Endmember-Guided Spectral–Spatial Feature Diffusion for Hyperspectral Anomaly Detection

Hyperspectral anomaly detection (HAD) aims at identifying the pixels whose spectral features notably deviate from their surroundings. However, when the spectral curves of targets are mixed with those of background in a pixel, existing HAD methods suffer from significant performance degradation due to the subtle spectral differences between targets and background. In this paper, We propose an endmember-guided feature diffusion network to address this problem. Specifically, a diffusion network is firstly proposed to regularize the spectral-spatial features of hyperspectral images (HSIs) extracted by an autoencoder. Afterwards, an unmixing module is designed to obtain the endmember features to represent the prototypes of HSIs. The proposed diffusion model subsequently aggregates the endmember features to update the HSIs features according to the spectral linear mixing model. Finally, the feature learning and diffusion are jointly optimized to ensure that the proposed method can deal with the negative effect of mixed pixels on HAD. Experimental results demonstrate that the proposed method has superior detection performance to several state-of-the-art HAD methods. In particular, when the target abundance is 0.01 in mixed pixels on the synthetic dataset, our method can better identify the anomaly targets.