Answers are grounded in this paper and cite their sources.
Accurately detecting subtle land-cover changes and delineating precise boundaries in high-resolution bitemporal remote sensing images remains challenging due to complex spatial structures and background variations. We propose a structural residual-edge density network (SREDNet) for end-to-end change detection. SREDNet integrates two complementary modules: a local structural residual module that suppresses smooth background responses and refines local structural residuals through learnable channel-wise gating and an edge-density perception module that jointly models edge-sensitive responses, regional structural context, and spatial attention to strengthen boundary-consistent representations. A hybrid difference construction strategy further combines complementary bitemporal discrepancy cues through adaptive weighting and a top-down VerticalFusion pathway aggregates hierarchical features for cross-scale interaction. Experiments on LEVIR-CD and WHU-CD demonstrate that SREDNet achieves competitive performance in F1 and Intersection over Union. Ablation studies, boundary-F1-score-based boundary evaluation, Wilcoxon signed-rank tests, and efficiency analyses further validate the complementary contributions of the proposed modules and the statistical significance of the observed improvements.
No takes yet. Share an insight, caveat, or question.
Gong et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: