A Hierarchical Analysis of the Impact of Methodological Decisionson Statistical Downscaling of Daily Precipitation and Air Temperatures
A new robust hierarchical framework for statistical downscaling is presented and used in a rigorous assessment of the relative importance of 1) model structure (generalized linear models, GLM v machine learning), 2) predictor type (grid cell v. indices of weather types) and 3) link functions to downscaling of air temperature and precipitation.
We show that nonlinear machine learning models (ANN) are generally more skillful than GLM and that use of predictors with a larger spatial footprint (via weather typing indices) also leads to more skillful prediction. This is particularly true for extremes, for which simple approaches typically exhibit the largest errors.
A systematic experiment across diverse climates is presented using a new model framework.