Hyperspectral Unmixing with Spectral Variability Using Endmember Guided Probabilistic Generative Deep Learning

Abstract

Spectral signatures of the pure constituent materials vary across the hyperspectral image (HSI) due to variable illumination, atmospheric conditions, and intrinsic variability. Using a single endmember to represent the target material (or endmember) with high spectral variability will lead to errors in estimating abundance. Therefore, we propose a probabilistic generative network (PGM-Net) architecture to learn the spectral variability from the HSI (hereinafter referred to as endmember-guided-probabilistic-model-network, EGPGM-Net). The PGM-Net is guided by endmember-network (E-Net) using the parameter sharing strategy. Experimental analysis was carried out on benchmark datasets to compare the performance of the proposed method with the state-of-the-art methods. Moreover, we have also demonstrated the application of EGPGM-Net for estimating the abundance of red and black soil over sparsely vegetated areas using airborne-visible-and -infrared -imaging-spectrometer-next-generation (AVIRISNG) sensor. The quantitative analysis reveals that the proposed method consistently achieves a better unmixing performance than other linear-mixing and deep learning based models in terms of spectral-angle-distance (SAD) and abunance-root-mean-square error (aRMSE). The proposed semi-supervised approach accurately delineated the abundances of red soil, black soil, crop residue, built-up areas and bituminous roads.