Urban flood monitoring and early warning by surveillance camera networks

Under the dual effects of global climate change and urbanization, urban flooding has become an increasingly severe development problem. Widespread surveillance cameras provide new opportunities to observe geographic phenomena such as extreme rainfall, flooding, and hurricanes. In this study, an urban flooding observation system is constructed using the surveillance camera network to alleviate traditional flooding sensing strategies' spatial and temporal resolution shortcomings. First, based on analyzing the visual features of flooding water in surveillance videos, a deep learning model is built to extract the flooded pixels from the surveillance video with the ResNet that incorporates an Attention mechanism as the backbone network. Second, a projection model, which maps the relationship between image space and geographic space, is employed to calculate the flooding range in the real world. After that, taking the geographical information of the surveillance scenarios into account, an evaluation index system is constructed to assess the flooding risk. For the training and testing of the deep learning models, an urban flooding visual dataset was produced. Extensive experimental results have demonstrated the effectiveness of the proposed method. Our approach deploys on the existing city surveillance resources and therefore has the advantage of low cost and high resolution.