Identification of River Defences from Digital Terrain Models using Deep Learning

Abstract

Flood defences play a central role in the quantification of flood risk. JBA Risk Management produces undefended hazard maps that are supplemented with defence information to provide risk practitioners with the most flexible view of risk. This requires knowledge of the locations of river defences so that they can be removed from the digital terrain models prior to flood modelling. We report on work to develop a predictive model for identifying river defences. This model was created using the U-Net deep neural network for image segmentation. The model was developed over a series of iterations, where the prediction outputs were refined and used to retrain the model. We have used this model to produce national maps of defences for a range of countries