DABF-NET: Difference Aggregation and Boundary-Aware Fusion for Change Detection

Abstract

Building change detection in high-resolution remote sensing imagery remains challenging because of weak difference representation, background interference, and blurred boundaries. We propose DABF-Net, a Difference Aggregation and Boundary-Aware Fusion Network with a shared weight MobileNetV2 FPN encoder. The Difference Aggregation Module (DAM) preserves explicit discrepancies while mixing bitemporal context and multiplicative correlations. The Change-Guided Adaptive Fusion Module (CGAF) reuses a coarse change prior to jointly regulate semantic fusion and detail enhancement. The Boundary-Aware Difference Refinement Module (BADRM) converts the local residual of a coarse prediction into boundary feedback for shallow temporal details without auxiliary boundary supervision. DABF-Net achieves F1/IoU scores of 92.05%/85.26% and 94.28%/89.18% on LEVIR-CD and WHU-CD, respectively. Ablation studies verify the individual and joint effectiveness of the three modules. Five complete runs on each dataset yield F1 scores of 91.90 ± 0.09% and 94.46 ± 0.32% on LEVIR-CD and WHU-CD, respectively.