HCDFNet: hybrid cross-attention with dynamic perception-aware fusion for remote sensing change captioning

Remote Sensing Image Change Captioning (RSICC) generates descriptive sentences for content differences in bitemporal remote sensing images. However, current methods have some weaknesses in dynamic spatial hierarchy modeling. To address limited multi-scale fusion and weak local-global interaction, we propose a Hybrid Cross-Attention Dynamic Fusion Network (HCDFNet), which leverages hybrid attention-guided feature interaction to enhance the representation and semantic coherence of change regions. First, ResNet101 was used to extract features from bi-temporal images, whereas a differential encoder selectively captured critical change information between temporal features; second, a stacked hybrid cross-attention transformer architecture was constructed. Through a dual-branch collaborative mechanism combining global contextual modeling and local detail refinement, this synergistic processing strategy not only improves the semantic coherence of remote sensing image change interpretation but also significantly enhances boundary delineation accuracy in complex scenarios; then, the dynamic perception-aware fusion module was designed, which implemented cross-scale information integration via feature splicing and attention-based weighting, thereby significantly improving expressive accuracy and contextual comprehension of change regions. Through extensive experiments conducted on the LEVIR-CC dataset, HCDFNet demonstrated superior performance over state-of-the-art methods in terms of computational efficiency and the comprehensive evaluation metric Sm*, achieving an improvement of 2.23 over the baseline in the Sm*. We will make our codebase publicly available to facilitate future research at https://github.com/chendengxiang/HCDFNet