b is arbitrarily chosen between 0.01 and 0.1 to avoid
higher value of e. Further, 'a' and . b' scaie the dynamic range of the grey values of
the output image with 0 and 255.
 The effect of the exponential transformation on the edges in an image is to compress
low contrast edges, while expanding high contrast edges. This generally highlights
the feature having higher grey values. The effect is just reverse of logarithmic
contrast transformation.

33
 6.5 IMAGE ENHANCEMENT TECHNIQUES
6.5.1 CONTRAST ENHANCEMENT
 Histogram equalization is widely used for contrast manipulation in digital image
processing because it is very simple for implementation. It needs minimum
information from the analyst. In this enhancement, the original histogram has been
readjusted to produce a uniform population density of pixels along the horizontal
grey value (DN) axis.

34
 6.5 IMAGE ENHANCEMENT TECHNIQUES
6.5.1 CONTRAST ENHANCEMENT
 The other method of contrast enhancement based on the histogram of input pixel
values and their manipulation is called Gaussion stretch. This is called Gaussion
stretch because it involves fitting of Normal Histogram to Gaussion Histogram.

35
 6.6 SPATIAL FILTERING TECHNIQUES
 A characteristic of remotely sensed images is a parameter called spatial frequency,
defined as the number of changes in brightness values per unit distance for any
particular part of an image.
 If there are few changes in brightness value over a given area it is termed as a low
frequency area. If the brightness values changes dramatically over very short
distances, this is called high frequency area.
 Algorithms which perform image enhancement are called filters because they
suppress certain frequencies and pass (emphasise) others.
 Filters that pass high frequencies while emphasizing fine detail and edges called high
frequency filters, and filters that pass low frequencies called low frequency filters.

36
 6.6 SPATIAL FILTERING TECHNIQUES

 Filtering is performed by using convolution windows. These windows are called

mask, template filter or kernel.
 In the process of filtering, the window is moved over the input image from extreme
top left hand corner of the scene.
 The discrete mathematical function transforming the original input image digital
number to a new digital value.
 First it will move along the line. As soon as the line is complete, it will restart for the
next line for covering the entire image
 The mask window may be rectangular (1 x 3, or 1 x 5 pixels) size or square (3 x 3, 5
x 5 or 7 x 7 pixels size). Each pixel of the window is given a weightage.
 For low pass filters all the weights in the window will be positive and for high pass
filter all the values may be negative or zero, but the central pixel will be positive
with higher weightage value. In the case of high pass filter the algebraic sum of all
the weights in the window will be a zero. Many types of mask windows of different
sizes can be designed by changing the size and varying weightage within the
window. The simplest form of mathematical function performed in filtering operation
is neighbourhood averaging.

37
 6.6 SPATIAL FILTERING TECHNIQUES
6.6.1 LOW PASS FILTERS
 low-pass filter to smooth the image in the process of registration with the technique
of mean filtering each pixel is sequentially examined and if the pixel digital number
(DN) is greater than the average brightness (DN) of its surrounding pixels by some
threshold (t), it is replaced by the means of the 3 x 3 pixels window.
 The window may be of any size, such as 5 x 5, 7 x 7 and so on . The larger the
window size, the more will be computational time. The following window
demonstrates the comparison of computation time of varying window.
 The central pixel 'X' has no correlation with the neighboring pixels and is replaced by
the mean value of surrounding pixels.

