Successful flood response and evacuation require timely access to reliable flood depth information in urban areas. However, existing flood-depth-mapping tools do not provide real-time flood depth information in residential areas. In this paper, a deep convolutional neural network is used to determine flood depth through the analysis of crowdsourced images of submerged stop signs. Model performance in pole length estimation is tested on a test set, achieving a root mean squared error of 10.200 in. (1 in. = 1 inch = 2.54 cm) on pre-flood photographs and 6.156 in. on post-flood photographs and an average processing time of 0.05 s. The performance of the developed model is tested on two case studies: Hurricane Ian in the USA (2022) and the Pacific Northwest floods in the USA and Canada (2021), yielding mean absolute errors (MAEs) of 4.375 and 6.978 in., respectively. The overall MAE for both floods is achieved as 5.807 in., which is on par with those from previous studies. Additionally, detected flood depths are compared with readings reported by the nearest flood gauge on the same date. The outcome of this study demonstrates the applicability of this approach to low-cost, accurate, scalable and real-time flood risk mapping in most geographical locations, particularly in places where flood gauge reading is not feasible.
