Detecting Spatial-temporal urban ponding distribution from surveillance videos based on computer vision

Detection of ponding levels timely and accurately during urban floods is the basis of effective disaster prevention and mitigation. New data sources such as social media and road surveillance video record the process of urban floods in the form of images, and the development of computer vision technology bring new opportunities for extracting ponding information from these image data. This study proposes a computer vision-based method to identify the spatial-temporal distribution of ponding levels in the scene of the road surveillance video. Firstly, a dataset of sedan images compiled from three sources was collected to train an object detection algorithm, You Only Look Once vision 3 (YoloV3). Then the trained YoloV3 model was adopted to identify the ponding levels whenever and wherever sedans were detected from the videos. Secondly, the outlier detection was employed to detect and delete the outliers of ponding levels in each time step. Finally, Inverse Distance Weighted was leveraged to estimate the ponding distribution in the scene. This method was employed for two urban flood events at a street crossing, Dongguan Street, in Dalian, China. The evaluation index mAP of the trained YoloV3 model reached 78%, which stated the model’s validity. The ponding level estimated by our method was validated well with the submerged depth of a static reference, and ponding process had a strong correlation with the rainfall time series. The results can be used to analyze the process of flood rising and receding, which contributes to arrange drainage facilities and improve urban flood management.