Hierarchical cross-temporal frequency network for remote sensing building change detection

Building change detection (BCD) using remote sensing images has become a crucial technique for urban planning. Although deep learning-based methods for BCD have made significant progress in the field of remote sensing imagery, they still encounter several challenges, including unclear detection of boundaries of buildings, susceptibility to noise interference, and missed detection of subtle changes. To address these issues, we propose a hierarchical cross-temporal frequency network (HCTFNet) to achieve more precise and refined building change detection. Specifically, we design a frequency-channel enhancement model, which incorporates a frequency-guided spatial attention block to capture different frequency components of input features, thereby enabling more accurate extraction of textural details and building edges, and a residual channel attention block to assign weights to features from a channel-wise perspective, which suppresses invalid channel features and enhances the model’s capability to detect change regions. We further introduce a cross-attention fusion block to establish spatial correlations between the two temporal phases, strengthen the representation of change regions, and fuse the features to highlight areas with the most significant changes. Experimental results show that HCTFNet achieves F1-scores of 92.10% on the LEVIR-CD dataset, 89.58% on the WHU-CD dataset, and 80.10% on the SYSU-CD dataset, outperforming most of the existing mainstream change detection methods.