DiffRegCD: Integrated Registration and Change Detection with Diffusion Features

The introduction of DiffRegCD marks a significant advancement in change detection (CD) methodologies, which are crucial for various applications such as environmental monitoring, disaster response, and urban development. Traditional CD models face challenges due to real-world imagery misalignments caused by parallax, viewpoint shifts, and temporal gaps. DiffRegCD addresses these issues by integrating dense registration and change detection into a single framework, reformulating correspondence estimation as a Gaussian smoothed classification task. This innovative approach not only achieves sub-pixel accuracy but also ensures stable training, making it robust against variations in illumination and viewpoint. Extensive experiments on datasets like LEVIR-CD, DSIFN-CD, WHU-CD, SYSU-CD, and VL-CMU-CD demonstrate that DiffRegCD consistently surpasses recent baselines and remains reliable under wide temporal and geometric variations, positioning it as a promising tool for future applications i…