Machine Learning Models of Flood Events on Multiple Topographies

This thesis investigates the potential of artificial neural networks to perform flood modelling on multiple domains, in order to address the issue of computational inefficiency of physically-based flood models. Using a synthetic training database, convolutional neural networks were trained on (1) modelling the spatially-distributed maximum water depth resulting from a flood event, and (2) modelling the spatio-temporal water depth during a flood event. Accurate maps of maximum water depth were produced on a variety of terrains without need for retraining. Spatio-temporal modelling of the water depth could not be achieved with sufficient accuracy. A number of methodological limitations were identified.