REDU-Net: Robust and Efficient Dynamic Unfolding Network for Abundance Estimation

In the field of hyperspectral unmixing (HU), accurately determining the spectral signatures of pure materials (endmembers) and their abundances presents significant challenges. Traditional methods often necessitate extensive parameter tuning, while some deep learning methods have faced critique for limited interpretability and model generalization. In this article, a new, robust, efficient, and interpretable dynamic unfolding network (REDU-Net) is designed to improve the accuracy of abundance estimation in HU. REDU-Net applies an improved version of the fast iterative shrinkage thresholding algorithm (FISTA), which combines the fast convergence of traditional optimization algorithms with the learning capabilities of deep neural networks. This combination enhances performance and opens new avenues for robust and efficient solutions. The architecture of REDU-Net unfolds each iteration of the optimization algorithm into separate phases in the network, where each phase uses dynamic Levenberg-Marquardt (LM), threshold denoising, and momentum modules to sequentially process and refine the output. In particular, the threshold denoising module comprises positive and inverse transformations to improve stability and efficiency. REDU-Net's critical parameters are obtained by training on the hyperspectral data rather than manually pre-setting them. Extensive experiments are conducted using both simulated and real datasets. The experimental results verify that the proposed REDU-Net exhibits effectiveness and robustness compared to state-of-the-art methods in the HU task.