Joint Bias Correction of Multiple Climate Model Outputs for Impacts

Abstract:

Climate model output often contains significant biases that can hinder its use in impacts analysis. Recent work has shown that, of the many bias correction methods in use, the best overall performance is provided by distribution mapping. Distribution mapping corrects bias via a transfer function that adjusts data points such that the cumulative distribution function (CDF) of the model output matches the CDF of the observational data. However, this method is not guaranteed to preserve the relationships between variables when applied to the variables individually.

We present a new method of bias-correcting multiple variables jointly based on simultaneous diagonalization of the covariance matrices. This process transforms the variables into an uncorrelated form, where each component can be corrected independently using distribution mapping; the corrected variables are then transformed back into their original form, restoring the correlations of their joint distribution.

We apply the method to model output from NARCCAP (the North American Regional Climate Change Assessment Program), bias-correcting 7 impacts-relevant variables (daily minimum and maximum temperature, precipitation, incoming solar radiation, specific humidity, and u- and v-winds) to match the University of Idaho's METDATA, which combines NLDAS-2 reanalysis with PRISM observations to derive a high-resolution daily gridded observational dataset for the contiguous United States. This joint multivariate bias correction produces results that better capture regional climate processes, such as the seasonal pattern of differences in moisture flux on dry vs rainy days and seasonal changes in diurnal heating on clear vs cloudy days.