Reliability-Enhanced State-Space Interaction for Robust Remote Sensing Change Detection
Abstract
Remote sensing change detection aims to identify land-cover changes from bitemporal images acquired over the same area. Although existing deep learning methods have achieved promising performance, they still suffer from pseudochanges caused by illumination variation, seasonal difference, shadows, sensor inconsistency, and slight spatial misalignment. Moreover, most bitemporal interaction strategies propagate temporal discrepancies without explicitly judging their reliability, which may amplify appearance-induced perturbations as change cues. To address these issues, we propose a reliability-aware change-guided state-space interaction framework for robust remote sensing change detection. Specifically, we first use temporal-consistent feature preparation (TCFP) to produce temporally comparable and locally refined hierarchical features. Then, a reliability-enhanced change-guided interaction module is designed to estimate change-evidence reliability and appearance-perturbation uncertainty from complementary state-space discrepancy representations. These reliability measurements are further introduced into the selective state-space interaction to regulate state writing, reading, and updating during bitemporal feature propagation. Finally, a lightweight hierarchical decoder combines the fused change representations and TCFP-prepared temporal detail cues to generate the final change map. Extensive experiments on six benchmark datasets demonstrate the effectiveness of the proposed method, achieving intersection over union/F1 scores of 90.61%/95.07%, 84.29%/91.48%, 90.39%/94.95%, 55.75%/71.59%, 91.67%/95.65%, and 53.50%/69.71% on BCDD, LEVIR-CD, GoogleBuilding, CLCD, DSIFN, and SECOND, respectively.