An Improved Hyperspectral Unmixing Approach Based on a Spatial–Spectral Adaptive Nonlinear Unmixing Network

The autoencoder (AE) framework is usually adopted as a baseline network for hyperspectral unmixing. Totally an AE performs well in hyperspectral unmixing through automatically learning low-dimensional embedding and reconstructing data. However, most available AE-based hyperspectral unmixing networks do not fully consider the spatial and spectral information of different ground features in hyperspectral images and output relatively fixed ratios of linear and nonlinear photon scattering effects under different scenarios. Therefore, these methods have poor generalization abilities across different ground features and scenarios. Here, inspired by the two-stream network structure, we propose a spatial–spectral adaptive nonlinear unmixing network (SSANU-Net) in which the spatial–spectral information of hyperspectral imagery is effectively learned using the two-stream encoder, followed by the simulation of the linear–nonlinear scattering component of photons using a two-stream decoder. Additionally, we adopt a combination of spatial–spectral and linear–nonlinear components using the optimized adaptive weighting strategy of learnable parameters. Experiments with several hyperspectral image datasets (i.e., Samson, Jasper Ridge, and Urban) showed that the proposed SSANU-Net network had higher unmixing accuracy and generalization performance compared with several conventional methods. This demonstrates that SSANU-Net represents a novel method for hyperspectral unmixing analysis.