Text-Prompt-Based AI-Generated Virtual Image Augmentation for Data-Scarce Flood Detection in Urban Flood-Prone Areas

Source abstract

Urban flood detection requires visual examples of flooded streets and alleys, but such event-state images are difficult to collect at scale. This study examines whether sparse real-image training sets can be strengthened using text-prompt-only AI-generated virtual imagery for ground-level flood detection in flood-prone urban areas. Building on AlleyFloodNet, the generated images were used only as condition-specific training augmentation, while validation and testing were conducted exclusively on real images. Across eight ImageNet-pretrained architectures and three random seeds, the results show that virtual imagery is not a substitute for real flood observations. When virtual images dominated the training set, performance declined. However, when a sufficient real-image anchor was available, virtual augmentation improved the highest observed fixed-test performance. The strongest mixed-condition result was obtained by EfficientNet-B2 under Real30_Aug70, reaching 91.00% accuracy, 89.48% flooded-class recall, and 90.79% macro-F1. These findings suggest that prompt-only virtual imagery can help mitigate real-image scarcity in selected training conditions, but its benefit depends on the real-to-virtual composition and model architecture.