TDGNet: Temporal Difference-Guided Semantic Change Detection in Remote Sensing Images

Abstract

Semantic change detection (SCD) aims to identify changed regions in bi-temporal remote sensing images and further recognize their semantic “from–to” transitions. However, this task remains challenging because real land-cover changes are often weak, spatially fragmented, and highly imbalanced against dominant unchanged areas. Existing SCD methods usually rely on implicit temporal fusion or uniformly applied cross-temporal interaction. Such designs may dilute subtle difference cues, blur change boundaries, and propagate pseudo changes caused by illumination, phenology, or imaging conditions. To address these issues, this letter proposes TDGNet, a temporal difference-guided network that explicitly uses bi-temporal difference information to guide feature enhancement, temporal interaction, and change refinement. Specifically, TDGNet consists of a difference-based edge–feature (DBEF) module, a change-gated cross-covariance transposed attention (CCTA) module, and a multiscale boundary-aware change (MBAC) module. To preserve discriminative change evidence, DBEF converts the temporal difference into channel-spatial attention, directional change cues, and a Laplacian boundary prior. Furthermore, CCTA performs efficient channel-wise cross-temporal interaction and uses the predicted change feature as a gate, enabling semantic reconciliation mainly in unchanged regions while avoiding excessive smoothing over true transitions. MBAC then amplifies weak and scale-varying change responses through edge-enhanced multiscale gates. Extensive experiments on SECOND and Landsat-SCD demonstrate the effectiveness of TDGNet. It achieves 23.69% separated kappa (SeK) and 73.54% mean intersection-over-union (mIoU) on SECOND, and 69.16% SeK and 92.17% $F_{scd}$ on Landsat-SCD, with only 1.63 M additional parameters and 17.39 G additional FLOPs over the baseline.