Estimating the spatial-temporal distribution of urban street ponding levels from surveillance videos based on computer vision

Timely and accurately estimating ponding levels during urban floods is the basis of effective disaster prevention and mitigation. Road surveillance videos record the urban flood process as images, and computer vision technology brings new opportunities for extracting ponding information from image data. This study proposes a computer vision-based method for estimating the spatial-temporal distribution of urban street ponding levels from surveillance videos. First, 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 model was adopted to identify the ponding levels whenever and wherever sedans were detected from the videos. Second, outlier detection was employed to detect and delete the outliers of ponding levels in each time step. Finally, the ponding level distribution was estimated by inverse distance weighted from the remaining ponding level points. This method was employed for two pluvial flood events at a street crossing, Dongguan Street, in Dalian, China. The mean average precision (mAP) of the trained YOLOv3 model reached 78%, which confirmed the validity of the model. The ponding levels estimated by our method were validated with the submerged depth of a static reference, and the ponding process had a strong correlation with the rainfall time series. Outlier detection improved the accuracy of ponding level estimation in cross-validation to 88% on average. The results can be used to analyze the progress of urban flood evolution, which contributes to arranging drainage facilities and improving urban flood management.