Impact de la complexité du modèle hydrologique sur la régionalisation des paramètres

Abstract

Climate change increases the vulnerability of human societies and ecosystems to hydrological risk. Despite their necessity in such a context, modelling tools are limited by the decline of hydrological data collection networks and the quality of the existing data is altered by human-induced changes to watersheds. Continuous streamflow prediction at ungauged sites addresses these two issues. In this study, spatial proximity, physical similarity and multiple linear regression regionalization methods are applied to 266 snow-covered basins located in the province of Québec, Canada. The main objective of this research project is to quantify the impact of hydrological model complexity. To achieve this objective, two specific objectives are pursued 1) to quantify the impact of the distribution, within the parameter space, of the transferred parameters; and 2) to verify the statistical significance of the regional linear model – computed to apply the multiple linear regression method. The main objective is met by comparing three hydrological model configurations – with respectively, 6, 9 and 15 free parameters – of the GR4J hydrological model coupled with the CemaNeige snow model. The first specific objective is reached by comparing two optimizations algorithms – SCE-UA and CMAES – and one random sampling optimization method. The second specific objective is reached by comparing the coefficient of determination R2, to the F-statistics, taken from the F-test, and the P-value, computed from the regional linear models. Results show that the physical similarity method is slightly more efficient than the spatial proximity method whereas the multiple linear regression method is the least efficient method. These results also confirm the importance of transferring entire parameter sets. Among the tested hydrological models, the 6 free parameters version of the model had the highest success rates while the 15 free parameters version of the model showed the highest values of the NSE efficiency criteria. This study makes the distinction between robustness – quantified by success rate – and performance – quantified by NSE efficiency criteria. The 9 free parameters model appears to be the best model to regionalize parameters under the prevailing conditions of this study. The three optimization methods presented a similar performance although the random sampling optimization method was the most robust. This later method also showed a greater uncertainty in streamflow prediction. Although it is possible to compute a regional linear model with a high coefficient of determination R2 for each hydrological model free parameter – meaning a strong correlation between the basins physical descriptors and the hydrological model free parameters –, the robustness of the regional linear model is not improved. Instead, the robustness of the regional linear model is generally inversely proportional to the strength of the coefficient R2. Moreover, when the coefficient of determination R2 reaches its optimum, the F-statistics indicates that the relationship between the explained and explanatory variables of the regional linear model is not significant. In conclusion, complex rather than parsimonious hydrological models and optimization algorithms generating parameter sets confined in a restricted region of the space parameters should be preferred to apply regionalization methods. Although multiple linear regression is a widely used as a regionalization method, its predictive ability is not statistically significant.

Date	20 Feb 2015
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	François Brissette (Supervisor)