The Capacity of Generative Models to Synthesize Regional Landslide and Non-Landslide Remote Sensing Imagery Under Data-Scarce Scenarios: Insights from Multimodal Foundation Models

Landslide interpretation based on remote sensing data is pivotal for efficient emergency response and risk management. However, the scarcity of high-quality landslide data remains a major bottleneck for data-driven landslide analysis. To address this challenge, this study investigates the potential of multimodal foundation models for generating high-quality synthetic landslide and non-landslide remote sensing images. The proposed regional remote sensing image synthesis framework based on Stable Diffusion models and Low-Rank Adaptation enables more controllable and interpretable remote sensing data augmentation under data-scarce scenarios. Based on the publicly available Bijie landslide dataset, landslide and non-landslide remote sensing image–semantic annotation databases can be separately constructed and subsequently utilized to fine-tune text-to-image diffusion models. By conducting comparative experiments across three Stable Diffusion backbones, the performance of the generative models in both landslide and non-landslide scenarios is systematically and quantitatively evaluated. Experimental results demonstrate that LoRA fine-tuning can effectively transfer landslide-specific visual knowledge into diffusion models, enabling the generation of high-fidelity synthetic remote sensing images with texture and structure closely matching real samples. Compared with the StyleGAN2 baseline with a minimum FID of 67.47 in the recent literature, the proposed SDXL-LoRA model achieves superior generation quality with a minimum FID of 54.70. In addition, the study indicates that the optimal diffusion backbone depends on semantic complexity. Accordingly, a heterogeneous backbone strategy should be adopted when constructing balanced synthetic datasets for downstream applications. The training configurations employed in this study also provide a practical reference for related research. This study exploratorily applies multimodal generative foundation models to landslide-related remote sensing data augmentation and provides a flexible and transferable solution for regional geohazard studies.