BWFNet: Bitemporal Wavelet Frequency Network for Change Detection in High-Resolution Remote Sensing Images

Accurate change detection in high-resolution remote sensing images is essential for a wide range of Earth observation tasks. However, current deep learning (DL)-based methods often struggle with detecting subtle changes, maintain limited robustness to temporal and environmental variations, and face challenges in balancing global semantic understanding with precise boundary localization. To address these issues, we propose bitemporal wavelet frequency network (BWFNet), a novel hybrid architecture that integrates four dedicated modules to systematically improve change detection performance. Specifically, the cosine directional convolution module (CDCM) enhances the extraction of directional and structural features, while the bitemporal cross-modulation mechanism (BCMM) adaptively fuses semantic information from bitemporal images to emphasize relevant changes. The multiscale feature refinement module (MSFM) aggregates and refines features at multiple scales for comprehensive spatial representation, and the wavelet frequency attention mechanism (WFAM) selectively highlights discriminative frequency components via wavelet decomposition to improve sensitivity to subtle and complex changes. Extensive experiments on four public remote sensing change detection benchmarks (LEVIR-CD, LEVIR-CD+, WHU-CD, and SYSU-CD) demonstrate that BWFNet achieves state-of-the-art performance and strong generalization across diverse and challenging scenarios.