Closing the Domain Gap: Can Pseudo-Labels from Synthetic UAV Data Enable Real-World Flood Segmentation?

We present a novel methodology for generating and filtering synthetic Unmanned Aerial Vehicle (UAV) flood imagery to enhance the generalization capabilities of segmentation models. Our framework combines text-to-image synthesis and image inpainting, using curated prompts and real-world segmentation masks to produce diverse and realistic flood scenes. To overcome the lack of human annotations, we employ an unsupervised pseudo-labeling method that generates segmentation masks based on floodwater appearance characteristics. We further introduce a filtering stage based on outlier detection in feature space to improve the realism of the synthetic dataset. Experimental results on five state-of-the-art flood segmentation models show that synthetic data can closely match real data in training performance, and combining both sources improves model robustness by 1-7%. Finally, we investigate the impact of prompt design on the visual fidelity of generated images and provide qualitative and quantitative evidence of distributional similarity between real and synthetic data.