Comparing global models of terrestrial net primary productivity (NPP): introduction

Improving the knowledge about broad-scale ̄uxes of carbon between the atmosphere and the biosphere is high on the scienti®c agenda for at least two reasons. First, productivity is fundamental to ecology. While it has been a central focus over the last three decades, beginning with the International Biological Programme (IBP) and continuing through the International Geosphere-Biosphere Programme (IGBP), there was, until recently, remarkably little progress in estimating productivity, or the carbon balance of large regions. Second, large-scale ̄uxes of CO2 between the land and the atmosphere are increasingly relevant to policy, because carbon storage by land ecosystems can play an important role in limiting the rate of atmospheric increases (IGBP Terrestrial Carbon Working Group 1998). Computer models have now gained broad acceptance for translating ecological hypotheses, derived from local observations, into estimates of regional, continental or even global outcomes of ecosystem processes. Indeed, such models are perhaps the only feasible method to make spatially detailed estimates for large regions. Such models are rather useless, however, if their behaviour cannot be evaluated. Ideally, productivity models should be evaluated based on comparions with observations. This creates a dilemma, however, for global ̄uxes such as net primary productivity (NPP). Direct observations of NPP are not available at the global scale, and direct validation is therefore not feasible. Model intercomparisons provide additional options for studying the behaviour of the models. Clear protocols, common data sets, and standardized output can help to ensure that results are comparable. Once a data base of model results has been created, it becomes possible to investiate speci®c features of the behaviour of participating models, including supporting or discarding some of the assumptions. Under the scienti®c sponsorship of the IGBP, such a model intercomparison has been carried out at the Potsdam Institute for Climate Impact Research. Seventeen terrestrial biosphere models (TBMs) participated in a study with common protocols, goals, and input data summarized in the overview paper by Cramer et al. (pp. 1±15). The models all simulate NPP of the land biosphere for an average year, with calculations based on a broad range of structures, complexity, and driving data. One of the key differences is that some of the models use satellite data (especially AVHRR data from the NOAA weather satellites), while others use data on climte and soils alone. It was not unexpected that no two models gave identical results. Still, all models agreed on basic features of the biosphere, such as low productivity in dry and/or cold regions and high productivity in humid tropical forests. Given the sparse and incomplete database of NPP observations, no model stood out as the `winner'. The bene®ts of the project, many of which are reported in the seven papers of this issue, are deeper but also richer than ®nding a single best model. First, practically all modelling teams experienced `surprises' during the intercomparison, many of which led them to reconsider speci®c aspects of the models. Improvements based on these reconsiderations did not emerge early enough in the comparison process for them to affect the simulations in the papers here. But opportunities for improvements were apparent to most workshop participants, and we postulate that they have indeed boosted the overall understanding of the terrestrial biosphere. Second, the analysis of Kicklighter et al. (pp. 16±24) identi®es the geographical regions with good model agreement as opposed to those with signi®cant differences (particularly boreal forests in summer and tropical evergreen forests in the dry season). This has direct implications for uncertainties in speci®c processes represented by the models. The paper by Schloss et al. (pp. 25±34) goes further by pointing out that NPP estimates are usually in closer agreement for regions where temperature is the main controlling variable. Water shortage over a signi®cant part of the season introduces additional uncertainty, a phenomenon analysed more in depth by Churkina et al. (p. 46±55). Both Kicklighter et al. and Schloss et al. stress that model performance is critically affected by the description of vegetation structure and whether that description comes from a map or is calculated by the model. The analysis by Bondeau et al. (p. 35±45) expands on this by comparing canopy changes through a given year against R