Sous-échantillonnage d'un ensemble de projections climatiques régionalisées | Sub-sampling ensembles of downscaled climate projections

Increasingly large ensembles of GCM simulations are becoming available for climate change impact studies, through multi-model and multi-run ensembles or perturbed physics ensembles. Moreover, advances in statistical downscaling make it now possible to downscale such large ensembles to the spatial resolution relevant for impact models. Additionally, many downscaling methods are themselves stochastic, which may again increase the size of downscaled projections ensembles. The computational cost of processing such large ensembles through impact models like physically-based distributed hydrological models may however be prohibitive. If recommandations for selecting GCM simulations from large ensembles have been recently proposed, there are still numerous open questions on how to adequately sample downscaled ensembles of such simulations. This work proposes a sub-sampling approach undertaken for providing a set of downscaled projections over the Durance catchment (southern French Alps) for building informed adaptation scenarios in water resource management. 30 transient runs from the ENSEMBLES Stream2 GCMs under the A1B emissions scenario have been downscaled over the Durance catchment by three variants of the K-nearest neighbours resampling approach: an analog method, a weather type method and a regression-based method (Lafaysse et al., 2013). 100 downscaled realizations have been stochastically generated by each method for all GCM runs (1 to 6 runs from 4 different GCMs). The approach selected here aims at preserving the relative contributions of the four different sources of uncertainties considered, namely (1) GCM structure, (2) large-scale natural variability, (3) structure of the downscaling method, and (4) catchment-scale natural variability. Given the relatively low sample size of the first three sources, this approach focused on sub-sampling 10 realizations of each downscaling method by applying a conditioned Latin Hypercube Sampling (see e.g., Christierson et al., 2012) and therefore preserving the statistical distribution of the 100 realizations. Many open choices---nature of the conditioning variables (future values/changes), associated temporal and spatial scales---have been carefully made by assessing their relevance for water resource management. The effect of conditioning the sampling on climate responses in temporally and spatially integrated variables---changes in catchment-scale winter/summer and interannual mean precipitation and temperature between two time slices---has been validated by assessing the response in more extreme independent variables like changes in the annual precipitation with an exceedance probability of 0.8 and in maximum consecutive dry days. A transient analysis of variance moreover confirmed the effectiveness of the approach in preserving the relative contribution of uncertainty sources for various climate variables. Critically, this approach allows to propagate the sources of uncertainty through impact models while reducing the associated computational burden. However, in order to meet actual constraints of the impact community, there is an urgent need for producing guidelines for sub-sampling multi-level ensembles of downscaled climate projections, i.e. 3D arrays of combinations in emissions scenarios, GCMs and downscaling methods.