A Global-to-Local Spectral-Spatial Attention-Based Nonlinearity and Scaled Endmember Variability Parametric Learning Network for Unmixing

Hyperspectral unmixing has attracted increasing attention in remote sensing applications. Unfortunately, significant unmixing residuals often arise from the coupled nonlinear mixing effects and spectral variability (SV), bringing challenges for reliably solving the underlying optimization problems in practical applications. Although deep autoencoder architectures have shown advantages in learning latent unmixing features from hyperspectral data, their performance in capturing accurate spectral-spatial information and interpreting both nonlinearity and SV remains limited. This paper proposes a novel autoencoder-based unmixing method that introduces a spectral-spatial attention mechanism to learn refined global-to-local semantic features of land covers. Leveraging the features, the model enables efficient parametric learning of nonlinearity and SV through the integration of physically interpretable second-order scatterings and scaled SV factors. Experimental results indicate that the proposed method has superior unmixing performance compared to the state-of-the-art methods.