Exploring machine learning for flood impact forecasting at the FFC

The Flood Forecasting Centre (FFC) conducted a study to investigate the application of machine learning (ML) techniques for predicting coastal and surface water flooding impacts in England and Wales. The primary objective was to understand whether FFC's manually-labelled flood impact database in combination with hydrometeorological datasets, could provide skilful ML-based flood impact predictions.

For coastal models, a key finding saw that incorporating tide gauge and wave buoy data significantly improved performance. For both coastal and surface water models, results show classification accuracy ranged from 70% to 90% across different regions. Results show regional differences, for example in the coastal model the South West region demonstrated the highest predictive accuracy, perhaps due to the higher number of observed impacts in the training set. Conversely regions with fewer observed events, like the North East, showed lower performance. The investigation highlights the potential of using ML approaches for forecasting coastal and surface water flooding events, though challenges with training data quality and class imbalance are likely to be important factors for further refinement of predictive skill in this area.

Future work will focus on improving data handling and model refinement. Further exploration is needed to include more feature datasets along with using other impact datasets as targets or using proxies for impacts. We would also like to explore operational implementation of similar ML approaches in forecast mode, using ensemble forecast data as input. This has the ambition that operationalised ML tools can enhance flood risk forecasting and support the protection of lives and livelihoods from flooding.