TDBRNet: A Lightweight Bi-Temporal Difference Modeling and Boundary Refinement Network for Remote Sensing Change Detection

Abstract

Remote sensing change detection requires accurate localization of changes in high-resolution bi-temporal imagery while maintaining a compact computational footprint. Existing lightweight methods often rely on simple absolute differencing and progressive upsampling, which may lead to pseudo-changes, blurred boundaries, and missed small objects. To address these issues, we propose TDBRNet, a lightweight bi-temporal difference modeling and boundary refinement network. The Bi-Temporal Difference Separation Module (TDSM) preserves a stable absolute-difference pathway while jointly modeling temporal discrepancy and consistency cues through cross-temporal interaction, attention-guided separation, and gated residual fusion. The Unified Boundary Refinement Module (UBRM) exploits prediction uncertainty to perform feature-level compensation and high-resolution logit refinement, thereby improving the localization of ambiguous boundaries and small changed regions. With only 2.70 million parameters and 3.30 GFLOPs, TDBRNet achieves 92.78% precision, 90.39% recall, 91.57% F1 score, and 84.45% intersection over union on the LEVIR-CD dataset. These results demonstrate that TDBRNet achieves a favorable balance between detection accuracy and computational complexity.