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

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Segmentation models struggle with diverse flood imagery.
• Lack of annotated real-world data hinders model generalization.
• Unmanned Aerial Vehicle (UAV) imagery offers valuable flood monitoring data.

Purpose of the Study:

• To develop a novel methodology for generating and filtering synthetic UAV flood imagery.
• To enhance the generalization capabilities of image segmentation models for flood detection.
• To improve model robustness by effectively combining real and synthetic training data.

Main Methods:

• Utilized text-to-image synthesis and image inpainting for synthetic data generation.
• Employed unsupervised pseudo-labeling for automatic segmentation mask creation.
• Implemented outlier detection in feature space for synthetic data filtering and realism enhancement.

Main Results:

• Synthetic data closely matched real data in training performance.
• Combining real and synthetic data improved model robustness by 1-7%.
• Prompt design significantly impacted the visual fidelity of generated images.

Conclusions:

• The proposed framework effectively generates realistic synthetic UAV flood imagery.
• Synthetic data augmentation enhances the generalization and robustness of flood segmentation models.
• The methodology provides a viable solution for data scarcity in remote sensing applications.