SCAFNet: A Semantic Compensated Adaptive Fusion Network for Remote Sensing Images Change Detection
Yunzuo ZHANG, Jiawen Zhen, Shibo Sun, Ting Liu, Lei Huo, Tong Wang
Abstract
Current CNN-Transformer hybrid methods for remote sensing change detection aim to address the limitations of CNNs’ constrained receptive fields and Transformers’ local detail insensitivity. However, these methods suffer from semantic misalignment and non-adaptive fusion between the dual branches, resulting in persistent sensitivity to pseudo-changes. To address these issues, we propose SCAFNet, a Semantic Compensated Adaptive Fusion Network, featuring three core components: 1) The Semantic Compensation Module (SCM) that aligns local-global features via cross-attention to resolve spatial-semantic mismatches; 2) The CNN-Transformer Feature Adaptive Fusion (CTFAF) module improving feature integration by dynamically balancing the local details of CNN and the global context of Transformer through cross-branch attention interactions and dynamic parameterization; 3) The Change Feature Identification Module (CFIM) computing channel and spatial weights, enhancing true changes while suppressing disturbances such as seasonal variations. Experiments on CDD and WHU-CD datasets demonstrate SCAFNet’s superior robustness and accuracy, outperforming existing methods through effective feature fusion. The source code and supplementary materials will be made available at https://github.com/Gyroprime/SCAFNet.