B. Tech.

V Semester (Civil Engineering)

2020-21 (Odd Semester)

CE: 15105 – GEOINFORMATICS

Unit-3: Satellite Image Processing [Part-III]

By Prof. RD Gupta, MNNIT Allahabad (India) 1

Image Classification

By Prof. RD Gupta, MNNIT Allahabad (India) 2

 Scatter Plot

Scatter plot / scatter diagram is a graphical plot between the BVs of one band to
that of another band.

• Let’s take pixel observations from

Band 4 (Near-IR) Digital Number

two bands digital image data set.

• The 2-dimensional BV attributed to

each pixel is known as
measurement vector.

• Say for a pixel, band 3 BV is 70

and band 4 BV is 15; then, the
measurement vector for this pixel
70
is (70, 15) in the feature space. + (70, 15)
15

Band 3 (Red) Digital Number

By Prof. RD Gupta, MNNIT Allahabad (India) 3

3-D Scatter Plot / Cluster Diagram

By Prof. RD Gupta, MNNIT Allahabad (India) 4

 Image Classification

IMAGE CLASSIFICATION

Unsupervised Supervised

By Prof. RD Gupta, MNNIT Allahabad (India) 5

 Image Classification

1. Unsupervised Classification:

We use Clustering techniques

 K-means clustering
 ISODATA (Iterative Self Organizing Data Analysis) Technique

Supervised Classification:

We use classification algorithms (i.e. Classifiers)

 Minimum Distance to Mean Classifier
 Parallelepiped Classifier
 Maximum Likelihood Classifier (MLC)
 Neuro Classifier (ANN based classifier)
 Fuzzy Classifier

By Prof. RD Gupta, MNNIT Allahabad (India) 6

Unsupervised Classification

By Prof. RD Gupta, MNNIT Allahabad (India) 7

 Image Classification: Unsupervised Classification

Unsupervised classifiers do not use training data as the basis for classification.

The classes that results from unsupervised classification are spectral classes or
spectrally separable classes.

The pixels in an image are aggregated into a number of classes based on the
natural groupings or clusters present in the image.

There are several clustering algorithms that can be used to determine the natural
spectral groupings- e.g.- K-means algorithm, ISODATA algorithm

Since these are based solely on natural groupings in the image, the identity of
these spectral classes will not be initially known.

These spectral classes are compared with some form of reference data (large
scale photograph, ground truth collection, maps)

Determine the identity and informational value of each class.

By Prof. RD Gupta, MNNIT Allahabad (India) 8

Any Questions …….

By Prof. RD Gupta, MNNIT Allahabad (India) 9

Supervised Classification

By Prof. RD Gupta, MNNIT Allahabad (India) 10

 Classification Algorithms/ Classifiers

• Minimum Distance to Mean Classifier

• Parallelepiped Classifier
• Maximum Likelihood Classifier (MLC)

By Prof. RD Gupta, MNNIT Allahabad (India) 11

 Minimum Distance To Mean Classifier

• Compute mean spectral

value in each band for each
category.
• These values are known
as mean vector for
each category.

Band 3 (Red) Digital Number

• Compute the distance
between the unknown pixel
and each of the category
mean.
• The unknown pixel is
classified into that class
from which it is closest,
i.e., at minimum distance.
• Computationally efficient
and mathematically simple,
however, Insensitive to the
variance in the spectral
response data.
Band 4 (Near-IR) Digital Number

By Prof. RD Gupta, MNNIT Allahabad (India) 12

 Gaussian Maximum Likelihood Classifier (MLC)

The MLC quantitatively evaluates variance and covariance of each

category/ class from the training data
It is assumed that the distribution of the cloud of points forming the
category training data is Gaussian

Under this condition, the distribution of a category can be completely

described by the mean vector and the covariance matrix

The resulting bell-shaped surfaces are called probability density

functions and there is one such function for each spectral category

By Prof. RD Gupta, MNNIT Allahabad (India) 13

 MLC ….

