Machine learning approaches to parameter calibration and uncertainty propagation for seasonal flood risk prediction

The increase in extreme weather associated with acute climate change is leading to more frequent and severe flood events.  In the window of months and years, climate change adaption is critical to mitigate risk on socio-economic systems. Mathematical and computational models have become widely used tools to quantify the impact of catastrophic flooding and to predict future flood risks. For decision makers to plan ahead and to select informed policies and interventions, it is vital that the uncertainties of these models are well estimated. Besides the inherent uncertainty of the mathematical model, uncertainties arise from parameter calibration and the driving observational climate data.

Here we focus on the uncertainty of seasonal flood risk prediction for which we treat uncertainty propagation as a two step process. Firstly through calibration of model parameter distributions based on observational data. In order to propagate parameter uncertainties, the posed calibration framework is required to infer model parameter posterior distributions, as opposed to a single best-fit estimate. While secondly uncertainty is propagated by the seasonal weather forecasts driving the flood risk prediction models, such model drivers have their own inherent uncertainty as predictions. Through handling both sources of uncertainty and its propagation we investigate the impacts of combined uncertainty quantification methods for flooding predictions. The first step focussing on the flooding models own characterisation of uncertainty and the second characterising how uncertain model drivers impact our future predictions.

In order to achieve the above features of a calibration framework for flood models we leverage concepts from machine learning. At the core we assume a minimisation of a loss function by the methods based on the supervised learning task in order to achieve calibration of the flood model. Uncertainty quantification is equally a growing field in machine learning or AI with regards the interpretability of parametric models. For this purpose we have adopted a Bayesian framework which contains natural descriptions of model expectation and variance. Through combining uncertainty quantification with the steps of supervised learning for parameter calibrations we propose a novel approach for seasonal flood risk prediction.