University of Bath

Process based Flood Frequency Analysis using Storm Tracking

This project aims to develop a flood frequency model utilising a new set of storm typologies to decrease engineering failures in future development projects.

This project aims to develop a flood frequency model utilising a new set of storm typologies generated through the combination of a modern particle trajectory system and a novel unsupervised classification methodology.

Project outline

Current approaches to flood frequency analysis rely on the assumptions of a stationary distribution and an approach which favours data over physical processes (Salas, Obeysekera, and Vogel 2018). With increasing amounts of high quality data (for example, the National River Flood Archive [NRFA]) and HYSPLIT (a particle trajectory system) there are new opportunities to be taken advantage of to develop the next generation of flood frequency analysis (FFA) techniques (Lamb et al. 2016). This project aims to classify flood events on their storm tracks, an example of which is given in figure 1. These classifications can then be combined to generate a new statistical model for FFA combining both physical and data-oriented approaches.

Beginning with the identification of floods this project will utilise the NRFA’s river flow archive to pick out flood events across the UK. The relevant storm tracks for these events will the be extracted using the HYSPLIT service from the National Oceanic and Atmospheric Administration (NOAA). Following this these tracks will be classified using a new neural network architecture combining deep learning with a topological preserving classification method. The flood distributions for each of these classifications will then be combined into a single model using statistical mixture techniques. Finally, this new FFA method will be compared with the traditional approaches to prove its effectiveness.


Through the combination of these new datasets a FFA method will be generated unlike any other before. As well as this a novel neural network architecture will be developed for the classification of these events resulting in advances in statistics, hydrology and computer science.


Through the provision of improved FFA estimates future development projects can be better informed in the magnitudes of events which will be present. This will result in a lower quantity of engineering failures if used correctly.

This PhD project is supervised by Dr. Thomas Kjeldsen from the Department of Civil Engineering and Architecture and Dr. Ilaria Prosdocimi from the Department of Mathematical Sciences.