Dual-Granularity Feature Alignment for Change Detection in Remote Sensing Images

Deep learning has emerged as the preferred method for remote sensing change detection owing to its ability to automatically extract discriminative features from bitemporal images. However, few methods simultaneously consider heterogeneous appearance of objects and affine geometric difference between bitemporal images, both of which contribute to pseudo change. In this paper, dual-granularity feature alignment (DgFA) is proposed to deal with these two issues. Specifically, bitemporal features extracted by transformer, along with learnable class tokens, are input into the proposed semantic alignment module to adjust the appearance of separate instances from same-category objects to ensure a cohesive style. Then a spatial alignment module is introduced to use the estimated transformation field to accomplish bitemporal feature registration. Finally, we develop a temporal contrast-based change detection head to infer the change map based on dual-granularity aligned bitemporal features and corresponding difference maps. To refine the change map, this head also constrains the feature similarity within changed and unchanged regions across bitemporal features via a contrastive loss. Experimental results demonstrate that DgFA outperforms several state-of-the-art methods on three public benchmark datasets, including LEVIR-CD, CDD, and SYSU-CD