Performance evaluation of various algorithms for hyperspectral image analysis
Student name: Ms Adrija Chatterjee
Guide: Dr P K Joshi
Year of completion: 2011
Host Organisation: Regional Remote Sensing Centre-North, NRSC, Indian Space Research Organisation (ISRO), Dept. of Spac
Supervisor (Host Organisation): Mr Kamal Pandey
Abstract: Hyperspectral data carries detailed spectral information of the ground surface that makes it possible
to make a descriptive study of a particular feature. Due to abundance of information contained in a
hyperspectral images, target detection is one of the most important usages of Hyperspectral Imaging.
The detection of a target in any hyperspectral image can be accomplished through several different
approaches. Hyperion image of Dehradun and its surroundings has been used for this study and four
targets, namely, dry grass, pine, sal (a species of tree dominant in the study area) and bare soil were
selected. The major steps involved for detection of any target in hyperspectral image are atmospheric
correction, dimensionality reduction, background characterization, and finally detection. This study
primarily aims at identifying the optimal algorithmic chain for the detection of the selected targets.
Several different algorithms for each step, like FLAASH, ATCOR, PCA, MNF, ICA, PPI, SAM and
OSP, were studied. Receiver operative characteristic (ROC) curves were generated for each chain
generated by these algorithms for each of the four targets in the different stages. Based on these ROC
curves, the performance of each chain of algorithm has been evaluated in order to find the best chain
of algorithm for the detection of the selected targets in the study area. Overall, there was no
particular algorithm for atmospheric correction which would work better for all the targets. For sal
and pine FLAASH performed better and for the rest ATCOR performed better. Similarly, in case of
analysis of the complete chain of algorithm, it is seen that no particular chain gives the best results
for all the four selected targets. Nevertheless, ATCOR-PCA-PPI-SAM is a chain of algorithm which
might yield satisfactory outputs in the detection of all the four targets i.e. sal, pine, dry grass and bare
soil, in the study area.
Keywords: Hyperion, atmospheric correction, algorithm chain, ROC, target detection.
