A Transformer-Based Multiscale Difference Enhancement Network for Change Detection

Abstract

Change detection (CD) is an important research field in remote sensing, aimed at identifying differences in multitemporal images. Despite the progress made by convolutional neural networks and Transformer architectures in visual analysis, challenges remain in achieving robust feature representation and global contextual understanding. To address these issues, we propose a transformer-based multiscale difference enhancement network (TMDENet). Our approach utilizes a multiscale feature extraction module to capture diverse feature representations, while incorporating a channel-spatial cooperation mechanism for refined detail enhancement. The extracted features are encoded into semantic tokens, which are processed by a Transformer to capture long-range dependencies. This is further complemented by a cross-hierarchical linear fusion module for multiscale feature integration. Additionally, a difference enhancement module isolates and emphasizes change-related features. Extensive evaluations on benchmark datasets (LEVIR, CDD, DSIFN, WHUCD) show that TMDENet achieves state-of-the-art results in boundary delineation and change localization. This study establishes TMDENet as a robust framework for high-resolution remote sensing CD, offering significant improvements in both precision and reliability.