38
39
 6.6 SPATIAL FILTERING TECHNIQUES
6.6.1 LOW PASS FILTERS
 Median Filter
 An alternative type of smoothing filter utilizes the median of the neighborhood rather
than the mean.
 The median filter is generally thought to be superior to the moving average filter for
two reasons.
 First, the median of a set of 'n' numbers is always one of the data values present in
the set, when n is an odd integer.
 Secondly, the median is less sensitive to errors or to extreme data values. Let us
consider a set of nine pixel values in the digital image, (2, 1, 25 , 7, 28, 5, 8, 30, 82)
(neighborhood) of 3 x 3 window, thus the median is the central value when the data
are ranked in an ascending or descending order of magnitude. In this example the
ranked values are { 1, 2, 5, 7, 8, 25 , 28, 30, 82} giving a median value of 8. The
mean 33.10, would be rounded up to the value of 33. The value 33 is not present in
the original data, unlike the median value 8.
 Also the mean value is larger than 8 observed values and may be thought to be
unduly influenced by the extreme data values, which might represent noise, are
removed by the median filter. The median filter preserves edges better than a mean
filter.
40
 6.6 SPATIAL FILTERING TECHNIQUES
6.6.2 HIGH PASS FILTERS
 High Pass Filters
 A simple high pass filter may be implemented by subtracting a low pass filtered
image (pixels by pixel) from the original, unprocessed image. The high frequency
component image enhances the spatial detail in the image at the expense of the large
area brightness information. High pass filtering can be performed by means of image
subtraction method or derivative based methods.

41
 6.6 SPATIAL FILTERING TECHNIQUES
6.6.3 FILTERING FOR EDGE ENHANCEMENT
 Edge or boundary between two features is very important for separating the different
features.
 Edge is characterized by high frequencies.
 In nature, edge between two features or objects may not always be distinct.
 hus edge enhancement or edge crispening is required for better interpretation of an
image.
 Edge enhancement filtering technique enhances the edge only.
 High frequencies signify the large degree of tonal variation within a small area or
within a small spatial distance.

42
 6.6 SPATIAL FILTERING TECHNIQUES
6.6.3 FILTERING FOR EDGE ENHANCEMENT

43
 6.6 SPATIAL FILTERING TECHNIQUES
6.6.3 FILTERING FOR EDGE ENHANCEMENT

44
 6.6 SPATIAL FILTERING TECHNIQUES
6.6.3 FILTERING FOR EDGE ENHANCEMENT

45
 6.7 IMAGE TRANSFORMATIONS

 The term 'transform' is used a little loosely in this chapter, for the arithmetic operators
of addition, subtraction, multiplication, and division are included although they are
not strictly transformations.
 All the transformations in image processing of remotely sensed data allow the
generation of a new image based on the arithmetic operations, mathematical statistics
and fourier transformations.
 The new image or a composite image is derived by means of two or more band
combinations, arithmetic's of various band data individually and/or application of
mathematics of multiple band data.
 The resulting image may well have properties that make it more suited to a particular
purpose than the original. For example, the division of band 1 and band 2 (B1/B2)

46
 6.7 IMAGE TRANSFORMATIONS
6.7.1 NDVI TRANSFORMATION

 The remote sensing data is used extensively for large area vegetation monitoring.
 Typically the spectral bands used for this purpose are visible and near IR bands.
 Various mathematical combinations of these bands have been used for the
computation of NDVI, which is an indicator of the presence and condition of green
vegetation.
 These mathematical quantities are referred to as Vegetation Indices. There are three
such indices, Simple Vegetation Indices, Rational Vegetation Indices, and Normalised
Differential Vegetation Indices.

47
 6.7 IMAGE TRANSFORMATIONS
6.7.1 NDVI TRANSFORMATION

 This NDVI is bounded ratio that ranges between -1 to +1. Clouds, water and snow
have negative NDVI since they are more reflective in visible than near IR wave
lengths.
 Soil and rock have a broadly similar reflectance giving NDVI close to '0', Only active
vegetation has a positive NDVI being typically between about 0.1 and 0.6 values at
the higher end of the range indicating increased photosynthetic activity and a greater
density of the canopy.
 This is explained by applying or transforming the digital numbers/ spectral
reflectance values/pixel values of NDVI using the formula given above for a part of
Hyderabad city by adopting the following algorithm:
 (i) Computation of NDVI values for the entire study area by conversion of spectral
reflectance values into NDVI values.
 (ii) Conversion of these NDVI values to a scaled channel values by using density
slicing method that measures apparent reflectance to sensor values.
 (iii) Display of image with NDVI and creation of a legend keeping the threshold
values and the ranges that are shown in the Plate 4.

