Machine Learning Approaches to Assessing Future Flood Risk

Lead Research Organisation: University of Cambridge
Department Name: Engineering


In the face of impending climate change, the need to understand the impact of extreme weather events is critical; whilst
large climate models provide broad detail, such as large scale patterns in surface temperature, they can be unsuitable for
understanding regional and localised impact and for predicting the potential impact of extreme events.
A range of machine learning and artificial intelligence approaches will be implemented, such as Gaussian Processes and
Neural Networks, as part of an investigation into whether or not credible predictions for potential future storm and flood
scenarios can be generated through these methodologies. The main predictive outputs that will be examined are total
precipitation, surface runoff, and river flow. In using these approaches, the potential development of novel aspects of
machine learning architecture, tailored to the problems of climate science, extreme weather, and hydrology, will be a
supplementary output.
Climate science techniques to be implemented alongside the machine learning approaches include bias correction,
statistical downscaling, and statistical weather generators for precipitation. Through these supplementary methodologies,
there is potential for novel mathematical functions and expressions in the field of climate science that may enable the
better prediction of extreme weather events, such as the introduction of a spatiotemporal model adjustment.
Finally, key to this project is the use of probabilistic and Bayesian techniques, in order to capture uncertainty around the
predictions, in order to enable policy and infrastructure decision makers to utilise the output of these machine learning
models more effectively.
At a more granular level, the objectives may thus be summarised as follows:
- Investigation of hydrological models and the development of machine learning analogues to capture the relationship
between climatic variables and hydrological response, critically river flow, to generate time histories for the duration of
a climate model run.
- Identification of key weaknesses with these machine learning models and investigation into the aspects of their
respective architectures with a view to tailoring components to the problem, such as the development of a custom
kernel function in a Gaussian Process.
- Development of the machine learning models with a view to determining whether or not they can generalise to
catchments upon which the models have not been trained, specifically to be used in data poor regions that may be at
more risk of extreme weather events.
- Correction of relevant climatic variables, employing further machine learning tools or developing a machine learning
pipeline, to cover bias correction and spatiotemporal downscaling, primarily to address precipitation prediction
- Determination of specific climatic drivers present within climate models to enable the prediction of extreme weather
events that may occur unseasonally or unexpectedly where precipitation projections from climate models may be


10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/R512461/1 01/10/2017 30/09/2022
2497035 Studentship EP/R512461/1 01/10/2017 30/09/2021 Robert Edwin Rouse