Superpixel Edge Refinement and Differential Feature Enhancement for Transformer-Based Remote Sensing Image Change Detection

Change detection in remote sensing images is a challenging task, particularly when identifying small and complex objects such as buildings. The task is often hindered by issues such as high fragmentation rates, blurred edges, and elevated false and missed detection rates. These issues arise due to the inherent complexity of the scenes and the problem of class imbalance. To address these challenges, we propose a superpixel edge refinement and differential feature enhancement for Transformer-based change detection networks. First, a region-enhanced Transformer encoder is introduced, incorporating a superpixel edge refinement module. This module leverages pixel-level features extracted by the Transformer to generate superpixel feature embeddings, thereby enhancing local and edge information. Second, a differential feature enhancement module is designed to emphasize true change regions while suppressing pseudo-change areas. This is achieved through a dual-branch structure comprising feature enhancement and change extraction, further augmented by a saliency module. Finally, a Dice loss function is employed as an auxiliary loss term, complementing the cross-entropy loss in the construction of an edge-optimized loss function. The resulting loss function aims to balance overall detection accuracy with edge refinement. The proposed algorithm was evaluated on the widely used datasets LEVIR-CD and DSIFN-CD. Compared to existing algorithms, the proposed method effectively improved edge detail and overall consistency while maintaining detection accuracy. Additionally, false detections and missed detections were reduced.