Segmented autoencoders for unsupervised embedded hyperspectral band selection

One of the major challenges in hyperspectral imaging (HSI) is the selection of the most informative wavelengths within the vast amount of data in a hypercube. Band selection can reduce the amount of data and computational cost as well as counteracting the negative effects of redundant and erroneous information. In this paper, we propose an unsupervised, embedded band selection algorithm that utilises the deep learning framework. Autoencoders are used to reconstruct measured spectral signatures. By putting a sparsity constraint on the input weights, the bands that contribute most to the reconstruction can be identified and chosen as the selected bands. Additionally, segmenting the input data into several spectral regions and distributing the number of desired bands according to a density measure among these segments, the quality of the selected bands can be increased and the computational time reduced by training several autoencoders. Results on a benchmark remote sensing HSI dataset show that the proposed algorithm improves classification accuracy compared to other state of the art band selection algorithms and thereby builds the basis for a framework of embedded band selection in HSI.