What are the limits to statistical error correction in land surface schemes when projecting the future?

Abstract

An artificial neural net is used to correct a land surface model (LSM) using present-day observational data. These data are split into a training and a testing set for the correction with the mean air temperature of each differing by a fixed amount. We investigate the success of the correction as this difference increases from 1°C to 20°C. We show that errors associated with the LSM's simulation of latent heat, sensible heat and net ecosystem exchange is substantially smaller if the model is corrected. The reduction in error remains large even as the temperature difference between the training and testing data sets reaches 10°C. We conclude that an artificial neural net can error-correct an off-line LSM under current and warmer climates, at least up to a warming of 10°C. The statistical correction of LSMs using present day data may therefore retain value in future climate applications. Copyright 2005 by the American Geophysical Union.