A robust framework for probabilistic precipitations downscaling from an ensemble of climate predictions applied to Switzerland

Abstract

[1] Rainfall is poorly modeled by general circulation models (GCMs) and requires appropriate downscaling for local-scale hydrological impact studies. Such downscaling methods should be robust and accurate (to handle, e.g., extreme events and uncertainties), but the noncontinuous and highly nonlinear nature of rainfall makes this task particularly challenging. This paper brings together and extends state-of-the-art methods into an integrated and robust probabilistic methodology to downscale local daily rainfall series from an ensemble of climate simulations. The downscaling is based on generalized linear models (GLMs) that relate monthly GCM-scale atmospheric variables to local-scale daily rainfall series. A cross-validation step ensures that the fitted models are correctly conditioned by the climate variables, and a statistical procedure is proposed to test whether the statistical relationships identified for the reference period also hold in a future perturbed climate (i.e., to test the stationarity assumption). Additionally, we propose a strategy to downweigh poorly performing GCM-GLM couples. The methodology is assessed at 27 locations covering Switzerland and is shown to perform well in reproducing historical rainfall statistics including extremes and interannual variability. Furthermore, the projections are consistent with the simulations of physically based dynamical models. Using an original visualization method based on heat maps, we show that although the downscaling models were fitted at each of the 27 sites independently, their projections follow a spatially coherent pattern and that regions exhibiting different climate change impacts can be identified.Citation: Beuchat, X., B. Schaefli, M. Soutter, and A. Mermoud (2012), A robust framework for probabilistic precipitations downscaling from an ensemble of climate predictions applied to Switzerland,

Transferability in the future climate of a statistical downscaling method for precipitation in France

“…As it is impossible to evaluate the capacity of the SD method to capture the future climate change signal based on observations, some authors have tried to evaluate the capacity of their method to reproduce future climate change using climate model results as pseudo‐observations [ Vrac and Stein , ; Frias and Zorita , ; Beuchat et al , ]. In this framework, the statistical downscaling method is built using present‐day model results instead of real observations.…”

Statistical Downscaling of General Circulation Model Outputs to Precipitation Accounting for Non-Stationarities in Predictor-Predictand Relationships

“…Statistical downscaling is based on the concept of developing empirical statistical relationships between the GCM outputs and the catchment scale hydroclimatic variables [7]. Statistical downscaling depends on the assumption that the relationships between the GCM outputs (predictors—inputs to downscaling models) and the observations of the catchment scale hydroclimatic variables (predictands—outputs of downscaling models) in the past are valid for the future under changing climate [8]. This assumption is called the stationarity assumption of the predictor-predictand relationships (PPRs).…”

A comparison of three predictor selection methods for statistical downscaling

“…The basic principle of statistical downscaling is to identify statistical relationships between an observed small‐scale predictand variable and larger‐scale predictor variables for a baseline period (Beuchat et al , ), and then applied these relationships to downscale future climate scenarios using global climate model (GCM) output predictors. In the past two decades, many studies have proposed statistical downscaling methods based on various algorithms including automated statistical downscaling (ASD), artificial neural network (ANN), stochastic weather generator (LARS‐WG), nonhomogeneous hidden Markov model (NHMM), statistical downscaling model (SDSM), support vector machine (SVM) and so on (Bates et al , ; Semenov et al , ; Wilby et al , ; Coulibaly et al , ; Hessami et al , ; Chen et al , ), some research has also compared the performance of different downscaling methods in capturing the characteristics of local‐scale climate change by means of uncertainty analyses, correlation analyses, distribution functions and other statistic methods and found that the algorithms of SDSM has an important influence on the downscaled results (Khan et al , ; Tryhorn and DeGaetano, ; Liu et al , ; Gutmann et al , ).…”