A Generative Change Detection Model Based on Difference-Feature Guided DDPM

Deep learning (DL) approaches, such as CNN and Transformer networks, have shown promise in bitemporal change detection (CD). However, these approaches have limitations in capturing long-range dependencies and incorporating 2D structure and spatial local information, resulting in inaccurate CD maps with discerning edges. To overcome these limitations, this paper presents a novel end-to-end DDPM-based model called change-aware diffusion model (CADM), which introduces three key innovations. Firstly, CADM directly generates CD maps as a generation model. It leverages variational inference, a powerful technique for learning complex probabilistic models, to facilitate the gradual learning and refinement of the model's data representation. This enables CADM to effectively distinguish subtle and irregular buildings or natural scenes from the background. Secondly, CADM introduces an adaptive calibration conditional difference encoding technique. This technique utilizes differences between multi-level features to guide the sampling process, enhancing the precision of the CD map. Lastly, CADM incorporates a noise suppression-based semantic enhancer (NSSE) to improve the quality of the CD map. The NSSE utilizes prior knowledge from the current step to suppress high-frequency noise, enhancing the differential information and refining the CD map. We evaluate CADM on four remote sensing CD tasks with different ground scenarios, including CDD, WHU, Levier, and GVLM. Experimental results demonstrate that CADM significantly outperforms state-of-the-art methods, indicating the generalization and effectiveness of the proposed model.