AI Generates Synthetic Sand Boil Images for Levee Inspection

Detecting safety-critical defects like sand boils on earthen levees is crucial, but pixel-level annotation for these defects is severely limited due to their infrequent occurrence. To overcome this data scarcity, a new diffusion-based synthesis pipeline has been developed to generate high-quality synthetic imagery of sand boils. The pipeline leverages Stable Diffusion XL, fine-tuned with DreamBooth, and employs a multi-branch ControlNet stack for precise conditioning. It generates synthetic inspection images from a small set of real reference images. A key innovation is a soft-mask inpainting protocol that preserves real defect pixels while seamlessly re-rendering the surrounding scene, avoiding visual artifacts common in other compositing methods. The system can also generate new boils within a specified mask, effectively creating segmentation labels by design. From the initial real training images, the pipeline produced over 1,000 synthetic candidates, with 815 passing a CLIP admissibility filter for quality. While various presets offer trade-offs between fidelity, diversity, and label reliability, the research focuses on image quality and diversity, leaving downstream segmentation for future work.