Comparitive Study of Wavelet based Feature Extraction Techniques for Classification of Hyperspectral Data
N, Prabhu; Arora, M. K.; R, Balasubramanian
IIT Roorkee, Roorkee, INDIA

The hyperspectral images are rich in spectral resolution, used in many applications like target/anomaly detection, vegetation classification, mineral mapping, etc. At the same time the huge volume of spectral details which are spread over few hundreds of bands are highly correlated to neighbouring bands. So it is an essential task to extract the data by reducing the redundancy without significant lose to the data. Here in this research paper, two feature extraction techniques based on wavelets are discussed. The behavior of the extracted data are classified by a per pixel classifier and a sub pixel classifier, both supervised, and the results are analyzed by reconstruction of the extracted features. The experiments are done on, two hyperspectral datasets, one Indiana Pine dataset acquired by hyperspectral sensor AVIRIS and the other Roorkee and its surrounding area acquired by Hyperion. The classification accuracies for the extracted features from Daubechies are better than that of Coiflets based feature extraction, for both per pixel and sub pixel classification.