Data Driven Joint Hyperspectral Band Selection and Image Classification

Abstract

Hyperspectral sensors acquire data with a large number of spectral bands. These large number of bands make the processing computationally expensive and difficult in many real-world applications. In addition, with the spatial dimensions, the volume of the data creates problems for cases where the applications permit only limited resources both in terms of hardware computational and storage requirements. To avoid these limitations, band selection plays very pivotal role for many applications. Existing techniques utilize redundancy, clus-tering, sparsity, ranking type criteria for band selection. We propose an end-to-end deep learning pipeline together with a constrained measurement learning structure to select bands in a data driven manner to optimize directly the final task, which is the classification accuracy for this paper. Our results on a publicly available hyperspectral dataset show that the proposed data-driven approach provides higher classification accuracy compared to the existing state-of-art methods for the same number of bands utilized.