Effects of Statistical Downscaling on Extreme Value Statistics

Statistical downscaling methods are often applied with the aim of addressing climate model biases and localizing the output. This is done through techniques that make use of observational data at a finer spatial resolution. A large number of statistical downscaling methods exist and they can exhibit different performance characteristics when analyzing the central tendencies versus the tails of the distributions. Here we explore extreme value statistics in the downscaled output from a small number of statistical downscaling methods. We calculate a range of return-level estimates on climate variables based on daily precipitation, maximum temperatures and minimum temperatures using generalized extreme value (GEV) for block maxima and generalized Pareto (GP) distributions for peak-over-threshold distributions. The performance of the different downscaling methods is evaluated by comparisons of return-levels computed from downscaled values vs. the observational “truth”. A perfect model framework allows analyses to be done both for the historical period and for model simulated end-of-21st-century projections. We investigate which of the methods may degrade under changing climatic conditions.