RSCDG: Remote Sensing Change/Damage Image Generator Based on Prior Foundation Model and Multimodal Reference Information

Abstract. In scenarios such as natural disasters, military conflicts, and rapid urban expansion, high-quality post-event remote sensing images are often difficult to obtain in a timely manner, limiting the training and application of change detection, damage assessment, and related interpretation models. To address this issue, this paper proposes RSCDG, a remote sensing change/damage image generation framework based on prior foundation models and multimodal reference information. Built on a pretrained latent diffusion model, RSCDG integrates three types of conditional information: a Pre-event Visual Prompt Adapter extracts structural priors from the pre-event image via Prithvi-EO-2.0 to preserve background stability in unchanged regions; a Spatial Location Control Pathway introduces the change/damage mask into a ControlNet branch to improve spatial precision; and a Generation Content Text Controller uses a CLIP text encoder to guide semantically consistent generation. In addition, a Mask Alignment Loss is introduced to align the change patterns of generated and real images under the supervision of a frozen change detection model. Experiments on the LEVIR-MCI change scenario and the CEBD earthquake damage scenario show that RSCDG consistently outperforms ControlNet. In the change scenario, it achieves an FID of 28.92, an IS of 9.62, and a KID of 0.0139; in the damage scenario, the corresponding values are 37.82, 8.05, and 0.0187, respectively. Ablation results further confirm the effectiveness of the Mask Alignment Loss. Overall, RSCDG provides a practical solution for post-event sample construction, change detection data augmentation, and controllable sample generation for damage assessment.