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

As climate change intensifies, extreme weather events like floods are occurring with increasing frequency. While data-driven deep learning methods are effective for extracting flood disaster information, their efficiency is constrained by the scarcity of postdisaster samples, the high cost of annotations, and the models’ strong dependence on both the quantity and quality of data. This study introduces an interactive semantic segmentation model based on multisource UAV flood images, incorporating four types of prompts. By embedding expert knowledge into the prompt design, the model reduces annotation costs and enhances generalization capabilities. First, a prompt encoder is developed to map different types of prompt information into a three-channel space using convolutional techniques, thereby reducing sample labeling costs. Moreover, an image encoder that integrates Mamba and convolution is developed to effectively extract global spatial and channel features from flood images while minimizing computational load. Finally, a spatial and channel attention module with residual connections is introduced to enable multiscale fusion and filtering of prompt information and image features across both spatial and channel dimensions, improving the utilization of prompt information. To validate the model's performance, we conduct experiments using UAV flood imagery collected from diverse regions, backgrounds, and angles. The results demonstrate that, under consistent prompt conditions, our model extracts flood areas more efficiently, reducing misclassification and omission errors. Compared with the next best benchmark model, the intersection over union for the flood category improves by at least 3.75%.