TRSR-Net: a lightweight remote sensing change detection network via temporal response fusion and context-edge calibration

Abstract Remote sensing change detection (RSCD) is a fundamental task for monitoring surface changes from multi-temporal remote sensing images. However, existing lightweight RSCD methods still face challenges in effectively modeling complex bi-temporal interactions and maintaining consistent multi-scale feature representations under diverse change patterns and heterogeneous scenarios. To address these issues, this paper proposes the Temporal Response and Structural Refinement Network (TRSR-Net), a lightweight RSCD framework that integrates temporal response fusion and context-edge calibration. Specifically, a Reciprocal Temporal Response Fusion (RTRF) module is designed to characterize complementary temporal responses between bi-temporal features by considering difference, energy, and agreement information. Through lightweight channel-spatial routing, RTRF selectively aggregates informative temporal cues and enhances the structured representation of bi-temporal interactions. Furthermore, a Context-Edge Calibration Bridge (CECB) is introduced to refine multi-scale features before decoding by jointly incorporating contextual aggregation, edge guidance, and semantic calibration, thereby improving feature consistency and structural representation quality. Extensive experiments on LEVIR-CD, SYSU-CD, and UAV-CD demonstrate that TRSR-Net achieves F1-scores of 91.66, 84.25, and 77.05%, respectively. With 4.25 M parameters and 3.99 G FLOPs, TRSR-Net achieves a favorable accuracy-efficiency trade-off, demonstrating its potential for resource-limited remote sensing change detection applications.