Hierarchical Diffusion Model for Remote Sensing Image Change Detection

Over several years, deep learning-based methods have made significant progress in the field of remote sensing image (RSI) change detection (CD). However, challenges such as multiscale object changes and cluttered backgrounds, including seasonal variations, shadows, and vegetation color changes, often result in false positives and missed detections, which substantially affect model performance. To tackle these limitations, we present a hierarchical diffusion model for CD (HDM-CD) in RSIs, which effectively integrates spatial and frequency domain information to eliminate false positives and pseudochanges in practical applications. Specifically, we propose a hierarchical feature representation diffusion model that analyzes the complementary characteristics of multilevel information in both spatial and frequency domains. This model successfully combines the capture of local details with the perception of global structures, achieving a comprehensive representation of multiscale spatial–frequency features. Furthermore, we present an uncertainty-guided CD model that directs the model’s focus on learning the features of change regions while simultaneously enhancing the network’s ability to represent uncertain regions and complex boundaries. Finally, we design a spatial–frequency joint optimization module (SFJOM) to mitigate the information loss in small object change areas and accurately detect object contours and sharp edges in the change areas. Extensive experimental results indicate that the proposed HDM-CD method attains state-of-the-art (SOTA) CD performance, exceeding existing competitors by +8.21%, +44.3%, and +17.1% in intersection over union (IoU) on LEVIR-CD, DSIFN-CD, and WHU-CD dataset, respectively.