48
6.7 IMAGE TRANSFORMATIONS
6.7.1 NDVI TRANSFORMATION

49
 6.7 IMAGE TRANSFORMATIONS
6.7.2 PCA TRANSFORMATION
 Principal components analysis, also referred to as PCA, has been proven to be of a
significant value in the analysis of remotely sensed digital data.
 The application of PCA on raw remote sensing data produces a new image which is
more interpretable than the original data.
 The main advantage of PCA is that it may be used to compress the information
content of a number of bands into just two or three transformed principal component
images.
 To perform PCA we apply a transformation to a corrected set of multispectral data.
 Adjacent bands in multispectral remotely sensed images are generally correlated.
Multiband visible / near infrared images of vegetated areas will show negative
correlation between the near-infrared and visible bands and positive correlations
among visible bands

50
6.7 IMAGE TRANSFORMATIONS
6.7.2 PCA TRANSFORMATION

51
 6.8 IMAGE CLASSIFICATION
 Image classification is a procedure to automatically categorize all pixels in an image of a
terrain into land cover classes.
 Normally, multispectral data are used to perform the classification of the spectral pattern
present within the data for each pixel is used as the numerical basis for categorization.
 This concept is dealt under the broad subject, namely, Pattern Recognition.
 Spectral pattern recognition refers to the family of classification procedures that utilizes this
pixel-by-pixel spectral information as the basis for automated land cover classification.
 Spatial pattern recognition involves the categorization of image pixels on the basis of the
spatial relationship with pixels surrounding them.
 image classification techniques are grouped into two types, namely
 supervised and unsupervised.
 The classification process may also include features,such as, land surface elevation and the
soil type that are not derived from the image.
 A pattern is thus a set of measurements on the chosen features for the individual to be
classified.
 The classification process may therefore be considered a form of pattern recognition, that is,
the identification of the pattern associated with each pixel position in an image in terms of
the characteristics of the objects or on the earth's surface.

52
 6.8 IMAGE CLASSIFICATION
6.8.1 SUPERVISED CLASSIFICATION

53
 6.8 IMAGE CLASSIFICATION
6.8.1 SUPERVISED CLASSIFICATION

 A supervised classification algorithm requires a training sample for each class, that is,
a collection of data points known to have come from the class of interest.
 The classification is thus based on how "close" a point to be classified is to each
training sample. We shall not attempt to define the word "close" other than to say that
both geometric and statistical distance measures are used in practical pattern
recognition algorithms.
 The training samples are representative of the known classes of interest to the analyst.
 Classification methods that relay on use of training patterns are called supervised
classification methods.
 The three basic steps (Fig. 6.23) involved in a typical supervised classification
procedure are as follows :

54
 6.8 IMAGE CLASSIFICATION
6.8.1 SUPERVISED CLASSIFICATION

 (i) Training stage: The analyst identifies representative training areas and develops
numerical descriptions of the spectral signatures of each land cover type of interest in
the scene.
 (ii) The classification stage: Each pixel in the image data set IS categorized into the
land cover class it most closely resembles. If the pixel is insufficiently similar to any
training data set it is usually labeled 'Unknown'.
 (iii) The output stage: The results may be used in a number of different ways. Three
typical forms of output products are thematic maps, tables and digital data files which
become input data for GIS. The output of image classification becomes input for GIS
for spatial analysis of the terrain. Fig. 6.24 depicts the flow of operations to be
performed during image classification of remotely sensed data of an area which
ultimately leads to create database as an input for GIS. Plate 6 shows the land use/
land cover color coded image, which is an output of image classification.

