IAP-25-108

Probabilistic flood prediction and modelling using machine learning

River flooding is a major natural hazard, with substantial human and economic impacts that are expected to increase due to changes in climate. Predictive modelling offers a way to mitigate these impact – long-term flood management strategies can be explored, tested and refined; while real-time forecasting allows resources to be deployed as effectively as possible during extreme events. However, the hydrologic response of a catchment is complex, with many interacting processes, and available data may not constrain all relevant aspects of the system. This creates significant uncertainties, and it is important to capture these within any predictions. This project will develop statistical machine learning models that enable probabilistic hydrological predications, and investigate how these can be used to reduce flood risks.

This project will explore a variety of machine learning methods that can be used to capture and model uncertainty – such as mixture density networks, variational autoencoders, and normalising flows. These will be trained using a combination of observed and simulated data, and tested against historical records obtained from partners at the Centre for Ecology & Hydrology and elsewhere. Work will initially focus on specific catchments for which substantial quantities of high-quality data are available; once this ‘proof-of-concept’ is established, we will explore how this can be applied more widely.

Year 1

• Initial familiarisation with machine learning models and/or hydrological theory & observations;
• Exploration/comparison of candidate modelling approaches and identification of approach to be employed;
• Compilation and exploration of available data.
• 1-week visit to CEH partners
• Attend a national conference e.g. British Hydrological Society symposium
• Attend a summer school e.g. University of Birmingham Catchment Science Summer School

Year 2

• Proof-of-concept study for one or more well-studied catchments: How well can the method perform in a known setting? Which environmental and hydrological parameters are most important for ensuring high-quality flood predictions?
• Initial development of strategy for using and communicating probabilistic results: How should this work be presented in order to maximise utiility for a range of audiences.
• 1-week visit to CEH partners
• Attend a national conference e.g. British Hydrological Society symposium

Year 3

• Scaling to less-known catchments: how should we handle missing or incomplete data? What are the limits on predictability?
• 1-week visit to CEH partners
• Attend an international conference e.g. European Geosciences Union

Year 3.5

• Finalisation of research and preparation of thesis

The student will receive training in relevant fields such as hydrology, machine learning, and statistics. Specific skills such as computer programming (e.g. in Python) and the use of standard ML libraries (e.g. Pytorch) will also be acquired as needed. A mix of formal and informal training will be provided. The student will attend a summer school such as the University of Birmingham Catchment Science Summer School in Year 1, to provide a solid grounding in domain-specific knowledge.