Stochastic downscaling of precipitationwith conditional mixture models

Statistical downscaling models (SDMs) seek to bridge the gap between large-scale variables simulated from General Circulation Models (GCMs) and small scale variables with high spatial variability such as precipitation. In this paper, we propose to model the distribution of precipitation conditional on large-scale atmospheric information with conditional mixture models (CMMs). CMMs are mixture models whose parameters are computed by a neural network based on large-scale atmospheric predictors. We consider three types of CMMs which differ in the type of continuous densities (Gaussian, Log-Normal or hybrid Pareto) they use as mixture components. We evaluate the three CMMs against the two-component mixture from Williams [3] at downscaling precipitation at three rain gauge stations in the French mediterranean area.