55
 6.8 IMAGE CLASSIFICATION
6.8.1 SUPERVISED CLASSIFICATION

56
 6.8 IMAGE CLASSIFICATION
6.8.2 TRAINING DATASET

 A training dataset is a set of measurements (points from an image) whose category

membership is known by the analyst.
 This set must be selected based on additional information derived from maps, field
surveys, aerial photographs, and analyst's knowledge of usual spectral signatures of
different cover classes.
 Selecting a good set of training points is one of the most critical aspects of the
classification procedure.
 However, an attempt is made to provide some guidelines based on the author's
experience. These guidelines are as following:
 (i) Select sufficient number of points for each class.
 (ii)Select training data sets which are representative of the classes of interest that
show both typical average feature values and a typical degree of variability.
 (iii) Check that selected areas have unimodel distributions (histograms).
 (iv)Select training sets (physically) using a computer-based classification system
 (v) The program should allow one to designate half of the points as training points,
and the other half to test the accuracy of the trained classifier.
 (vi) Separability of classes

57
 6.8 IMAGE CLASSIFICATION
6.8.3 UNSUPERVISED CLASSIFICATION

 Unsupervised classification algorithms do not compare .points to be classified with

training data. Rather, unsupervised algorithms examine a large number of unknown
data vectors and divide them into classes based on properties inherent to the data
themselves.
 In particular, use is made of the notion that data vectors within a class should be in
some sense mutually close together in the measurement space, whereas data vectors
in different classes should be comparatively well separated.
 If the components of the data vectors represent the responses in different spectral
bands, the resulting classes might be referred to as spectral classes, as opposed to
information classes, which represent the ground cover types of interest to the analyst.

58
 6.9 PERFORMANCE ANALYSIS OF IRS-BANDS FOR LAND USE/LAND COVER
CLASSIFICATION SYSTEM USING MAXIMUM LIKELIHOOD CLASSIFIER

 Remote sensing has become a powerful tool for the regional mapping of natural
resources and geological features.
 Starting with the use of image during the early stages of development of remote
sensing in the mid-seventies, enough progress has been achieved in data
interpretation with the easy availability of digital data.
 Digital processing of remote sensing data have gained momentum in the last decade.
 In India, with the establishment of remote sensing centres all over the country in
recent years, attention has been focused on the large scale data processing for natural
resources evaluation.
 One important aspect in remote sensing is the characterization and classification of
spectral measurements taken from satellites into various features of the land surface.
 Pattern recognition can be carried out if appropriate procedures are adopted for
classification.

59
 6.9 PERFORMANCE ANALYSIS OF IRS-BANDS FOR LAND USE/LAND COVER
CLASSIFICATION SYSTEM USING MAXIMUM LIKELIHOOD CLASSIFIER

 In classification studies, it is often desirable to know how well the classes can be
separated by observing the values of some feature vector for a set of samples.
 In other words, one wants to know how much information the features provide for
distinguishing the classes.
 To answer these questions a measure is needed to quantify the amount of information
on the features.
 The objective of this study is to improve the classification accuracy of obtaining
accurate and cost effective information about the features. This study utilises a
multiband data set to determine the effectiveness in improving the classification.
 The study assesses the utility of the multi-band data for the study of the urban
environments, the land covers which are often difficult to examine accurately with
remotely sensed data.
 It also attempts to examine the classification accuracy of a number of land cover
classes for different band combinations and the potential of the classification method.
 The emphasis is on the use of Maximum Likelihood Calcification and the derivation
of meaningful confidence level to all land cover classes.

60
 6.9.1 CLASSIFICATION METHODOLOGY

 During the testing of a maximum likelihood classifier for land use classification in
the Hyderabad region, the need arises for an account of the application of statistical
confidence level assessments in remote sensing.
 Such an account is necessary in evaluation of the classification which consists of the
following components:

 (i) The acquisition of data.

 (ii) A decision to the level of class separability desired and attainable.
 (iii) The selection of training areas which will suit a given computer-based
classification software package.
 (iv) The effective operation of the package.
 (v) The selection of an appropriate threshold for each class to apply
likelihood distribution of that class.
 (vi) The creation of appropriate output production.

