Effects of systematic predictor selection for statistical downscaling of rainfall in Hawai'i

Abstract

While there has been rapid advancement in the development and application of statistical downscaling methods for climate projection, determining the best predictive large‐scale climate information for the targeted local climate variable remains a challenge. The choice of predictor variables is one of the most influential steps of model development and has the potential to lead to varying results, contributing to the total uncertainty of future projections. Despite being a well‐known problem, predictor selection often does not receive adequate attention and the development of a straightforward and feasible prescreening process is needed to provide guidance to simple (e.g., linear regression) and complex (e.g., machine learning) modeling tools alike. In this project, Akaike information criterion (AIC) and leave one out cross validation are used to evaluate sets of predictor variables common in statistical downscaling models of precipitation (e.g., temperature, geopotential height, moisture transport). Because thousands of predictor sets were found to be competitive in their statistical skill in projecting rainfall, results suggest that an ensemble of predictor sets should be used to account for the resulting variance associated with predictor selection. This ensemble approach is applied to make improvements to the most recent statistical downscaling model developed for rainfall projection in Hawai'i. An ensemble validation study is performed by projecting future wet season rainfall in Hawai'i for scenario RCP4.5 using 17 CMIP5 GCMs with ensembles of highly ranked and poorly ranked predictor sets. Results show significantly less variance and improved agreement in the projected sign of precipitation change for the highly ranked predictor set ensemble. In conclusion, it is recommended that statistical downscaling models are run in an ensemble‐mode with multiple combinations of predictors instead of using a single model, allowing for additional quantification of model uncertainty and a “safety net” for when the true physically best‐suited predictor set is unknown.