Data assimilation of SAR derived flood extent maps into flood forecasting models via particle filters

Abstract

Every year, floods cause global socio-economic losses of more than US$100 billion with an increasing trend due to growing economic activities in flood-prone zones and changes in the climate patterns. Flood warning can assist in reducing these losses based on flood forecasting models that simulate the physical rainfall-runoff-inundation processes in river basins. These models have to be as accurate and reliable as possible. However, the equations, the model parameters, the boundary conditions and the inputs are all affected by inherent uncertainty. Data assimilation (DA) techniques integrate ground based or satellite observations into the model in order to reduce the uncertainty associated with both model results and measurements. In this PhD study, flood extent maps derived from Synthetic Aperture Radar (SAR) observations are assimilated into flood forecasting models by using newly developed filtering techniques based on variants of the Particle Filter.The thesis is organised into three parts. In the first, as a proof of concept of a DA framework previously introduced in the scientific literature, we evaluate a Sequential Importance Sampling (SIS) variant of a Particle Filter in a controlled environment where synthetically generated rainfall and SAR observations represent the only sources of uncertainty.We show that DA indeed improves the model performances in terms of waterlevel, discharge and flood extent predictions. However, the state-of-the-art filter used in part 1 shows a tendency for degeneracy issues, as the number of particles with a significant weight reduces after the assimilation.For that reason, in the second part of the thesis, we enhance the existing SIS to mitigate the degeneracy and sample impoverishment issues. We develop a DA framework based on a Tempered Particle Filter (TPF) and evaluate it in a synthetic twin experiment,where rainfall and SAR observations are known to be the only sources of uncertainties.The analysis finds that the model results are improved and the degeneracy issue is mitigated.In the third part of the PhD study, the TPF is applied in a real-world experiment,where the uncertainties are no longer controlled. Three flood events of the River Severn in the UK are used as test cases. The filter performances differ depending on the gauging stations considered as a reference for the evaluation. For the gauging stations located downstream of the confluence of the three main rivers (i.e., Theme, Avon and Severn) close to the downstream boundary of the hydraulic model, the model results in terms of water level simulations are substantially improved. However, the improvements are less significant in terms of inflows at the upstream boundaries of the model, due to compensation effects between the contributing tributaries.Overall, the study is considered an important step towards an enhanced Particle Filter by reducing degeneracy and sample impoverishment and improving predictions for the downstream gauging stations. The comparison of the real-world test case with the synthetic one shows that the assumption of the rainfall and SAR observations as the only sources on uncertainty could be too simple, given the complexity of the hydrological processes involved, and it is suggested to take additional sources of uncertainty into account in future studies.