61
 6.9.2 THE LAND USE AND LAND COVER CLASSIFICATION
SYSTEM

 From multi-spectral data, one needs to identify and isolate particular objects.
 To proceed in a smooth and systematic manner, the data need to be grouped in a
suitable framework.
 The framework should not only be flexible in nomenclature and definitions but also
be capable of incorporating new information obtained from the same source or
different sources.
 We have used level-1 classification.
 The classification categories at level-1 are identified in the study area given below:

(i)Water (ii) Scrub (iii) Forest

(iv) Vacant (v) Commercial (vi) Residential

62
 6.9.2 THE LAND USE AND LAND COVER CLASSIFICATION
SYSTEM

Training samples

 The validity of any training sample depends on two factors and namely, the size and
the representativeness of the sample.
 Sample size is related to the number of spectral bands whose statistical properties are
to be estimated.
 The user should have a prior knowledge of the data and the study area.
 The number and statistical characters to be extracted from training samples can be
identified by their geographical location using maps and ground trulh data.
 Once they are selected accurately, the results could be estimated after using
classification decision rules.

63
 6.9.2 THE LAND USE AND LAND COVER CLASSIFICATION
SYSTEM

Need of a confidence level

 Any computer-derived classification, that will lead ultimately to a ground-cover

thematic map, is based on ground truth data gathered by the user from the selected
training areas.
 This applies whether unsupervised or supervised classification is employed and
whether parametric or non-parametric techniques are used.
 The accuracy of the thematic map depends on the ability to extrapolate data
successfully from the training areas to the whole mapped area.
 Unless we have some statistical measure of the efficiency of extrapolation process,
the level of confidence in the classification cannot be estimated.
 Once a confidence level is so quantified, then the user of the classification data can
relate it by means of the probability of the correct classification, to actuality over the
whole classified area.
 The classification is thus related to the ground actuality by the confidence level.

64
 6.9.3 DATA ANALYSIS

Need of a confidence level

 General statistical parameters, such as, mean, standard deviation, variance-covariance

matrix and correlation coefficients of different classes like water, scrub, forest,
vacant, industrial, and residential, different band data have been analysed
supplementary to the test of performance.
 The sample data are subjected to classification by the application of the Maximum
Likelihood Classifier for all four bands.
 Variance-covariance matrices and correlation coefficient matrices for the three band
data (1,2,3) reflect the accuracy of the classification in each of the classes

65
6.9.3 DATA ANALYSIS

67
 6.9.4 CLASSIFICATION ACCURACY APPROACH

 From the study of the Maximum Likelihood Classifier in the performance analysis of
IRS-bands for land use/land cover classification system it can be deduced that:

 (i) The statistical background to the derivation of a confidence level based on the
probability disfrlbution function can be taken for a classification exercise.

 (ii) This method not only has the advantage as the adaptive feature selected and the
sequential classification by the usual layered classifier, it also improves the accuracy
of the performance and reduces the amount of computation needed. So, to some
extent, we can obtain reasonably good classification results without using excessive
computer time.

68
 6.9.4 CLASSIFICATION ACCURACY APPROACH

 (iii) This study has concluded that the use of multi-band datasets for the land use /
land cover classification system will provide more accurate results than an
independent set of sensor data. The best classification occurred when the data
contained all the major portions of an electromagnetic spectrum.

 (iv) The method of classification accuracy approach reduces the. Processing time
substantially by providing more classes on the pre-classification stage and it may be
comparable to that of the other classifications to get more accurate results.

 (v) More work is required to validate this and additional tests on other areas having
the high spatial complexity needed In order to understand the limits of remote sensor
data and the performance of variables, should be checked for the level-II
classification system.

69
6.9.4 CLASSIFICATION ACCURACY APPROACH

70
6.9.4 CLASSIFICATION ACCURACY APPROACH

71
6.9.4 CLASSIFICATION ACCURACY APPROACH

72