Flood Modelling in Large Catchments using Open-Source Data and Data-Driven Techniques

Abstract

Flooding has severe and devastating consequences across the globe; climate change means that the frequency and severity of these events will increase. Advances in flood modelling and prediction methods and developments in open source data and computing capabilities mean that flood modelling methods should be more accessible. However, the diversity in modelling approaches and available data sources makes it hard to determine the approach most suited to a study area. This thesis uses satellite data products and data-driven techniques to explore methods suited for large catchments with limited available data. Three global DEMs at different resolutions are used to create one dimensional hydrodynamic river network models, calibrated using streamflow derived from a satellite gauge estimate to determine whether, at a large-scale, accuracy can be improved using a more recent or higher resolution Digital Elevation Model (DEM). The application of Artificial Neural Networks (ANNs) has also been explored, combined with satellite precipitation data for training to predict streamflow. An ensemble of hybrid Genetic Algorithm Neural Networks (GANN) was also applied to streamflow prediction, trained on rainfall data from gauges located throughout the upper catchment area. Three input variable selection methods (IVSM) were evaluated to determine the influence of training data selection. The main findings were that DEMs of higher vertical accuracy and horizontal resolution does not significantly improve large-scale models' accuracy. For large complex catchments, data-driven methods such as neural networks can be used where a physically-based hydrological rainfall-runoff model would be too computationally expensive and require extensive calibration. Training ANN models on satellite precipitation data proved more effective than observed rain gauge data, and the ANN outperformed an ensemble of GANNs. Training data selection was found to significantly impact the models' accuracy, and consequently, a model-based selection method was proposed.