AdaptFormer: An Adaptive Hierarchical Semantic Approach for Change Detection on Remote Sensing Images

Change detection (CD) in remote sensing (RS) aims to consistently track alterations in specific regions over time. While current methods employ hierarchical architectures to analyze semantic details, they often miss crucial changes across different semantic levels, resulting in partial representations of environmental shifts. Addressing this, we propose AdaptFormer, uniquely designed to adaptively interpret hierarchical semantics. Instead of a one-size-fits-all approach, it strategizes differently across three semantic depths: employing straightforward operations for shallow semantics, assimilating spatial data for medium semantics to emphasize detailed interregional changes, and integrating cascaded depthwise attention for in-depth semantics, focusing on high-level representations. The experimental evaluations reveal that AdaptFormer surpasses many leading benchmarks, showcasing exceptional accuracy on LEVIR-CD and DSIFN-CD datasets. AdaptFormer showcases impressive performance with F1 and intersection over union (IoU) scores of 92.65% and 86.31% on the LEVIR-CD dataset, and 97.59% and 95.29% on the DSIFN-CD dataset, respectively. The datasets are available at https://github.com/aigzhusmart/AdaptFormer.