The probability density

functions are used to
classify an
unidentified pixel by
computing the
probability of the
pixel value belonging
to each category
After evaluating the
probability in each
category, the pixel
would be assigned to
the most likely class
(highest probability
value)
 Stages in Supervised Classification

Three Stages in Supervised Classification:

Training Stage, Classification Stage & Output Stage
By Prof. RD Gupta, MNNIT Allahabad (India) 15
 Supervised Classification

1. Select an appropriate classification scheme (e.g., Anderson’s)

2. Select the land use/ land cover classes (i.e., features of interest)
carefully depending upon the area/ terrain.
3. Select representative training sites for each class.
4. Develop a numerical description of spectral patterns/ spectral attributes
for each feature in the image.
5. Extract statistics from the training site spectral data
(Check for separability between various classes)
6. Select appropriate bands (channels) for classification

7. Select the appropriate classification algorithm

(Minimum Distance to Mean Classifier, Parallelepiped Classifier, Maximum
Likelihood Classifier-MLC)

8. Classify the satellite image into “m” classes.

9. Statistically evaluate the classification accuracy.
(Confusion Matrix, Kappa Coefficient)

By Prof. RD Gupta, MNNIT Allahabad (India) 16

 Rectified Image

IRS-1C LISS-III FCC IRS-1C PAN FCC

By Prof. RD Gupta, MNNIT Allahabad (India) 17

 Classified Image

Land Use and Land Cover Map

By Prof. RD Gupta, MNNIT Allahabad (India) 18
Classification Accuracy

By Prof. RD Gupta, MNNIT Allahabad (India) 19

 Classification Accuracy

• Classification accuracy evaluates the performance of a classifier

• For accuracy assessment, a set of test data is used for which the
correct class assignments is known by ground observations/ air
photograph interpretation/ map study

• Accuracy is assessed by generating a “k x k” matrix known as

Confusion Matrix/ Contingency Matrix/ Error Matrix
 Confusion Matrix

Known Training Data Sets (Known Row • Element i of row i (i.e., ith diagonal
category Cover Types) Total element) contains the number of pixels
type (∑ xi+) which are correctly classified.
F U S W
F 50 3 0 0 53 • The other elements of row i give the
number and distribution of pixels that
U 4 62 3 0 69 have been incorrectly classified.
S 3 5 70 2 80
• The overall classification accuracy is
W 0 0 0 64 64 computed by diving the total number of
correctly classified for all classes by the
Column 57 70 73 66 266
total number of reference pixels.
Total (∑ x+i)

• Classification error of Omission (exclusion) : along non-diagonal column elements

• Classification error of Commission (inclusion): along non-diagonal row elements

• Overall Classification Accuracy = (50 + 62 + 70 + 64)/ 266

= 246/ 266
= 92.5 %
 Confusion Matrix

• User’s Accuracy (indicates the

Known Training Data Sets (Known Row
probability that a pixel classified into
category Cover Types) Total
a given category actually represents
type (∑ xi+)
F U S W that category on the ground)
F 50 3 0 0 53 Class F = 50/ 53 = 94.3%
U 4 62 3 0 69
Class U = 62/ 69 = 89.9%
S 3 5 70 2 80
Class S = 70/ 80 = 87.5%
W 0 0 0 64 64
Class W = 64/ 64 = 100%
Column 57 70 73 66 266
Total (∑ x+i)
• Commission (inclusion) Error:
along non-diagonal row elements

• Producer’s Accuracy (indicates how well

training set pixels of the given cover
type are classified)

