SI2FM: SID Isolation Double Forest Model for Hyperspectral Anomaly Detection

Hyperspectral image (HSI) anomaly detection (HSI-AD) has become a hot issue in hyperspectral information processing as method for detecting undesired targets without priori against unknown background and target information, which can be better adapted to the needs of practical applications. However, demanding environment with no prior small targets, well large data high redundancy HSI itself, make study HSI-AD very challenging. First, we propose an based on nonsubsampled shearlet transform (NSST) domain spectral divergence isolation double forest (SI2FM) this paper. Further, excavates intrinsic deep correlation properties between NSST subband coefficients two ways provide synergistic constraints guidance prediction abnormal coefficients. On one hand, “difference band” guide, global local models are constructed (SID) attribute values difference band low-frequency high-frequency subbands, scores determined by evaluating path lengths binary tree nodes model obtain progressively optimized map. other relationship spatial-spectral dimensions, three-dimensional structure is realize co-optimization multiple maps obtained from forest. Finally, suppresses noise interference certain extent, enhancing separability background. The two-branch collaborative optimization coefficient mining enables sample gradually improved perspectives, effectively improves accuracy detection. effectiveness algorithm verified comparing real datasets captured four different scenes eleven typical algorithms currently available.