Content Adaptive Transformer for Change Detection in High-Resolution Remote Sensing Images

Transformer-based methods have demonstrated remarkable performance in change detection (CD) of highresolution remote sensing images (HRRSIs), with window-based self-attention mechanisms gaining prominence due to its lower memory requirements. However, these methods are fundamentally limited by their fixed window configurations, which are inherently data-agnostic and lack content adaptability, thereby limiting the interaction of intra- and inter-diverse geographical content. To overcome these limitations, we propose a novel content-driven framework, called content adaptive Transformer network (CATNet), which mainly consists of region-level content adaptive Transformer (RCAT) and global-level content adaptive Transformer (GCAT) to extract local and global features based on image content. The RCAT facilitates the interaction of local self-similarity information by separately processing intra-similar content of images. This method can adaptively represent complex spatial structures by focusing exclusively on regions of similar content and excluding dissimilar content, which is critical for accurate HRRSI CD. The GCAT efficiently captures global dependencies of entire images by modeling spatial relationships from inter-similar content. This manner extracts more representative global features by processing similar content of images rather than individual pixels. It captures rich context information of images while reducing pseudo changes of image content. Extensive experiments across three benchmark datasets verify that CATNet consistently outperforms eleven state-of-theart methods, achieving better performance in quantitative metrics and visual quality assessments