Image Classification using

multi-spectral and
multi-temporal
satellite data

Project by:
Mr. Vinod Jadhav
Ms. Deepali Gadkari
Ms. Sangita Warade
 Approach
 Introduction
 Aims & objects
 Methodology
 Observations & Results
 Conclusion
 Introduction
 The project is divided in two parts:
– Image classification
– Change detection
 Introduction

Image Classification
 It is replacement of visual analysis of the
image data with quantitative techniques for
automating the identification of features in a
scene.
 It categorizes all pixels in an image into land
cover classes.
 It distinguishes as supervised and unsupervised
classification.
 Introduction

Image Classification
 In supervised the image analyst supervises the
pixel categorization process by specifying
training areas.
 In unsupervised the image data are first
classified by aggregating them into the natural
spectral groupings, or clusters, present in scene.
 Introduction

Change Detection
 Change Detection involves the use of multi-
temporal data sets to
– discriminate areas of land cover change between
dates of imaging.
– Change in identify crop health pattern
 Aims & objects
 Supervised and unsupervised classification
of multi-spectral data for land-use mapping
 To decide which classifier is better for
Landuse / Landcover mapping
– By comparing statistical area
– By pixel-by-pixel comparison
 Change detection by using multi-temporal
data
 Study Area
Area of interest:
–Location: Part of Patna, Gaya and Aurangabad
Districts of Bihar
 Patna Town
Study Area
FCC
IRS-1D LISS-III
February-2002

Son River

Hills
 Methodology

 Image classification by applying

Supervised & Unsupervised methods
 Supervised Classification
– Subsetted LISS-III image
– Image enhancement
– Collection of training sets
– Applications of different classifiers
• Minimum Distance
• Maximum Likelihood
 Methodology

 Quantitative Expressions of category

separation
 Signature Separability
 Scatter Plots
 Area
 PCT values
 Confusion matrix
 Application (Running) of Classification
 Methodology

 Unsupervised Classification
– Defining number of classes
– Applications of K-mean classifier
 Source of Data
A) Image Classification:
Satellite: IRS-1D
Sensor: LISS-III
Spatial Resolution: 23.5 meters
Spectral Resolution:
• Green: 0.52-0.59mm
• Red: 0.62 - 0.86 mm
• Near Infrared: 0.77 - 0.86 mm
• Middle Infrared: 1.55 - 1.70 mm

Date of image: February 2002

 Source of Data
B) Change Detection:
Satellite: IRS-1D
Sensor: WiFS
Spatial Resolution: 188 meters
Spectral Resolution:
• Red: 0.62 - 0.86 mm
• Near Infrared: 0.77 - 0.86 mm

Dates of image:
• January 2000
• February 2000
 Software
 Geomatica 8.2.1
– Focus
– PCI Modeler
– OrthoEngine Core
 Classifications
 Unsupervised Classification
– K-mean Classification (5,10,15,20 Classes)
– Fuzzy
– ISODATA
 Supervised Classification
– Maximum Likelihood Classifier
– Minimum Distance Classifier
Unsupervised –5 Classes Unsupervised –10 Classes
Unsupervised –15 Classes Unsupervised –20 Classes
Unsupervised –Fuzzy Unsupervised –ISODATA
Unsupervised: 12 Classes Unsupervised: 6 Classes(Aggregated)
 Supervised Classification
 Classes (Training sites)
– Agriculture-1
– Agriculture-2
– River
– Sand
– Fallow Land
– Hills
– Settlement
Supervised –Min Distance Supervised –Max Likelihood
 Supervised –Parallepiped with Tie
Supervised –Parallepiped breaker
 Parameters used for Classification
Interpretation
 Scatterplot
 Signature Separability
 Area
 PCT values
 Confusion matrix
 Scatter Plot: Minimum distance classification with default threshold

Agriculture2

Hills Sand

Fallow River
Land
Settlements
Scatter Plot: Minimum distance classification while applying threshold

Fallow Land

Settlements
 Scatter Plot: Minimum distance classification after applying threshold

Agriculture2

Sand
Hills

Fallow
Land River
Settlements
Signature Separability: Minimum distance classification
 Minimum Distance Classification: Pixel distribution

Name Code Pixels Image Thres Bias

Agriculture-1 1 0 0.00 3.00 1.00
Agriculture-2 2 1639264 32.23 3.00 1.00
River 3 44893 0.88 3.00 1.00
Sand 4 81202 1.60 3.00 1.00
Settlements 5 711024 13.98 3.00 1.00
Fallow Land 6 2508629 49.32 3.00 1.00
Hills 7 101088 1.99 3.00 1.00
NULL 0 0 0.00
Total 5086100 100.00
 Minimum Distance Classification: Confusion Matrix

Name Code Pixels 2 3 4 5 6 7

--------------------------------------------------------------------------
Agriculture-2 2 583 99.31 0.00 0.00 0.00 0.69 0.00
River 3 837 0.00 83.03 2.63 13.14 0.12 1.08
Sand 4 999 0.00 0.00 98.80 1.20 0.00 0.00
Settlements 5 1080 0.00 0.00 0.00 92.96 5.46 1.57
Fallow Land 6 676 0.00 0.00 0.00 11.83 68.93 19.23
Hills 7 578 0.00 0.00 0.00 21.28 14.01 64.71

