Sensitivity analysis for a flood and inundation mapping forecasting modelling chain
Résumé
The development of a hydrometeorological flood forecasting chain adapted to quantify predictive uncertainty is
challenged by the very short time available to run its models, evaluate flood probabilities, alert the population
at risk and manages emergency responses. Forecasting systems often integrate several models in a chain of
input-output simulations that altogether provide the hydrological information needed for decision-making. Each
model is a component of the system that needs to be understood and evaluated in terms of its contribution to
enhance the accuracy of the final prediction. In this study, we investigate a flash flood forecasting system based
on two existing simulation models: the hydrological model GRSD (de Lavenne et al., 2016), which provides
streamflow time series at gauged and ungauged catchments, and the MHYST model (Rebolho et al., 2018), which
is a flood inundation model that aims to map inundation extents at the river reach scale. Each model has its own
set of parameters that might influence differently the results of the simulations, modifying this way the overall
performance of the forecasting system. Here, we use the Sobol sensitivity analysis to investigate the influence of
the individual model parameters but also the interactions sensitivities of the parameters within each model and
when considering the modelling chain that integrates both models. We evaluate sensitivity with regard to flood
peaks, flood peak timing and flood inundation extent. Our case study is the Loing river, a tributary of the Seine
river in France, which was severely affected by floods in May-June 2016. Results highlight sensitivity rankings for
both models and how the real-time information acquired can be useful to define calibration and data assimilation
strategies based on both point and spatial data during the forecasting.