Statistical downscaling of precipitation through mixture-model clustering and nonhomogeneous transition probabilities for circulation and precipitation patterns

Downscaling is a general term to describe methods that attempt to derive local-scale values or characteristics from large-scale information such as AOGCM outputs. Hence, downscaling can be very useful to assess projected changes in climate from a local point of view, particularly for purposes such as regional climate assessments that require continuous time series for multiple scenarios and AOGCM drivers, a computational feat currently beyond the range of most dynamical downscaling or regional modeling efforts. Here, we develop an advanced statistical categorization and transition modeling method that provides accurate and rapid simulations of local-scale precipitation features at low computational cost. This statistical method is based on a stochastic weather typing approach to downscale precipitation on 37 raingauges in Illinois. Two different kinds of weather states are defined: "circulation" patterns - developed by a mixture model applied to large scale NCEP reanalysis data - and "precipitation" patterns - developed by a hierarchical ascending clustering method applied directly to the observed rainfall amounts on Illinois. By modelling the transition probabilities from one pattern to another by a nonhomogeneous Markov model - i.e. influenced by some large scale atmospheric variables such as geopotential heights, humidity and dew point te