Detection of Urban Flood Inundation from Traffic Images Using Deep Learning Methods

Abstract

Urban hydrological monitoring is essential for analyzing urban hydrology and controlling storm floods. However, runoff monitoring in urban areas, including flood inundation depth, is often inadequate. This inadequacy hampers the calibration of hydrological models and limits their capacity for early flood warning. To address this limitation, this study established a method for evaluating the depth of urban floods using image recognition and deep learning. This method utilizes the object recognition model YOLOv4 to identify submerged objects in images, such as the legs of pedestrians or the exhaust pipes of vehicles. In a dataset of 1,177 flood images, the mean average precision for water depth recognition reached 89.29%. The study also found that the accuracy of flood depth recognition by YOLOv4 is influenced by the type of reference object submerged by the flood; the use of a vehicle as the reference object yielded higher accuracy than using a person. Furthermore, image augmentation with Mosaic technology effectively enhanced the accuracy of recognition. The developed method extracts on-site, real-time, and continuous water depth data from images or video data provided by existing traffic cameras. This system eliminates the need for installing additional water gauges, offering a cost-effective and immediately deployable solution.