Discriminative Spectral Analysis via Dual-Window Coupled Modeling for Hyperspectral Anomaly Target Detection

In hyperspectral anomaly detection for noncooperative regions with complex noise, obtaining prior knowledge in advance is impractical. Furthermore, subpixel spectral mixtures caused by low spatial resolution hyperspectral imagery and intrinsic environmental noise significantly degrade the detection efficacy of discriminative spectral analysis algorithms. This study presents a novel anomaly detection method, drawing methodological inspiration from the Reed–Xiaoli (RX) dual-window framework. An innovative endmember learning mechanism is employed to extract the endmembers for the designated pure background and mixed environmental areas, facilitating an effective discriminative hybrid spectral analysis model for anomaly target detection. Comparative experiments were conducted with three classical methods, in publicly and field-collected hyperspectral datasets, the proposed method consistently achieved higher AUC scores on images with resolutions ranging from low to high, enabling effective unsupervised anomaly detection in hyperspectral remote sensing.