Spatially distributed calibration of a hydrological model with variational optimization constrained by physiographic maps for flash flood forecasting in France

The estimation of storage and fluxes in surface hydrology is an essential scientific question related to major socio-economic issues, especially when forecasting extreme floods and droughts with the undergoing climate change. Advanced spatially distributed modeling tools are critically needed to perform reliable and skillful local forecasts. Nevertheless, hydrological modeling remains a challenging task because of limited observations of physical processes and modeling uncertainties. In particular, given the spatial sparsity of constraining discharge data, hydrological modeling is faced with the challenge of producing predictions at ungauged locations based on the regionalization of the model parameters. Despite the overparameterization problem in spatially distributed modeling, Jay-Allemand et al. (2020) presented promising results for estimating the spatial variability of the distributed parameters within a catchment using only downstream discharge observations. However, providing better spatial constrains on the estimated parameters patterns, inside or outside calibration catchments in a regionalization perspective, remains a challenge.

This contribution presents a regionalization approach based on: (i) the SMASH (Spatially distributed Modelling and ASsimilation for Hydrology) hydrological modeling and assimilation platform (Haruna et al., 2021) underlying the French national flash flood forecasting system Vigicrues Flash (Javelle et al., 2019); (ii) the variational assimilation algorithm from Jay-Allemand et al. (2020), adapted to high dimensional inverse problems; (iii) spatial constraints added to the optimization problem, based on masks derived from physiographic maps (e.g., soil occupation and nature, bedrock type, terrain slope); (iv) multi-objective optimization which targets independent watersheds. This method gives a regional estimation of the distributed parameters over the modeled area. Performances of the model and the parameters robustness are evaluated on a large sample of French catchments and flash floods in spatio-temporal extrapolation based on cross-validation experiments. Effects of the spatial constraints (regularization and multi-objective optimization) are discussed in the light of adjoint sensitivity maps. Further work aims to improve the global search of prior parameter sets and to better balance the adjoint sensitivity with respect to the spatial constraints resolution and catchment characteristics. This will ensure a better consistency of simulated fluxes variabilities and enhance the applicability of the regionalization method at higher spatial scales.