Sample Similarity-Guided Spatial-Spectral Masked Autoencoder for Hyperspectral Anomaly Detection

Hyperspectral anomaly detection aims to identify targets that are significantly different from the surrounding background within hyperspectral image (HSI). The lack of prior information poses a significant challenge in effective designing abnormal detectors. Most existing algorithms focus on reconstructing the background and detecting anomalies through reconstruction errors. However, due to the complexity of the background and the interference caused by abnormal pixels, it is difficult to accurately recover the high-quality background, resulting in limited detection accuracy. To solve these problems, a novel anomaly detector based on sample similarity guidance and spatial-spectral masked autoencoder (SSMAE) is proposed in this paper. Firstly, to alleviate the impact of abnormal pixels on background restoration, a spatial-spectral masked (SSM) strategy is adopted to reduce abnormal pixels and minimize their interference on subsequent feature extraction. In addition, we design a multi-scale residual convolution block to extract multi-scale features of the background. This enables the model to capture diverse aspects of the background, enhancing its ability to reconstruct it accurately. Recognizing that the application of SSM may lead to the loss of some background information, we guide the decoder by calculating the local similarity features of the unmasked patches. This guidance significantly improves the accuracy of background reconstruction during the decoding phase. Finally, anomalies are detected on the residual between reconstructed background and input HSI. By comparing SSMAE with eight detectors on two real-world datasets, the experimental results demonstrate its superior detection performance and accuracy.