Integrating methods of machine learning into computational models of flood risk assessment

The impacts of flooding and society’s vulnerability to it are increasing under climate change, with more than 79 million people affected, 5,500 fatalities and US$33.5 billion of damages each year, according to the EM-DAT database. In order to manage and reduce these impacts, flood risk assessments are required, which enable spatial planning of communities – e.g. to define flood zones or other mitigation measures. These assessments are often completed using computational numerical modelling of surface water flow hydraulics, in a process which integrates large volumes of data for boundary conditions (i.e. water inflows, topography, bathymetry and land cover), model accuracy assessment (e.g. maps derived from SAR images from satellite remote sensing) and data for impact quantification (e.g. building locations and types, or data of other vulnerable infrastructure). However, while these methods can provide excellent assessments, even reduced complexity numerical models are computational expensive, meaning that a challenging trade-off is required between model spatial resolution and extent, model structure, and number of simulations. In addition, full assessment of uncertainty is rarely completed, due to the high computational demands which results from a requirement to complete many 100s or 1000s of simulations. On top of this, the data used in flood risk assessments are rarely without significant gaps or errors, which add to this uncertainty and reduce the reliability of estimates. The integration of machine learning methods into flood risk assessments can, potentially, provide solutions which address some of these issues and lead to a drastic improvement in flood risk assessment. For example, it may enable improved data processing of model boundary conditions, improve the quantification of uncertainties associated with flood risk assessments, or facilitate the rapid near-real time prediction of flooded areas during an event. This PhD will assess the suitability of machine learning for flood risk assessment, and explore methods for integrating these methods into numerical models.