Enhancing the efficiency of terrestrial biosphere model simulations by reducing the redundancy in global forcing data sets

Abstract

Data sets of climatic variables and other geographic characteristics are becoming available in increasingly higher resolutions, resulting in substantial computing burdens for simulation models of the terrestrial biosphere. But by how much do higher resolutions of forcing data actually contribute to higher accuracy in model predictions? I investigated this question using the Cramer-Leemans climatology as an example for a high resolution forcing data set and a model of net primary productivity (NPP). I first used cluster analysis to reduce the complete grid of the climatology to a few grid points, each representative of regions with similar values. A global map of NPP was reconstructed by using the simulated values of the representative grid points for the respective regions. I then compared the reconstructed map of NPP to the one obtained from all grid points. The results show that a high accuracy in simulating the high resolution pattern and magnitude can be achieved by only considering a comparatively small subset of representative grid points. What this suggests is that, while high resolution data sets provide the necessary means to determine the typical regions, they do not add much accuracy to the overall outcome of model simulations because they contain many grid points with similar values. By reducing this redundancy, the methodology used here allows model simulations to be considerably more computing-time efficient while still retaining the accuracy in predicted quantities.