Average accuracy = 84.63

Overall accuracy = 86.37

KAPPA COEFFICIENT = 0.83360 Standard Deviation = 0.00605

Confidence Level :
99 +/- 0.01560
95 +/- 0.01185
90 +/- 0.00995
 Scatter Plot: Maximum Likelihood Classification with default threshold

Agriculture2

Sand
Hills

River
Fallow
Land
Settlements
Scatter Plot: Maximum Likelihood Classification while applying threshold values

Fallow
Land
Settlements
 Scatter Plot: Maximum Likelihood Classification after applying threshold

Agriculture2

Hills Sand

Fallow
Land River
Settlements
Signature Separability: Maximum Likelihood classification
 Maximum Likelihood Classification: Pixel Distribution

Name Code Pixels Image Thres Bias

Agriculture-1 1 3396814 66.79 3.00 1.00
Agriculture-2 2 632902 12.44 3.00 1.00
River 3 35026 0.69 3.00 1.00
Sand 4 40182 0.79 3.00 1.00
Settlements 5 142596 2.80 3.00 1.00
Fallow Land 6 746469 14.68 3.00 1.00
Hills 7 92111 1.81 3.00 1.00
NULL 0 0 0.00
Total 5086100 100.00
 Maximum Likelihood Classification: Confusion Matrix

Name Code Pixels 1 2 3 4 5 6 7

--------------------------------------------------------------------------
Agriculture-2 2 583 7.55 92.45 0.00 0.00 0.00 0.00 0.00
River 3 837 14.22 0.00 85.78 0.00 0.00 0.00 0.00
Sand 4 999 5.91 0.00 0.00 94.09 0.00 0.00 0.00
Settlements 5 1080 5.19 0.00 0.65 0.00 90.09 1.94 2.13
Fallow Land 6 676 2.81 0.00 0.00 0.00 0.74 83.58 12.87
Hills 7 578 4.67 0.00 0.00 0.00 1.90 6.23 87.20

Average accuracy = 88.87

Overall accuracy = 89.19

KAPPA COEFFICIENT = 0.87065 Standard Deviation = 0.00530

Confidence Level :
99 +/- 0.01367
95 +/- 0.01038
90 +/- 0.00871
 Scatter Plot: Maximum Likelihood Classification with default threshold values

Agriculture2
Hills
Sand

Fallow River
Land
Settlements
Scatter Plot: Maximum Likelihood Classification while applying threshold

Fallow
Land
Settlements
 Scatter Plot: Maximum Likelihood Classification after applying threshold

Agriculture2

Hills Sand

Fallow
Land River
Settlements
Signature Separability: Maximum Likelihood classification
Comparison of Maximum Likelihood & Minimum Distance classification
Pixel-by-pixel Method
Comparison of Maximum Likelihood & Minimum Distance classification
Pixel-by-pixel Method
Area wise Comparison of classifications
(In %)
Supervised Unsupervised
Max. Likelihood Min. Distance K-Means
Agriculture-1 66.79 0 49.04
Agriculture-2 12.44 32.23 20.30
River 0.69 0.88 1.43
Sand 0.79 1.6 0.47
Settlements 2.8 13.98 7.73
Fallow Land 14.68 49.32 12.29
Hills 1.81 1.99 8.74
 Observations

 Area under each class is another criteria to

compare two classifiers and two
classification schemes.
 Conclusion
 MXL gives better results than Min.Dist.
 Supervised Classification is giving better
results than Unsupervised classification
 Settlements,Hills and Fallow land are the
classes with close reflectance patterns.
 Band combination,threshold value, bias
and tolerance are of critical importance
for Separability of closely resembling
classes.
 Change Detection
WiFS-Image1 - NIR
WiFS-Image1 - R
 NDVI-1 = (NIR-R) / (NIR+R)

WiFS-Image2 - NIR
WiFS-Image2 - R
 NDVI-2 = (NIR-R) / (NIR+R)

 Image Subtraction = (NDVI-1) - (NDVI-2)

Modeler for NDVI of image-1 & image- 2
NDVI = (Near Infra Red – Red) / (Near Infra Red + Red)
Resultant of image-1: NDVI-1
Resultant of image-2: NDVI-2
Modeler for Change Detection
Resultant of image-1:(NDVI-1) – (NDVI-2)
Classified (NDVI-1) – (NDVI-2)
Statistics (NDVI-1) – (NDVI-2)
(Feb., 2000 – Jan. 2000)
Classification Algorithm: K-Means Unsupervised
Classification Input Channels: 1
Classification Result Channel: 2
Number of Clusters: 5
Cluster Pixels Mean Position Std Dev
( 1) 26280 -0.10982 0.07458
( 2) 125943 0.04669 0.03788
( 3) 261951 0.15975 0.03032
( 4) 254666 0.26006 0.03207
( 5) 156760 0.38541 0.04524
------------
Total 825600
 Observations & Results
 Differences in NDVIs ranges from –
0.109 to 0.385
 96.82% area is classified under positive
values.
 3.18% area is classified under negative
values. i.e. crop was having peak
vegetative growth in the month of Jan. and
reached maturity in the month of Feb.
 Conclusion
 NDVI of multi-date satellite data may be
used to separate different crops / crop
growth stages.
Thank You !