Evaluation of operational tools for rainfall, discharges and flood forecasting in coastal catchment of south-eastern France

A wide range of methods and tools are available to anticipate precipitation, rainfall-runoff and associated flooding. The aim of this thesis is to identify whether the tools tested in this thesis are precise and robust enough for use in flood risk anticipation and crisis management by local operational entities such as the SMIAGE (co-financier of this thesis with the Sud region). The studies carried out during this thesis focused on three main subjects: precipitation forecasting, rainfall-runoff forecasting and flood modelling. For each subject, methods from the scientific literature and operational models were implemented and analyzed, focusing on related phenomenon. The study of precipitation focused on (i) an analysis of the spatial and temporal characteristics of heavy rainfall events that have affected southeastern France in recent years, based on COMEPHORE precipitation data (Tabary et al., 2012), (ii) the building of a weather patterns classification indicating the average atmospheric circulations associated with days showing similar rainfall spatial variability and (iii) the evaluation of the RainNet artificial neural network (Ayzel et al., 2020) for the heavy rainfall nowcasting. The results of the 158 heavy precipitation events selected characteristics analysis showed that all heavy precipitation events that lasts more than 30 hours and all heavy precipitation events with a cumulative basin rainfall of more than 150 mm were associated with at least one high-flow event in a gauged basin in south-eastern France during the 2007-2020 period. In addition, a classification into 8 weather patterns was constructed from 8 rainfall classes with contrasting rainfall locations. Finally, unfortunately, RainNet's precipitation forecasts were not satisfactory, the model did not seem useful for operational use.The river discharges study focused on (i) comparing simulations from three rainfall-runoff models GRP (Tangara, 2005; Berthet, 2010), GR5H (Le Moine, 2008) and GR5H-RI (Astagneau, 2022) without and with CemaNeige (Valery, 2010; Riboust, 2018) on 39 gauged outlets, (ii) the evaluation of simulations from a semi-distributed GR5H model that takes into account water uses for the modelling of Siagne river discharges at Pégomas and (iii) the evaluation of GRP and GR5H-RI models for flood forecasting with data assimilation on the 31 gauged outlets not influenced by snow. The results of the hydrological modeling tools evaluation showed that, overall, the GR5H-RI model was the most precise in simulating and forecasting floods in the basins of south-eastern France, with a 24-hour horizon calibration. This is particularly the case for heavy rainfall events with a short rising time. In the specific case of discharges modelling on the Siagne at Pégomas, the results showed that the GR5H semi-distributed model, which takes into account the influence of the dam, provides a better discharges modelling than the GR5H global model.Finally, the flood study examined (i) the sensitivity of a 1D-2D HEC-RAS hydraulic model of the downstream part of the Brague river, to parameterization and runoff modelling method and (ii) a method combining hydrological modeling, frequency analysis and hydraulic modeling to construct flood extents maps for dscharges of different return periods. Finally, the results of the analysis of flood simulations downstream of the Brague showed that water level simulations are strongly influenced by injected river discharges, hydraulic structures representation and runoff modelling method. Finally, an atlas of flooding extents was built from water levels simulated from HEC-RAS 1D-2D model of the Brague and flood hydrographs from the October 2015 flood, adjusted so that the peak flow reaches the 10-, 20-, 30-, 50- and 100-year return period discharges on the Brague and its tributaries. Peak discharges were calculated by fitting a Gumbel law to discharges simulated by a semi-distributed GR5H hydrological model.