The influence of natural variability and interpolation errors on bias characterization in RCM simulations

Abstract

Climate model simulations are routinely compared to observational data sets for evaluation purposes. The resulting differences can be large and induce artifacts if propagated through impact models. They are usually termed “model biases,” suggesting that they exclusively stem from systematic models errors. Here we explore for Switzerland the contribution of two other components of this mismatch, which are usually overlooked: interpolation errors and natural variability. Precipitation and temperature simulations from the RCM COSMO-Community Land Model were compared to two observational data sets, for which estimates of interpolation errors were derived. Natural variability on the multidecadal time scale was estimated using three approaches relying on homogenized time series, multiple runs of the same climate model, and bootstrapping of 30 year meteorological records. We find that although these methods yield different estimates, the contribution of the natural variability to RCM-observation differences in 30 year means is usually small. In contrast, uncertainties in observational data sets induced by interpolation errors can explain a substantial proportion of the mismatch of 30 year means. In those cases, we argue that the model biases can hardly be distinguished from interpolation errors, making the characterization and reduction of model biases particularly delicate. In other regions, RCM biases clearly exceed the estimated contribution of natural variability and interpolation errors, enabling bias characterization and robust model evaluation. Overall, we argue that bias correction of climate simulations needs to account for observational uncertainties and natural variability. We particularly stress the need for reliable error estimates to accompany observational data sets.