Unsupervised Nonlinear Hyperspectral Unmixing Based on an Extended Multilinear Mixing Model-Inspired Dual-Stream Network

To address the challenges posed by mixed pixels in hyper-spectral images (HSIs), significant advancements have been made in both model- and deep learning-based unmixing methods. However, issues such as complex computational processes and limited interpretability, hinder the improvement of unmixing performance. In this paper, we propose an extended multilinear mixing (EMLM) model-inspired dual-stream network for unsupervised nonlinear hyperspectral unmixing. The alternating direction method of multipliers (ADMM) for the EMLM-based unmixing is unfolded to construct an encoder, connected to an EMLM-based decoder, creating an autoencoder-like architecture. Within a weight-sharing dual-stream framework, two same network branches process original HSIs and superpixel-averaging-based coarse HSIs, respectively. Estimates from both branches contribute to local spatial similarity regularizers, enhancing the network’s loss function. Experimental results on two hyperspectral datasets demonstrate the superior performance of our proposed method over other state-of-the-art methods.