Nonlinear Unmixing via Deep Autoencoder Networks for Generalized Bilinear Model

Hyperspectral unmixing decomposes the observed mixed spectra into a collection of constituent pure material signatures and associated fractional abundances. Because universal modeling ability neural networks, deep learning (DL) techniques are gaining prominence in solving hyperspectral analysis tasks. The autoencoder (AE) network has been extensively investigated linear blind source unmixing. However, mixing model (LMM) may fail to provide good performance when nonlinear effects nonnegligible complex scenarios. Considering limitations LMM, we propose an unsupervised spectral method, based on architecture. Firstly, is employed as encoder extract low-dimension feature pixel. Then, generalized bilinear (GBM) used design decoder, which part one. coefficient can be adjusted by set learnable parameters, makes method perform well both data. Finally, some regular terms imposed loss function alternating update strategy utilized train network. Experimental results synthetic real datasets verify effectiveness proposed show very competitive compared with several existing algorithms.