부산대학교 도서관

Change detection (CD) methodologies have garnered substantial attention owing to their capability to monitor alterations in geographic spaces across temporal intervals, especially with the acquisition of high-resolution remote sensing images. However, challenges persist due to dissimilar imaging conditions and temporal windows. Although deep-learning architectures have shown promise in addressing challenges in CD, many existing methods struggle to capture long-range dependencies and local spatial information effectively. The current CD methods rely heavily on pure CNNs and transformers, which employ only single-pass forward propagation. This approach leads to inadequate utilization of feature information, resulting in inaccurate CD maps, particularly when discerning edges. To overcome these limitations, we propose a transformer-based conditional generative diffusion method for CD, named TransC-GD-CD, tailored for RS data. This approach leverages the numerous sampling iterations of the DDPM, contributing to the generation of high-quality CD maps. In addition, the frequency cross transformer mechanism seamlessly amalgamates CD condition with the noise feature within the DDPM. The innovative mechanism effectively bridges diffusion noise and conditional semantic terrains. Moreover, a novel multitype difference extraction module, named appear–disappear–concat, is devised to partition the CD task to optimize both segmentation extraction and CD classification, overcoming the persistent challenge of information loss endemic to conventional CD algorithms, such as simple subtraction. We demonstrate the superiority of TransC-GD-CD by comparing the experiment results against various algorithms across three widely used CD datasets, namely CDD, WHU, and LEVIR.