Résumé
The accuracy of gauge-based, satellite-based and reanalysis precipitation products for streamflow simulation has rarely been investigated in data-scarce and meso-scale karst catchments characterized by infra-daily response time due to the predominance of quick flow processes. This study evaluates and compares the reliability of gauge- and satellite-based precipitation products (CPC, E-OBS, PERSIANN-CDR, IMERG-LR, SM2RAIN-ASCAT, CHIRPS) and reanalysis products (SAFRAN, COMEPHORE, ERA5-Land) in simulating the daily flow of the Baget karst catchment (13.25 km2), located in the Southwestern French Pyrenees. The assessment was conducted over the 2006-2018 period using the semi-distributed karst hydrogeological model ISPEEKH, integrated with a PEST framework for model calibration, global sensitivity analysis using the Morris method, and parameter estimation using an iterative ensemble smoother form of the Gauss-Levenberg-Marquardt algorithm. The discharge coefficients and emptying exponents of the epikarst-to-conduit and conduit-to-spring quick flows were the most sensitive model parameters irrespective of the input precipitation, and ISPEEKH successfully reproduced the non-linear conduit flow dynamics in the catchment. Yet, simulated streamflow was significantly underestimated under the ensemble of precipitation products (up to 32-79 % in the calibration period and up to 28-70 % in the validation period), and the reanalysis products outperformed the gauge- and satellite-based products. Downscaling of the CPC, IMERG-LR, ERA5-Land and E-OBS products, and merging of the CPC and IMERG-LR datasets at 1-km spatial resolution did not improve the model predictive performance. Finally, the study showed that watershed-scale precipitation correction can effectively improve the hydrological simulation performance in the catchment, particularly under the French reanalysis precipitation product COMEPHORE. This result emphasizes the need to install representative rain gauge stations at different altitudes in studied karst catchments of similar scale and hydrodynamics characteristics, and apply observation-based correction methods in order to reduce the errors in regional reanalysis precipitation database and optimize the karst discharge simulation.