Climate sensitivity of global terrestrial net primary production (NPP) calculated using the reduced-form model NNN

Abstract

ABSTRACT: In order to assess the potential impact of climate change on terrestrial equilibrium net primary production (NPP), information about the sensitivity of terrestrial NPP to climate change is needed. A novel approach to the definition of climate sensitivity is introduced; it does not depend on specific (and uncertain) scenarios, but rather describes the worst-case climate sensitivity of NPP, as measured by the magnitude of the gradient of NPP, as a function of the climate variables. The metric in climate space necessary for the determination of the gradient is calculated using the existing spatial variation of the climate variables as a measure of the potential for climate change, taking into account the unchanging determinants of the climate‹latitude and altitude. The current correlations between the climate variables are preserved using principal component analysis. The resulting map of global NPP sensitivity shows especially high values, e.g. in the US Midwest, southern Africa, Australia, western Kazakhstan, the Maghreb and Spain. The sensitivity is aggregated to the country level, excluding the effects of the very insensitive deserts, in order to make these results applicable for policy analysis. The reduced-form model ŒNNN¹ predicts annual terrestrial NPP of potential natural vegetation in equilibrium on the basis of a climatology including monthly values for temperature, precipitation and light. The very short computing time of this model is a prerequisite for the above-described multidimensional sensitivity study. To construct NNN, the average of the global NPP results of 7 advanced climatology-driven functional vegetation models was used to obtain a Œbest guess¹ NPP field at a 0.5° x 0.5° spatial resolution. With the underlying climatology (36 values for monthly mean temperature, precipitation and light intensity per grid element), 62483 points of an R36 Ø R1 mapping are defined. A subset of these was used to train a neural network yielding a good reproduction of the spatial pattern with a mean absolute error of 0.026 kgC m-2 yr-1, which is significantly less than the mean uncertainty of the NPP average (mean absolute deviation: 0.097 kgC m-2 yr-1). Furthermore, it is shown that the simple model calculates moisture-limited regions correctly, indicating that functional properties of the original models are reproduced. We hope with the NNN to make a contribution to other research which needs a very fast reduced-form NPP model. The NNN model is accessible in FORTRAN or as a C-subprogram available at: http://www.pik-potsdam.de/~oliver/NPP.h tml.