Future hydrological scenarios are usually simulated by forcing an impact model with high-resolution meteorological scenarios, get from statistical downscaling models (SDMs). These SDMs are expected to fill the gap between the poor resolution and the bias of General Circulation Models (GCMs) scenarios and the requirements of impact models. Numerous SDMs are currently developed worldwide. Their validation usually focus on their ability to reproduce the main statistical characteristics of the observed climatology (e.g. probability distribution functions and seasonality of reconstructed variables) when they are forced by large scale fields of meteorological reanalyses. This is a necessary but insufficient condition for their temporal transferability, particularly in a changing climate context. Then, future hydrological projections usually account for the emission scenarios and GCMs uncertainties. However, they often ignore the downscaling related uncertainty. We present here an evaluation methodology to illustrate the possibilities and/or the difficulties to transfer in time these algorithms. We next illustrate the uncertainties in future meteorological and hydrological projections that can result from this imperfect transferability.

For evaluating the time transferability of SDMs, we consider the ability of the models to simulate the low-frequency variations of paste climate as a sort of “natural” climate change. For this, the chronology of generated scenarios, aggregated on several months or years, should explain a large part of the observed chronology and especially of observed trends if any. Similarly, the capacity of the model to give similar performance for a validation period with quite different atmospheric conditions than that observed for the calibration one should be tested. We use therefore an evaluation framework similar to that usually applied by hydrologists. A poor time transferability could suggest a poor model performance. It could also result from time heterogeneities in both large scale fields and local meteorological data. As data heterogeneities would however lead to similar results for different SDMs, a multi-model evaluation can be useful to isolate the respective effects of model and data. The same may be achieved with different parametrizations of a same SDM (e.g. different large scale predictors).

The ability of SDMs to generate relevant correlations between the meteorological variables required as input of the behavioural model of the studied hydrosystem is a last important feature to test. The non-linearity between precipitation and discharge can emphazise or lessen some weaknesses of SDMs. A hydrological evaluation of the SDMs, where reconstructed and observed hydrological series are compared at different aggregation time steps, should be additionally required.

The proposed methodology is more severe than classical climatological validations and allows better highlighting the strengths and weaknesses of the considered SDMs. It has been applied on a mesoscale alpine basin in Southern French Alps (the Upper Durance basin, 3580 km2) with three different SDMs : ”dsclim”, based on a weather regimes classification and regional precipitation indices (Boé et al. 2006), ”analog”, a k-nearest neighbours resampling model for selecting an analog day from large scale predictors fields (Obled et al, 2002) ”ddwgen” based on generalized linear models for simulating mean areal precipitation and temperature from atmospheric circulation indices (Mezghani et Hingray, 2009). For “dsclim” and “ddwgen”, several sets of large scale predictors are considered. Meteorological reconstructed scenarios are used as input of the physical-based hydrological model ISBA-Durance (Lafaysse et al, 2011) to simulate discharges.

The same SDMs and hydrological model are then used to simulate a large number of future hydrological projections, based on the ENSEMBLES european project GCMs. This allows for comparing the GCMs and the downscaling related uncertainties by a classical variance analysis. The dispersion in hydrological regime changes due to the choice of both downscaling model and large scale predictors is much higher than expected with our previous transferability evaluation. This suggests that the reasons for changes at climatological temporal scale can significantly differ for processes involved in interannual variability during the last 50 year. Progress in comprehension of physical processes involved in climatological changes at the catchment scale is therefore necessary to improve the SDMs robustness. Nevertheless, in this particular case of a high mountainous catchment, significant river flow changes due to the snow-rain ratio decrease are simulated for all GCMs/SDMs configurations despite the high uncertainties in total precipitation changes.

This paper presents final results from a PhD research funded by the French National Weather Service (Météo-France). It is linked to the research project RIWER2030 "Regional ClImate, Water, Energy Resources and uncertainties from 1960 to 2030" (http://www.lthe.fr/RIWER2030/) funded by the French National Research Agency (ANR).

References

Boé, J., Terray, L., Habets, F. and Martin, E. (2006). A simple statistical downscaling scheme based on weather types and conditional resampling. J. Geophys. Res., 111.

Lafaysse, M., Hingray, B., Etchevers, P., Martin, E. and Obled, C. (2011). Influence of spatial discretization, underground water storage and glacier melt on a physically-based hydrological model of the Upper Durance River basin. J. Hydrol., 403(1-2):116–129.

Mezghani, A. and Hingray, B. (2009). A combined downscaling-disaggregation weather generator for stochastic generation of multisite hourly weather variables over complex terrain: Development and multi-scale validation for the Upper Rhone River basin. J. Hydrol., 377(3- 4):245–260.

Obled, C., Bontron, G. and Garcon, R. (2002). Quantitative precipitation forecasts : a statistical adaptation of model outputs through an analogues sorting approach. Atmos. Res., 63(3-4):303–324.