Scribd
Upload
3 views19 pages

M5L1

The document discusses supervised classification in digital image processing, particularly in remote sensing. It covers the selection of training samples, various statistical feature selection methods, and different classifiers such as the Parallelepiped, Minimum Distance to Means, and Maximum Likelihood classifiers. Additionally, it addresses the evaluation of classification accuracy using confusion matrices and Kappa statistics.

Uploaded by

prithwish.roy
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPT, PDF, TXT or read online on Scribd
3 views19 pages

M5L1

The document discusses supervised classification in digital image processing, particularly in remote sensing. It covers the selection of training samples, various statistical feature selection methods, and different classifiers such as the Parallelepiped, Minimum Distance to Means, and Maximum Likelihood classifiers. Additionally, it addresses the evaluation of classification accuracy using confusion matrices and Kappa statistics.

Uploaded by

prithwish.roy
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPT, PDF, TXT or read online on Scribd
You are on page 1/ 19

Digital Image Processing-Information

Extraction
1

(i) Supervised Classification

Remote Sensing: M5L1 D. Nagesh Kumar, IISc


Objective
2

 Introduction
 Selection of Training Samples
 Statistical Feature selection methods
 Parallelepiped classifier
 Minimum Distance to means classifier
 Maximum Likelihood classifier
 Evaluation of Classification accuracy

Remote Sensing: M5L1 D. Nagesh Kumar, IISc


Introduction
3

 In remotely sensed images, a range of digital


numbers represents a single surface class

 Classification is used to smooth out small,


insignificant variations and simplify an image into a
thematic map of land cover.

 Pixels that have similar spectral characteristics and


which are consequently assumed to belong to the
same class are identified and assigned a unique
colour

Remote Sensing: M5L1 D. Nagesh Kumar, IISc


Multispectral Image Classifcation
4

Differentiation of surface classe


A, B and C in bands 1, 2 and 3.

Remote Sensing: M5L1 D. Nagesh Kumar, IISc


Sequence of steps in
Classification process
5

Remote Sensing: M5L1 D. Nagesh Kumar, IISc


Selection of Training Samples
6

 Training sites within an image are representative


of individual land cover type
 Statistical classifiers depend on a good quality of
training data to generate excellent results
 Training area selection is employed to collect
spectral statistics for each of the land cover type.
 Number of ways to select training data either using
in-situ information or selecting on screen polygonal
training data or seeding of training data on screen.

Remote Sensing: M5L1 D. Nagesh Kumar, IISc


Selection of Training Samples
7

 Most of the image processing softwares employ a tool that


enables selection of region of interest (ROI).
 The ROI tool is used to select pixels corresponding to each of
the land cover types as observed on screen

 The users may also select a specific location within the image
using cursor. The seed program can then be used to evaluate
and expand to neighbouring pixels like an amoebae till it
stops finding pixels with similar characteristics as the one
originally selected.

Remote Sensing: M5L1 D. Nagesh Kumar, IISc


Statistical Feature Selection
Methods
8

 Divergence
Widely used statistical measure of separability of
remotely sensed data. Using training data of a
supervised classification, the degree of divergence or
separability between two classes (say a and b) can
be computed using the expression:
1
 1

Diverab  Tr (Va  Vb )(Vb 1  Va 1  Tr (Va 1  Vb 1 )( M a  M b )( M a  M b ) T
2 2

Tr = Trace of matrix
Va ,Vb = Covariance matrices of two classes a and b
Ma,Mb = Mean vectors of the two classes a and b
Remote Sensing: M5L1 D. Nagesh Kumar, IISc
Distance measures
9

 Bhattacharyya Distance measure


This measure separates two classes at a time
assuming that both classes are Gaussian in nature
and that their means and covariance matrices are
available
Va  Vb
det
1 ' (Va  Vb ) 1 2
Bhattab 
 Ma  Mb  ( M a  M b )  log e
8 2 2 det(Va ) det(Vb )

Here det refers to the determinant of matrix


considered.

Remote Sensing: M5L1 D. Nagesh Kumar, IISc


Jeffreys Mattusita Distance
10

 This measure provides a saturating behaviour with


increasing class separation similar to transformed
divergence. The expression is given as:

JM ab  2(1  e  Bhatt ab )

 But this measure is not as computationally effective


as transformed divergence measure.

Remote Sensing: M5L1 D. Nagesh Kumar, IISc


Parallelepiped Classifier
11

 Size of the box is decided based on standard deviation


 Pixels outside the box are not assigned to any class
 Pixels falling in more than one box may be left unclassified

Remote Sensing: M5L1 D. Nagesh Kumar, IISc


Minimum Distance to means
Classifier
12

 Every pixel is assigned to a class based on its


distance from the mean of each class

Remote Sensing: M5L1 D. Nagesh Kumar, IISc


Maximum Likelihood
Classifier
13

 Assumes a normal distribution for the training areas


 Probability contours are created around each training
area and a pixel assigned to a class depending upon
the value of the probability contours that encompass it
 The maximum likelihood classifier is generally
considered to be the most powerful but is also
considered the most computer intensive
 Using this algorithm, pixel1 belongs to class A, pixel2
to class B and pixel4 to class C. Pixel3 has a higher
probability of belonging to class B than class C.

Remote Sensing: M5L1 D. Nagesh Kumar, IISc


Maximum Likelihood
Classifier
14

Remote Sensing: M5L1 D. Nagesh Kumar, IISc


Supervised Classification
15

FCC
R-TM4
G-TM5
B-TM3

Training areas
for supervised
Classificaton

(background
image
blacked out)
Remote Sensing: M5L1 D. Nagesh Kumar, IISc
Supervised Classification
16

Supervised classification
representation: Look up Table
Density sliced classified image

Remote Sensing: M5L1 D. Nagesh Kumar, IISc


Evaluation of Classification
Accuracy
17

 Confusion matrix / Error matrix


Traditional approach is to compare classification results
with corresponding reference or ground data and tabulate
the findings in the form of a square matrix known as
confusion matrix/classification error matrix or contingency
table. This matrix shows how well the classification has
categorized pixels corresponding to each land cover type

Producer’s accuracy = Number of correctly classified


pixels in each category/ number of training set pixels used
for that category

Remote Sensing: M5L1 D. Nagesh Kumar, IISc


Evaluation of Classification
Accuracy
18

User’s accuracy = Number of correctly classified


pixels in each category/Total number of pixels
classified in that category

Kappa statistics
r r
N  xii  x i * x i
i 1 i 1
K hat  r
N  2
x
i 1
i * x i

Where r is the number of rows in the errorxmatrix, is the number of


ii
observations in row i and column i, are the marginal totals for row i and
xi  , x 1
column i . And N represents the total number of observations.

Remote Sensing: M5L1 D. Nagesh Kumar, IISc


Thank You

Remote Sensing: M5L1 19 D. Nagesh Kumar, IISc

You might also like