A deep learning approach for urban pluvial flood depth estimation using citizen-generated data and physics-based validation

ABSTRACT Development of calibrated urban flood models in many cities is constrained by the lack of data from actual flooding events. The use of citizen-generated data and deep learning (DL) approaches presents significant potential for achieving more accurate urban flood risk assessment in data-scarce environments. In this research, a DL convolutional neural network (CNN) utilising the Mask R-CNN model architecture was developed using Python v3.10.11 and applied to detect flooded areas and visible submersion of passenger service vehicles during flooding conditions. A total of 537 images and videos from recent extreme flooding events, collected during post-flood field surveys and from crowd-sourced social media platforms, were used to estimate average flood depths. Model validation was undertaken through comparison of Mask R-CNN simulated flood depths with corresponding values estimated using manual flooded image segmentation and a physics-based coupled 1D–2D model. The computed Nash–Sutcliffe Efficiency (NSE) Index, root mean square error (RMSE), and mean absolute error (MAE) values of 0.960 and 0.040 and 0.034 m suggested superior performance of the Mask R-CNN model. The developed approach can enable more reliable estimation of flood depths in data-scarce cities using crowd-sourced images and videos during and after occurrence of extreme flooding events.