Geostatistics for assessing the efficiency of a distributed physically-based water quality model: application to nitrate in the Seine River
Résumé
This article shows how geostatistics can be used to reduce distributed physically based model (DPBM) uncertainties when assessing nitrate concentrations along a 250-km stretch of the Seine River. First of all, co-kriging is used to build a set of boundary conditions (BCs) (inlet concentrations from major tributary rivers) consistent with validation data. It partially addresses the issues of errors in data used as model input and errors in data to be compared with model outputs. Then the analysis of temporal variograms (simple variograms of observed and simulated concentrations, and their cross-variogram) reveals a clear mismatch between simulated and observed values that was not detected by classical objective functions [root mean squared error (RMSE), etc.]. Variograms appear to include three components representing three different time scales: sub-weekly, monthly and annual scales. Mismatches between simulated and observed values are analysed as (i) wrong quantification of inputs to the river (especially during the rainy period: combined sewer overflow (CSO), waste water treatment plant (WWTP) and tributary contributions to the nitrate fluxes in the Seine River) and (ii) wrong description of physical processes within the river. Finally, the modelling of simple and cross-variograms appears to be a sensitive analysis tool which can be used to describe and reduce modelling uncertainties