S2OMNet: A Spectral-Saliency Guided Orthogonal Mamba Network for Hyperspectral Image Change Detection

Abstract

Hyperspectral image change detection (HSI-CD) identifies land-cover changes by exploiting bitemporal spatial-spectral information. Although state space models (SSMs) provide efficient long-range sequence modeling, generic spatial scanning strategies mainly construct sequences over the spatial grid and do not explicitly organize spectral-feature responses as state-space sequences or emphasize subtle bitemporal spectral differences of hyperspectral data. To address these issues, we propose the spectral-saliency guided orthogonal Mamba network (S2OMNet), a dual-stream framework that separates explicit spectral-difference enhancement from spatial-spectral dependency modeling. In the difference enhancement stream, the spectral difference saliency encoding block refines change-sensitive difference features through efficient channel attention-based channel recalibration and nonlinear projection. In the interaction stream, the orthogonal spatial-spectral interaction module first factorizes spatial context into two orthogonal directional representations and then performs bidirectional selective scanning along the spectral-feature dimension, modeling long-range spectral dependencies while retaining directional spatial context. The two representations are integrated through a lightweight dual-context aggregation module. Experiments on three public HSI-CD datasets show that S2OMNet achieves the highest overall accuracy (OA), Kappa, and F1-score among the evaluated methods on all three datasets, with average values of 97.02%, 0.8760, and 89.93%, respectively. Its F1-score exceeds the best baseline result on each dataset by an average of 2.00 percentage points, while maintaining low computational cost among the evaluated patch-wise methods.