Back
INDEX
Next

Abstract

1. Introduction

1.1 Land cover and Land use

1.2 Hyperspectral Remote Sensing

1.3 Research Objectives

1.4 Outline of the Report

1.5 Data

   

1.5.1 Instrument Background

  1.6 Study Area
2. Preprocessing of Hyperspectral Image Data
2.1 The Challenges for Hyperspectral Processing
    2.1.1 Calibration
    2.1.2 Atmospheric Correction
    2.1.3 Radiometric Correction
    2.1.4 Data Normalisation
  2.2 Data Characteristics
  2.2.1 Data Volume
 

2.2.2 Redundancy

 

2.3 Feature reduction / Band reduction / Data dimensionality reduction techniques

  2.3.1 Principal Component Analysis (PCA)
 

2.3.2 Minimum Noise Fraction (MNF)

3. Hard Classification Techniques

3.1 Hard Classification Techniques

3.2 Supervised Classification

  3.2.1 Gaussian Maximum Likelihood Classifier (GMLC)
 

3.2.2 Spectral Angle Mapper (SAM)

 

3.2.3 Neural Network

 

3.2.4 Decision Tree Approach

  3.3 Unsupervised Classification
  3.3.1Clustering
  3.3.2 K – Means Algorithm
  3.3.3 Drawbacks of Clustering
  3.4 Supervised versus Unsupervised Classification
  3.5 Validation of the result
    3.5.1 Classification Error Matrix
4. Soft Classification: Spectral Unmixing
4.1 Introduction
4.2 Linear Unmixing
4.3 Constraint Least Squares method (CLSM)
4.4 Endmember Extraction
  4.4.1 Pixel Purity Index (PPI)
 

4.4.2 Scatter Plot

  4.5 Literature review: Spectral Unmixing
 

4.6 Summary

5. Results
 

5.1 Land cover Mapping using LISS-3 MSS

  5.1.1 Georeferencing and Geometric Correction
 

5.1.2 Land Cover Analysis

  5.2 Land Cover Mapping using MODIS Data
    5.2.1 Georeferencing and Geometric Correction
    5.2.2 Land Cover Analysis
  5.3 Classification of high resolution LISS-3 MSS data
 

5.3.1 Unsupervised Classification

  5.3.2 Supervised Classification
  5.4 Classification of MODIS data
    5.4.1 Classification of MODIS Bands 1 to 7
    5.4.2 Classification of Principal Components (PC's) of MODIS Bands 1 to 36
    5.4.3 Classification of Minimum Noise Fraction (MNF) components of MODIS Bands 1 to 36
  5.5 Soft classification: Spectral Unmixing of MODIS imagery
  5.5.1 MNF Transformation
  5.5.2 Endmember Collection
  5.5.3 Class spectral characteristics of the Endmembers
 

5.5.4Linear Spectral Unmixing

6. Accuracy Assessment
  6.1 LISS-3 MSS Classification Accuracy
 

6.2 MODIS Classification Accuracy

  6.2.1 Accuracy Assessment of hard classification for MODIS
  6.2.2 Accuracy Assessment of soft classification for MODIS
  6.3 Discussion
  6.3.1 Hard Classification
  6.3.2 Soft Classification using Linear Spectral Unmixing
  6.3.3 Classification Accuracy
7. Conclusions
8. References
Annexure A
 
L1B Reflective Calibration - The MODIS reflective calibration algorithm is designed to determine the at-aperture spectral radiance of the Earth scene and the bidirectional reflectance of the Earth scene with their respective associated uncertainties. Level 1A data is Earth-located raw sensor digital numbers and Level 1B data is Earth-located, calibrated data in physical units. The Solar Diffuser, Spectroradiometric Calibration Assembly (SRCA) and Space View are used periodically to determine calibration coefficients for the reflective bands. The Space View is used every scan along with the periodic calibration results to calibrate the reflective bands. The on-orbit reflective band calibration is a one-point method adjusted by data from a two-point periodic method to fit a linear detector response [2].
Annexure B
  Similarity Metrics and Clustering Criteria
 

Automatic Cluster Detection

Annexure C
Annexure D
Figures Tables

Figure 1. 1 : Study area – Kolar district, Karnataka State, India

Figure 2. 1 : Atmospheric effects influencing the measurement of reflected solar energy

Figure 2. 2 : Atmospheric correction processing thread flow chart [49]

Figure 2. 3 : Rotated coordinate axes in PCA [19].
Figure 3. 1 : Equiprobability contours defined by a maximum-likelihood classifier

Figure 3.2 : Spectral Angle Mapping concept. (a) For a given feature type, the vector corresponding to its spectrum will lie along a line passing through the origin, with the magnitude of the vector being smaller (A) or larger (B) under lower or higher illumination, respectively. (b) When comparing the vector for an unknown feature type (C) to a known material with laboratory-measured spectral vector (D), the two features match if the angle ‘a ' is smaller than a specified tolerance value. (After Kruse et al., 1993) [34].

  Figure 3.3 : Example of an artificial neural network with one input layer, two hidden layers and one output layer [19].
 

Figure 3. 4 : Example of Decision Tree

  Figure 3.5 : LISS-3 Preprocessing and Classification
 

Figure 3.6 : Processing and Hard Classification of MODIS data.

