A Regional Adaptive Autoencoder Network for Hyperspectral Unmixing

Hyperspectral unmixing is the foundation and key component of hyperspectral image refinement applications. In recent years, since they were proposed, autoencoder networks have attracted much attention in the field of hyperspectral unmixing because of their powerful and general abstract representation capabilities. The encoder designs of autoencoders are often single nonlinear or linear networks; however, the hyperspectral images encountered in practical application scenarios include both linear and nonlinear mixing types. A single coding method cannot accurately unmix all mixing types, so we propose a regionally adaptive autoencoder-based unmixing network, which generates linear and nonlinear mixing-type masks by unmixing and mixing the reconstruction process to fit the original image. The mixed-type masks are fused with the autoencoder network and used to guide the autoencoder in performing region-adaptive matching to adaptively adjust the encoding category according to the mixed-type mask in the actual area and ultimately achieve accurate hyperspectral image unmixing results. Moreover, the proposed method has some physical interpretability because of the integration of mixed-type masking into the network. The proposed method is experimentally verified on a typical dataset; its unmixing accuracy is better than that of the current state-of-the-art algorithms, and the effectiveness of its structure is proven via ablation experiments.