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

Deep learning (DL)-based methods have recently shown great promise in bitemporal change detection (CD). However, most existing methods are ineffective in simultaneously capturing long-range dependencies and exploiting local spatial information, resulting in inaccurate CD maps with discerning edges. To overcome these obstacles, a novel Denoising Diffusion Probabilistic Model (DDPM)-based generative CD approach called GCD-DDPM is proposed for remote sensing data. More specifically, GCD-DDPM is designed to directly generate CD maps by leveraging variational inference, which enables GCD-DDPM to accurately distinguish subtle and irregular buildings or natural scenes from the background. Furthermore, an adaptive calibration conditional difference encoding technique is proposed for GCD-DDPM to enhance the CD map through guided sampling of the differences among multi-level features. Finally, a noise suppression-based semantic enhancer (NSSE) is devised to cope with the high-frequency noise incurred in the CD map by capitalizing on the prior knowledge derived from the current step. Extensive experiments on four CD datasets, namely CDD, WHU, Levier and GVLM, confirm the good performance of the proposed GCD-DDPM.