Guided diffusion-based image editing for landslide image synthesis in remote sensing change detection

Landslide change detection (CD) based on supervised learning plays an important role in natural disaster monitoring and post-disaster assessment. However, its performance is often limited by the scarcity of high-quality landslide change datasets. Even within existing datasets, many image regions contain no landslide changes and thus provide limited supervision for learning landslide change patterns. Furthermore, existing models are often difficult to apply directly to new landslide scenarios, where labeled samples are usually insufficient for effective model training. To address these issues, this paper proposes a task-oriented guided landslide synthesis framework for landslide CD, which converts non-landslide samples into positive landslide-change samples via image editing, thereby alleviating the shortage of positive samples for supervised CD training. Specifically, we introduce a diffusion-based Guided Landslide Generator (GLG) that integrates landslide-specific textual prompts and spatio-temporal metadata to guide landslide appearance generation, moving beyond conventional generic prompt-guided synthesis. In addition, DEM-informed masks are employed in the editing process to constrain the spatial geometry of the generated landslides and improve terrain consistency. Experiments on the large-scale GVLM dataset demonstrate that the proposed method generates diverse and realistic landslide appearances and improves CD performance across multiple networks, with a maximum IoU gain of 2.08 percentage points over the baseline without synthetic data. On the newly constructed HHDL dataset, the synthesized data also improve few-shot CD accuracy, with a maximum IoU gain of 3.63 percentage points over the corresponding baseline, validating the effectiveness of the proposed method. The related code and dataset will be available at https://github.com/yixin28/HHDL-Dataset .