TriCDNet: A Multi-scale Tri-stream Interaction Network for Building Change Detection

Abstract. Building change detection (BCD) from multi-temporal remote sensing imagery plays a vital role in urban monitoring and land management. However, existing deep learning-based methods still suffer from insufficient semantic differentiation, weak multi-scale consistency, and limited global context modeling. To address these issues, we propose TriCDNet, a multi-scale tri-stream interaction network for accurate and robust building change detection. The network integrates three complementary feature streams—bitemporal features and their normalized difference map—and performs stage-wise feature interaction through a multi-layer Difference-guided Cross-temporal Interaction Module (DCIM). A top-down Feature Pyramid Network (FPN) is employed to aggregate multi-scale information, while a lightweight Transformer-based decoder captures long-range spatial–temporal dependencies for global reasoning. Experiments on three public datasets (LEVIR-CD, WHU-CD, and SYSU-CD) demonstrate that TriCDNet achieves superior accuracy and structural consistency, with IoU of 85.2%, 86.91%, and 71.54%, respectively. The results confirm that each component contributes positively to performance, and the proposed tri-stream framework effectively balances local detail preservation and global semantic coherence, showing strong generalization capability in complex urban scenes.