Verifying the time-invariance assumption of statistically downscaled precipitation datasets for future climate

Given the coarse resolution of the global climate models (GCMs), downscaling techniques are often needed to generate finer scale projections of variables affected by local scale processes such as precipitation. However, classical statistical downscaling experiments for future climate rely on the time-invariance assumption as one cannot know the true change in the variable of interest, or validate the models with data not yet observed.

Our experimental setup involves using the Canadian regional climate model (CRCM) outputs as pseudo-observations to estimate model performance in the context of future climate projections by replacing historical and future observations with model simulations from the CRCM, nested within the domain of the Canadian GCM (CGCM). In particular, we validated statistically downscaled daily precipitation time series in terms of the Pierce skill score, mean absolute errors, and climate indices. Specifically, we used a variety of linear and nonlinear methods such as Bayesian neural networks, support vector regression, regression trees, multiple linear regression, and k-nearest neighbours to generate present and future daily precipitation occurrences and amounts. We obtained the predictors from the CGCM 3.1 20C3M simulation (1970-1999) and the SRES A2 simulation (2040-2069), and precipitation outputs from the CRCM 4.2 forced with the CGCM 3.1 boundary conditions as predictands. In general, ensembles of Bayesian neural network models or regression trees outscored the linear models and the simple nonlinear models in terms of precipitation occurrences, while the models exhibit a heterogeneous performance in terms of precipitation amounts.