 

Figure 4.1 : Four cases of mixed pixels [116]

 

Figure 4.2 : Scatter plots between two bands typically show a triangular shape, with the data radiating away from the shade-point and A, B, C as the endmembers

  Figure 4.3 : Overall methodology of linear unmixing process
 

Figure 5.1 : NDVI of LISS-3 MSS based on bands 3 (Red) and 4 (NIR).

 

Figure 5.2 : NDVI generated using MODIS bands 1 and 2

  Figure 5.3 : Unsupervised classification of LISS-3 image
  Figure 5.4 : (A) False Colour Composite of the LISS-3 image, (B) Kolar district with training data set, (C) Supervised Classified image of LISS-3 MSS
  Figure 5.5 : Different MODIS inputs to the hard classification algorithms
 

Figure 5.6 : Unsupervised Classification on MODIS Bands 1 to 7

 

Figure 5.7 : Plot of training RMS vs. iterations of NN applied on MODIS bands 1 to 7

  Figure 5.8 : Supervised Classification using (A) MLC, (B) SAM, (C) NN and (D) Decision Tree Approach on MODIS Bands 1 to 7.
 

Figure 5. 9 : Plot of training RMS vs. iterations of NN applied on PCA bands (MODIS Bands 1 to 36).

  Figure 5.10 : Supervised Classification using (A) MLC, (B) SAM, (C) NN and (D) Decision Tree Approach on PC's
 

Figure 5.11 : Plot of training RMS vs. iterations of NN applied on MNF Components

  Figure 5.12 : Supervised Classification using (A) MLC, (B) SAM, (C) NN and (D) Decision Tree Approach on MNF components
  Figure 5.13 : 3-Dimensional visualisation of the Endmembers showing their separability
  Figure 5.14 : (A) Gray scale Abundance Maps for Agriculture and (B) Built up land (Urban / Rural )
  Figure 5.15 : Gray scale Abundance Maps for Evergreen / Semi-Evergreen Forest (A) Plantations/orchards (B) Wasteland/Barren rock/Stone (C) and Waterbodies (D). Bright pixels represent higher abundance of 50% or more stretched from black to white
  Figure 5.16 : Overall RMSE for MODIS. Bright pixels represent high error
  Figure 6.1 : Chikballapur taluk in Kolar district where field data were collected
  Figure 6.2 : Best (left) to worst (right) classification algorithms for mapping land cover classes in Chikballapur taluk. LISS-3 MSS classified image is used for comparison
Tables

Table 1.1 : Specifications of MODIS sensor [19]

Table 3.1 : Genesis, advantages and the disadvantages of the classification techniques

Table 5.1 : Transformed Divergence matrix of the spectral classes in LISS-3. Values greater than 1.9 indicate a very good separability

Table 5.2 : Class wise percentage statistics for Unsupervised and Supervised classified maps using LISS-3 MSS

Table 5.3 : Transformed Divergence matrix of the spectral classes in MODIS Bands 1 to 7.

Table 5.4 : Knowledge based Land cover classification of MODIS data (Bands 1 to 7).

Table 5.5 : Percentage wise distribution of land cover classes for MODIS bands 1 to 7 classified using K-Means, MLC, SAM, NN and Decision Tree Approach

Table 5.6 : Eigenvalue for the PCA analysis of MODIS 36 bands.

Table 5.7 : Transformed Divergence matrix of the spectral classes in PCs.

Table 5.8 : Knowledge based LC classification of PC's

Table 5.9 : Percentage wise distribution of LC using MLC, SAM, NN and DTA on PC's

Table 5.10 : Eigenvalue for the MNF analysis of MODIS 36 bands.

Table 5.11 : Transformed Divergence matrix of the spectral classes in MNF

Table 5.12 : Knowledge based LC classification for MNF components
Table 5.13 : Percentage wise distribution of LC classes obtained from MNF components classification using MLC, SAM, NN and DTA

Table 5.14 : Land cover classes compared on the basis of different algorithms versus percentage area.

Table 6.1 : Area statistics of each taluk in Kolar district.

Table 6.2 : Producer's accuracy, user's accuracy and overall accuracy of land cover classification using LISS-3 MSS data for Chikballapur Taluk

Table 6.3 : Overall Accuracy of classified MODIS Data of Chikballapur taluk.

Table 6.4 : User's Accuracy of classified MODIS Data of Chikballapur taluk.

Table 6.5 : Producer's Accuracy of classified MODIS Data of Chikballapur taluk.

Table 6. 6 : Land Cover statistics for Chikballapur Taluk

Table 6.7 : Overall Accuracy obtained from pixel to pixel analysis with LISS-3 image comparison for Chikballapur Taluk
Table 6.8 : User's Accuracy obtained from pixel to pixel analysis with LISS-3 image comparison for Chikballapur taluk

Table 6.9 : Producer's Accuracy obtained from pixel to pixel analysis with LISS-3 image comparison for Chikballapur taluk.

Table 6.10 : Land Cover details of fraction images for Chikballapur taluk .

Table 6.11 : Validation of land cover classes in Chikballapur

Acknowledgements
PDF