An Interactive Prompt Based Network for Urban Floods Area Segmentation Using UAV Images

Abstract:

Rapid and accurate detection of flood-affected areas is critical for effective disaster response and urban management. This study presents an interactive prompt-based network for segmenting urban flood regions from Unmanned Aerial Vehicle (UAV) images. The proposed framework combines color-space analysis in HSV and CIE-Lab color models with prompt-guided attention to precisely identify water-covered regions in complex urban scenes. Foreground and background prompt points are automatically sampled from ground-truth masks and encoded as a three-channel prompt map, which guides the segmentation decoder. A multi-scale dilation strategy progressively refines flood boundaries by aggregating contextual features at multiple spatial scales. Fusion weights between HSV and Lab feature streams are optimized on a validation set to maximize Intersection-over-Union (IoU). Data augmentation including flipping, rotation, and random scaling expands training diversity. Gradient-weighted Class Activation Maps (Grad-CAM) provide visual interpretability of learned feature regions. Experiments demonstrate that the model achieves strong quantitative performance across accuracy, precision, recall, and F1 score, confirming its suitability for real-world UAV-based urban flood monitoring and emergency decision support systems.

Keywords— UAV image segmentation, Urban flood detection, Prompt-based network, multi-scale dilation, Disaster remote sensing.