A Lightweight Difference-Guided Gated Network for Building Change Detection in Remote Sensing Imagery

Abstract

Remote sensing building change detection is critical for urban monitoring and disaster assessment, yet existing deep learning methods suffer from three limitations: the heavy computational cost of Transformer models, the information loss induced by naive differencing, and the tendency of convex gating to degenerate into uniform averaging. To address these issues, we propose LDGNet, a lightweight difference-guided gated network built on a shared-weight U-Net encoder. The network integrates a multi-scale difference-aware context module for bi-temporal feature enhancement and a deep gated modulation module with channel attention for selective fusion. With only 8.25 million parameters, LDGNet achieves competitive performance on the LEVIR-CD and WHU-CD datasets, attaining F1-scores of 91.68% and 92.23%, respectively. Comparative results further show that LDGNet outperforms several advanced methods while using substantially fewer parameters than Transformer-based ChangeFormer. These findings confirm that difference-space context modeling combined with residual gated fusion provides an efficient and lightweight solution for high-precision change detection.