Abstract. Rapid and accurate acquisition of urban flood information is crucial for flood prevention, disaster mitigation, and emergency management. With the development of mobile internet, crowdsourced images on social media have emerged as a novel and effective data source for flood information collection. However, selecting appropriate targets and employing suitable methods to determine flooding level has not been well investigated. You Only Look Once version 8 (YOLOv8) is a convolutional neural network-based computer vision model that has been widely applied in image recognition tasks due to its end-to-end architecture and high computational efficiency. This study proposes a method to assess urban flood risk levels based on the submerged status of buses captured in social media images. First, a dataset containing 1008 images in complex scenes is constructed from social media. The images are annotated using Labelimg, and expanded with a data augmentation strategy. Four YOLOv8 configurations are validated for their ability to identify urban flood risk levels. The validation process involves training the models on original datasets, augmented datasets, and datasets representing complex scenes. Results demonstrate that, compared to traditional reference objects (e.g., cars), buses exhibit greater stability and higher accuracy in identification of urban flood risk levels due to their standardized height and widespread presence as they remain in service during flood events. The data augmentation strategy enhances the model's mAP50 and mAP50-95 metrics by over 10 % and 20 %, respectively. Additionally, through comparative analysis of YOLOv8 configurations, YOLOv8s demonstrates superior results and achieves an effective balance between accuracy, training time, and computational resources, recommended for the identification of urban flood risk levels. This method provides a reliable technical foundation for real-time flood risk assessment and emergency management of urban transportation systems, with substantial potential for practical applications.
Qiu et al. (Tue,) studied this question.