Transformer-Based Differential Superpixel Selection Network for Remote Sensing Images Change Detection

In remote sensing images change detection, the results of Transformer often face the low precision and edge fitting problems, which are caused by the highly complex background of building categories and the imbalance of sample ratio. To address these challenges, we propose a Transformer-based differential superpixel selection network. First, a differential superpixel selection module is proposed which introduces difference selecting to refine the edge feature representation of the object region by using the boundary information of the selected superpixel. Second, we propose a multi-scale feature-aware enhancement module, which fuses multi-scale feature differences with the features extracted by the hierarchical Transformer encoder. This module enhances difference awareness while suppressing interference from complex backgrounds. Finally, we construct a composite loss function that combines the focal loss function with the binary cross-entropy loss function. The composite loss function can pay more attention to hardly classifying samples and suppress the influence of easily classifying samples. Our proposed algorithm is tested on the widely used LEVIR-CD and DSIFN-CD datasets. The results show that it effectively improved detection accuracy and edge fitting compared to existing methods.