DBAF-Net: Dual Branch Alignment and Fusion Network for Remote Sensing Change Detection

The field of remote sensing change detection (RSCD) has witnessed remarkable progress. Most existing RSCD methods currently follow the encoder-feature interaction-decoder architecture. However, these methods often treat spatial details and semantic information in a combined manner during the feature interaction stage, leading to insufficient extraction of spatial and semantic features, while overlooking the distribution inconsistency between these features. Moreover, there is a lack of a unified alignment and fusion strategy. To address these issues, we propose a novel dual branch alignment and fusion network (DBAF-Net). Specifically, the network is designed with a dual-branch structure in the feature interaction stage to achieve efficient differential feature extraction, where the enhancement space difference module and semantic-aware difference module separately extract spatial and semantic features from the image. To alleviate the feature distribution discrepancy between the two branches, we design a cross-branch alignment fusion decoder, which leverages domain adaptation and differential attention mechanisms for the deep fusion of spatial and semantic features. Finally, the refined features are progressively aggregated to generate the final change map from deep to shallow layers. Comprehensive evaluations performed on three well-established benchmark datasets—LEVIR-CD, SYSU-CD, and UAV-CD—indicate that the proposed DBAF-Net achieves superior change detection performance compared to current state-of-the-art methods.