Résumés
Résumé
Une procédure d'évaluation régionale des risques de contamination des eaux souterraines par des pesticides a été développée et appliquée à une partie de la plaine du Rhône valaisanne. La combinaison d'une application stochastique (Monte–Carlo) de modèles déterministes simulant localement le devenir de pesticides et des techniques d'interpolation géostatistique permet d'évaluer également les incertitudes entachant les prédictions effectuées. Les divers types de modèles utilisés (solution analytique et résolution numérique de l'équation de convection-dispersion, modèle capacitif) conduisent en général à des résultats très similaires. Les cartes obtenues montrent que le risque de contamination est très élevé. Les incertitudes sont d'un ordre de grandeur similaire, i.e. ± 0.2-0.3 pour des indices de risque compris dans l'intervalle [0,1]. Ces incertitudes proviennent à raison d'environ 40-50 % des propriétés des pesticides et d'environ 30-40 % de la profondeur de la nappe phréatique, le 20 % restant étant dû aux incertitudes entachant les caractéristiques des sols, essentiellement leurs teneurs en carbone organique.
Mots-clés:
- Pesticides,
- contamination,
- incertitudes,
- Monte-Carlo,
- géostatistique
Abstract
A method to assess groundwater vulnerability to pesticide contamination on a regional scale has been developed and applied to the Rhone river valley in Valais. Uncertainties regarding vulnerability predictions were accounted for by combining local stochastic simulations, i.e. one-dimensional deterministic pesticide fate models used in a Monte–Carlo process, and interpolation by geostatistical tools. Due to the characteristics of the dataset, several preliminary operations were required including:
- a statistical analysis of soil sample data to derive a probabilistic relationship between soil classes and hydrodynamic properties,
- an estimation of the relevant water table depths at soil profile locations by cokriging the minimal and maximal water table heights measured in neighboring wells,
- and an analysis of the climate data provided by two stations located close to the study area.
In the Monte–Carlo process, soils were described by the probability density functions of particle size distribution and organic carbon content as derived for each soil class from soil sample data and by an empirical uniform distribution of dispersivity. The parameters of moisture retention curves and hydraulic conductivity versus water content relationships were derived using various pedotransfer functions. Water table depth was described by a uniform distribution within the range defined by its minimum and maximum values while empirical uniform distributions outlined the properties of three generic pesticides. The important spatial variations in climate along the valley were accounted for by linear interpolation between the data from the two stations. The three selected simulation models were an analytical solution of the convection–dispersion equation (attenuation factor), a tipping bucket model (Leach–A) and a numerical solution of the convection–dispersion equation (Leach–M). Model sensitivity analysis using Latin hypercube sampling along with multiple regression showed that pesticide properties (degradation rate, partition coefficient), organic carbon content and water table depth are the most important variables regarding cumulative (10 year) pesticide fluxes to the groundwater. This analysis also stressed the weak effect of soil hydrodynamic characteristics. Besides, the Latin hypercube sampling technique proved to be very effective in reducing the number of simulations required by the Monte–Carlo process to something manageable. The five simulated cases were:
1. all uncertainties included,
2. fixed pesticide properties,
3. fixed pesticide properties and water table depth,
4. same as 3 but with empirical distributions for soil characteristics
5. same as 3 but with constant average climate conditions.
In all cases, the fractiles of the locally simulated distributions show the same spatial pattern, i.e. all their variograms and cross variograms are proportional to the same model. Due to this intrinsic coregionalization property, interpolation may be achieved by kriging independently the various fractiles rather than by cokriging simultaneously all of them. However, integration of spatial uncertainty by sequential gaussian simulation has not been achieved due to excessive computer lasting. The resulting maps show that groundwater vulnerability is very high. Uncertainties are almost of the same order of magnitude, i.e. ± 0.2 for vulnerability indices ranging within the [0, 1] interval. Uncertainties on pesticide properties and water table depth account each for some 40 % of the resulting uncertainty while the variability in organic carbon content accounts for the remaining 20 %. All model outcomes were quite similar, except in the case of the less persistent pesticide with the attenuation factor. The dominant variables are the degradation rate and the partition coefficient of the pesticide, the organic carbon content of the soil and water table depth. Climate is of course the governing process, but 10 to 20 % variations do not significantly affect the predicted cumulative pesticide fluxes. Availability of soil sample data regarding particle size distribution as well as pedotransfer functions has little effect on the resulting vulnerability assessments.
Keywords:
- Pesticides,
- contamination,
- uncertainties,
- Monte-Carlo,
- geostatistics