基于 ADMM 和深度生成先验的高光谱解混方法

The presence of mixed pixels restricts the accuracy of hyperspectral image classification and object detection. To improve the accuracy of mixed pixel decomposition and accurately analyze the composition of mixed pixels, this study proposes a hyperspectral unmixing method that combines an optimization method with deep generative priors, thereby achieving an organic combination of data-driven and model-driven approaches. In recent years, deep neural networks have been widely used in hyperspectral unmixing; however, these methods often act as “black boxes” lacking physical interpretability. Conversely, traditional mathematically optimized hyperspectral unmixing methods use manually selected priors to introduce intrinsic information and improve the accuracy of results. However, computing a complex regularizer needs difficult algorithms, and some information cannot be modeled mathematically. In this study, we propose a hyperspectral unmixing method that integrates the alternating direction method of multipliers (ADMMs) with deep generative priors to combine the strengths of both approaches. Specifically, we use ADMM to decompose the data-fitting term and generative priors, and the decoder of a VAE pre-trained by abundance calculated using conventional methods is applied as the generator. This study uses simulated and real remote-sensing datasets to evaluate the effectiveness of the proposed method.