Nonlocal and Local Feature-Coupled Self-Supervised Network for Hyperspectral Anomaly Detection

Abstract

Hyperspectral anomaly detection (HAD) aims to locate targets deviating from the background distribution in hyperspectral images (HSIs) without requiring prior knowledge. Most current deep learning-based HAD methods struggle to effectively distinguish anomalies due to limited utilization of supervision information and intrinsic nonlocal self-similarity in HSIs. To this end, this article proposes a novel nonlocal and local feature-coupled self-supervised network (NL2Net) tailored for HAD. NL2Net employs a dual-branch architecture that integrates both local and nonlocal feature extraction. The local feature extraction branch (LFEB) leverages centrally masked and dilated convolutions to extract local spatial-spectral features, while the non-LFEB incorporates a simplified self-attention module to capture long-range dependencies. Furthermore, an improved center block masked convolution strengthens NL2Net ’s focus on surrounding background features, enhancing the background modeling effect. By reconstructing pure backgrounds and suppressing anomalous features, NL2Net achieves precise anomaly separation and superior HAD performance. Experimental results demonstrate its ability to effectively integrate multidimensional features and enhance HAD accuracy, surpassing state-of-the-art methods.