Class F Class U Class S Class W

= 50/ 57 = 62/ 70 = 70/ 73 = 64/ 66

= 87.7% = 88.6% = 95.9% = 96.9%

• Omission (exclusion) Error :

along non-diagonal column elements
 Kappa Coefficient

• It a statistical measure of accuracy assessment

 It is derived from the contingency matrix

xii = Diagonal entries of the confusion matrix

r r
N ∑ xii −∑ xi + x + i xi + = Sum of the row i

κ ==i 1 =r
i 1
x+i = Sum of the column i of the confusion matrix

N −∑ xi + x + i
2
r = Number of categories/ classes
i=1
N = Total Number of elements of confusion matrix

For the confusion matrix shown above, the kappa coefficient is:

= (266 * 246 – 17915)/ {(266)2 - 17915} = 0.89

• A value of 0.0 indicates no agreement while a value of 1.0 shows perfect agreement between
the classifier output and the reference data

• A value of kappa of 0.75 or grater shows a “very good to excellent” classifier performance
while a value of less than 0.4 is “poor”
Applications

By Prof. RD Gupta, MNNIT Allahabad (India) 24

 Remote Sensing Applications

This can be used for a wide range of applications, including:

 Urban Planning/ Developmental Planning

 Land Use & Land Cover Mapping

 Forestry, Agriculture

 Ocean Resources

 Environmental Impact Analysis (EIA of River Valley Projects, Mining)

 Water Resources Monitoring & Management

 Mineral Prospecting, Desertification, Soil mapping

 Natural Hazards Mapping & Monitoring (Floods, Landslides, Earthquakes, Volcano, etc.)

 Natural Resource Inventorying, Monitoring & Management

 Remote Sensing Applications

• Remote sensing for Natural Resources Management

• Remote Sensing and Environmental Impact Assessment (EIA)
 EIA of Major River Valley Projects
 EIA of Mining

• Remote Sensing and Desertification

• Remote Sensing Applications for Soil Mapping

• Monitoring of Land Resources

 Watershed Management [water (surface & ground), Soil erosion]

• Remote Sensing for Natural Hazards Monitoring

 Floods
 Landslides
 Earthquakes
 Volcanoes
 Land Subsidence

By Prof. RD Gupta, MNNIT Allahabad (India) 26

 Remote Sensing
for
Natural Resources Mapping, Monitoring & Management

Natural resources are dynamic in nature

and change over a period of time

For planning, management and monitoring of natural resources

Timely and reliable information

about land use/land cover for a region is needed

Further, for formulating any land based information system

Accurate and updated information about natural resources

with reference to their nature and distribution is essential

By Prof. RD Gupta, MNNIT Allahabad (India) 27

 Mapping & Monitoring of Natural Resources using Remote Sensing

Four distinct advantages as compared to conventional methods:

Synoptic view: Satellite images are "big-picture" views of large areas of the
surface of earth. The positions, distribution, and spatial relationships of
features are clearly evident.
Repetitive coverage: Repeated images of the same regions, taken at
regular intervals over a period of time provide databases for
recognizing and studying particular changes.
Multispectral data: Satellite sensors are designed to operate in many
different portions of the electromagnetic spectrum. For this, visible,
infrared and microwave energy is used to extract information about
earth's surface and/or atmosphere.
Low-cost data: Near-global, repetitive collection of data is far cheaper
using satellite sensors than collecting the same type and quantity of
data through conventional ground surveys.

By Prof. RD Gupta, MNNIT Allahabad (India) 28

 Flow Chart for Resource Mapping

Statement of the problem

Data Acquisition (Remote Sensing/ Collateral)

Raw Satellite Data (PAN/LISS-III Image of IRS-1C/1D)

Satellite Image Rectification

Generation of Spectral Signatures for each class

Edit / Evaluate Signatures of each class

Classify Image (Supervised/ Unsupervised Classification)

Evaluate Classification (Classification Accuracy Assessment)

Land related Plans & Projects

By Prof. RD Gupta, MNNIT Allahabad (India) 29

 Any Questions ?

Any Questions ?

--- from Satellite Image Processing Portion ---

By Prof. RD Gupta, MNNIT Allahabad (India) 30