Supervised data-driven approach for hyperspectral band selection using quantization

Abstract

Hyperspectral image (HSI), acquired in narrow and contiguous bands, contains redundant information in neighbouring bands. In order to overcome the processing overhead of this data like Hughes phenomenon, dimension of HSI is reduced either by using feature extraction or selection technique. Most of the existing dimensionality reduction methods depend on the user input to provide the reduced set of bands. However, perceiving the required number of bands in advance is difficult, since this number varies with dataset. In order to cope up with this limitation some wrapper-based methods are suggested which are again computationally inefficient. Hence, the present research focuses on a supervised data-driven approach that enables in selecting required number of bands without any user intervention. In proposed method, signature patterns with minimum and maximum reflectance values are extracted for each class from the labeled data, which are subsequently quantized. The quantization process continues till unique patterns are obtained for each class. Finally, bands having maximum correlation and minimum variance are eliminated to ensure minimum redundancy among selected bands. The experimental result shows that, an improved classification accuracy is obtained while using the proposed method on two real HSIs as compared to other state-of-the-art methods.