Dual-Scale Pixel Aggregation Transformer for Change Detection in Multitemporal Remote Sensing Images

Transformers have recently been applied to change detection (CD) of multitemporal remote sensing images because of their ability to model global information. However, the rigid patch partitioning in vanilla self-attention destroys spatial structures and consistency in observed scenes, leading to limited CD performance. In this paper, we propose a novel dual-scale pixel aggregation transformer (DSPA-Former) to mitigate this issue. The core of DSPA-Former lies in a dynamic superpixel tokenization strategy and bidirectional dual-scale interaction within the learned feature space, which preserves semantic integrity while capturing long-range dependencies. Specifically, we design a hierarchical decoder that integrates multiscale features through specialized mechanisms for pixel superpixel dialogue, guided feature enhancement, and adaptive multiscale fusion. By modeling the homogeneous properties of spatial information via superpixel segmentation, DSPA-Former effectively maintains structural consistency and sharpens change boundaries. Comprehensive experiments on the LEVIR-CD, WHU-CD, and CLCD datasets demonstrate that DSPA-Former achieves superior performance compared to state-of-the-art methods, particularly in preserving the structural integrity of complex change regions.