Investigations of band selection techniques and endmember extraction algorithms for hyperspectral unmixing

Abstract

Hyperspectral image processing is an emerging concept in the field of remote sensing Hyperspectral sensors collect hundreds or thousands of bands at different wavelengths resulting in a multidimensional imagecube and hence the spectral resolution of hyperspectral images is very high compared to multispectral images of 4 10 bands. Because of large volumes of data collected by imaging spectrometer, hyperspectral image processing has received the considerable interest in the recent years These data are collected either by a satellite or an airborne instrument and sent to the ground station for subsequent processing Because of huge dimension it provides abundant information than multispectral images The thesis focuses on development of a novel algorithm by combining band selection with endmember extraction for spectral unmixing in hyperspectral images The major issue in hyperspectral images is its high dimensionality nature Hence dimensionality reduction is an essential task in all hyperspectral image processing. Dimensionality reduction can be achieved either by feature extraction or feature selection Feature extraction method transforms high dimensional hyperspectral imagecube into a lower dimensional space where the originality of the data is perturbed On the other hand feature selection also called as band selection selects the most informative bands without disturbing the originality of data Prior to band selection there is an issue of determination of number of bands to be selected The issue can be resolved by two ways namely intrinsic dimension and virtual dimensionality Intrinsic dimension is well suited for multispectral images whereas virtual dimensionality is appropriate for hyperspectral images. The eigenthresholding based method which is derived from the concept of Neyman Pearson detection is used to provide reliable estimation of virtual dimensionality newline