Explicit Background Endmember Learning for Hyperspectral Anomaly Detection

Hyperspectral anomaly detection (HAD) is challenging especially when anomalies appear in the form of subpixels. Since the spectral signatures of anomalies in mixed pixels are mixed with those of other materials in the background, the anomalies are more difficult to be separated from the background. However, most existing deep learning-based methods neglect the negative effect of mixed pixels on HAD. Therefore, their applications are restricted. To handle this problem, an explicit background endmember learning (EBEL) detection framework is proposed for HAD. Specifically, a spectral-spatial reconstruction error-based background estimation (BE) strategy is employed to obtain the pseudobackground of hyperspectral images (HSIs). Then, the inherent relationships of cross attention mechanism in the Transformer with endmember learning and spectral linear mixing model (LMM) are constructed. Therefore, the cross-attention mechanism is introduced to learn the endmember features from the pseudobackground. Meanwhile, the HSI features can be linearly represented by the background endmember features through a cascaded cross-attention-based module. In this way, the anomalies are detected by the reconstruction errors since the background can be well represented by the background endmember features, while the anomalies cannot. Experimental results on one synthetic dataset and four real datasets demonstrate the superior detection performance of our method to existing state-of-the-art HAD methods.