Density-guided blind-spot mamba network for hyperspectral anomaly detection

To address the challenges in hyperspectral anomaly detection, including insufficient background sample purity and anomaly information leakage in traditional blind-spot reconstruction methods, and the failure of existing feature modeling networks to balance the capture of long-range dependencies and computational efficiency, this paper proposes a novel hyperspectral anomaly detection network based on density-guided adaptive blind-spot generation and pyramidal Mamba feature reconstruction (DBSMNet). First, an adaptive blind-spot generation strategy based on DBSCAN density clustering is designed. It screens high-density core samples with strong spectral consistency from the neighboring background via unsupervised clustering to fill blind-spot regions, preventing anomaly information from contaminating background modeling at the data source. Second, a four-level pyramidal Mamba feature reconstruction network is constructed, which integrates the selective state space model with a multi-scale downsampling–upsampling structure. This network accurately models multi-scale spatial-spectral long-range dependencies in hyperspectral data, breaking through two critical bottlenecks, the limited receptive field of convolutional networks and the prohibitive computational cost of self-attention mechanisms. DBSMNet achieves the highest AUC on five datasets and the best average performance across all six datasets.