Improving flood hazard mapping using global flood model and machine learning techniques

Abstract

Flooding is one of the most frequent and dangerous natural disasters worldwide. Global flood models are now considered as a practical tool for flood hazard assessment, with modelling results validated against benchmark flood hazard maps from regional flood models in data-rich regions. To date, most of the Earth’s land surface remains ungauged or covered by sparse observation networks. Consequently, accurate flood hazard assessment in data-sparse regions remains a challenging topic globally. Recently, machine learning techniques have been widely used to support physically based hydrological modelling in data-sparse regions, but very few studies look at improving flood hazard mapping coupled with machine learning and global flood models. This thesis brings together a global flood model (the FATHOM model) and machine learning techniques to produce improved flood hazard maps considering three data-sparse scenarios. First, I built a design flood dataset for global river networks using a hybrid machine learning framework. The derived design flood estimates can provide better discharge boundary conditions for driving global flood models, and this is likely to yield a significant improvement in flood hazard mapping for regions without discharge observations. Second, I considered the impact of dams on extreme floods and large-scale inundation simulation by coupling machine learning techniques and the FATHOM model. I investigated the flood attenuation effect for dams in ungauged regions over the conterminous United States and successfully tested the reliability of the improved FATHOM model for simulating flooding downstream of dams in two case studies. Finally, I estimated the flood defence standards for ungauged levees in the conterminous United States and England using machine learning and publicly available datasets. I then incorporated the estimated flood defence standards in flood hazard mapping of the FATHOM model and demonstrated the model results in three case studies by comparing with official flood hazard maps.

Date of Award	21 Jun 2022
Original language	English
Awarding Institution	University of Bristol
Supervisor	Paul D Bates (Supervisor) & Jeff Neal (Supervisor)