How do we improve Earth system models? Integrating Earth system models, ecosystem models, experiments and long-term data


1st INTERFACE workshop, Captiva Island, FL, USA, 28 February–3 March 2011

Global land surface models – terrestrial models within more comprehensive Earth system models that typically also include atmospheric, ocean, and cryospheric components – are the primary tool to predict future changes in global terrestrial ecosystems and land–surface interactions with climate. Experiments and long-term observations can provide critical glimpses into ecosystem responses to global change and thereby improve the land surface component of Earth system models. To be most effective, experimental manipulations and observational studies should, ideally, be designed to confront key uncertainties in models and current knowledge, and provide data that have value in data-model comparison and assimilation. Experimental and observational scientists also rely on results from Earth system models to guide the development of new research. Therefore, a feedback between Earth System modelers and empirical scientists should emerge. With a common focus on understanding and forecasting the future of global terrestrial ecosystems and feedbacks with global change, direct and frequent interactions between scientists working with models, experiments and observations are needed. Facilitating such interactions was the primary goal of the first meeting of INTERFACE, the Integrated Network for Terrestrial Ecosystem Research on Feedbacks to the Atmosphere and ClimatE. The workshop focused on three primary topics: nutrient limitation of carbon storage on land, the role of changes in the hydrologic cycle in regulating climate feedbacks, and the role of acclimation in moderating responses of carbon uptake and loss to atmospheric and climatic change.

Modelling with experiments in mind, experimenting with modelling in mind

Global change experiments and long-term data collections are focused on two central goals: to gain an understanding of the ecosystem processes and to determine the future net ecosystem response to forcing variables. Peter Reich (University of Minnesota, Minneapolis, MN, USA), in his keynote talk, argued that it is critical to design experiments and observational studies to gain process-level understanding, rather than focusing primarily on the net response. Without the process-level knowledge that explains an ecosystem response, models may be developed or tuned to explain the observed experimental response using a set of processes that might differ from those operating in nature (i.e. they could inadvertently be made to get the right answer for the wrong reason). Reich identified plant phenology as a key process in need of a thorough understanding because failure to model phenology correctly can have profound impacts on productivity and respiration of terrestrial ecosystems. He used an example from an in situ forest warming experiment (B4WARMED), where six species of tree seedlings were warmed by 3.6°C. There was a highly species-specific increase in growing season of up to 30 d under warming, illustrating that a generalization of future phenology changes in Earth system models is not adequate. Reich encouraged the modelling community to update their models with newly gained process understanding from experiments, even if the models initially perform less well – such a ‘modelling with experiments in mind’ strategy will lead to ‘one step back followed by two steps forward’ because models will increasingly incorporate robust process-level understanding.

One key to improving Earth system models is to identify, through field experiments and model sensitivity studies, those processes that are most influential for a given ecosystem and at a given point in time. This means that the sensitivity of a response variable (e.g. net primary productivity, respiration, mortality) to the spectrum of individual forcing variables and their combinations (e.g. warming, elevated atmospheric carbon dioxide (CO2) concentrations, precipitation changes, nitrogen (N)-deposition] has to be understood. This will eventually facilitate the identification of potential tipping points in vegetation cover and geographic regions where models are particularly sensitive to model parameterization and structure. Although such tipping points are extremely important in the tropics (Cox et al., 2008; Meir & Woodward, 2010) and the arctic (Dorrepal et al., 2009), those biomes are drastically under-represented in experimental research. Also, many examples were discussed in which the forcing spectrum was in the nonsensitive area of the response variable: for example, Marcelo Sternberg (Tel Aviv University, Israel) reported from a long-term study with drought treatments in the Mediterranean, in which the imposed 30% reduction of precipitation was well within the natural variability of the rainfall at that site, so not surprisingly, the vegetation responded very little to the treatment (Talmon et al., 2011). Alan Knapp (Colorado State University, Fort Collins, CO, USA) showed that annual net primary productivity (ANPP) responses varied dramatically both within and among grassland types and were dependent on the local climate history and the forcing (rainfall pattern) applied to the system (Heisler-White et al., 2009). Claus Beier (Risø National Laboratory for Sustainable Energy, Roskilde, Denmark) asked whether we typically fail to explore the extreme ends of the forcing spectrum, and hence often miss the sensitive regions of the response variables. For example, in the case of precipitation experiments we might need more studies in which ecosystems are pushed to their limits, even though this may result in the death of the vegetation found there. Marcelo Sternberg concluded this discussion with the fundamental statement that ‘vulnerability to climate change decreases with increasing long-term climatic variability’, so that, for example, mesic sites, which often have a less variable climate, would be more vulnerable to perturbation than semi-arid or arid regions, which naturally experience greater climatic variability. In designing experiments, we therefore need to carefully plan the forcing that is applied to a given ecosystem, keeping in mind the models that will eventually make use of the data.

Killing trees – an example that needs more work

Jim Randerson (University of California, Irvine, CA, USA) asked whether ‘we know how to kill trees’ (in our models). In fact, while tree mortality under drought may be one of the most important quantitative components of the global carbon cycle in the future, we still lack a thorough mechanistic understanding of the phenomenon (McDowell et al., 2008; Adams et al., 2009; Leuzinger et al., 2009; Sala et al., 2010). This is clearly due to a lack of studies excluding all precipitation and inadequate methodology to identify the processes that ultimately lead to plant death. Margaret Torn (Lawrence Berkeley National Laboratory, Berkeley, CA, USA) emphasized that even if mortality can be correctly modelled in the present, the concurrent consideration of several global change drivers might complicate predictions of tree mortality substantially. For example, ecosystem responses to warming may change substantially when considered under concurrent drought and elevated atmospheric CO2. Claus Beier added that we generally lack experiments testing interactions between multiple global change drivers, and it is unlikely that the individual responses are additive (Leuzinger et al., 2011). It is therefore questionable whether we will be able to gain full mechanistic understanding of higher-order interactions on, for example, tree mortality, or whether we will have to rely on experimental and observational datasets for phenomenological parameterizations rather than mechanistic modelling. In any case, it was concluded that we need more experiments to elucidate patterns of tree mortality to more realistically model this fundamental process.

Role of ecosystem models as a process level test-bed

Plot- to landscape-scale ecosystem models can serve as the bridge between the process-level understanding gained from experiments or observations and Earth system models. Processes that are difficult to parameterize at the global scale can be developed in ecosystem models before being scaled-up for use in Earth system models. The ways in which ecosystem models could be developed for a number of key ecosystem processes were discussed at this meeting. The importance of clarifying experimental needs and focussing the representation of the process into a modelling framework before implementation into Earth system models was also highlighted. For example, Margaret Torn described the possible use of highly resolved ecosystem models as a test-bed for including current knowledge in relation to soil decomposition. Specifically, that molecular structure on its own cannot be used to predict soil organic carbon turnover times and that leaf and root litter do not compose at equal rates. In another example, Edward Rastetter (Marine Biological Laboratory, Woods Hole, MA, USA), used a theory-based ecosystem model to show that plant acclimation to optimal resource use has ecosystem-level impacts, including dynamics not captured when using the Liebig’s Law of the Minimum approach that is currently used in many Earth system models (Rastetter, 2011).

Key conclusions and future challenges emerging from the workshop

In order to improve the quality of predictions of the impact of global change on the terrestrial biosphere, we need both experiments and models to be involved early in the research process. Importantly, communication between the modelling and the empirical communities is necessary before experiments are implemented (e.g. Parton et al., 2007). This will help facilitate model development and enable simulations to be carried out using empirical data, thus aiding the design of empirical studies so that advancements in process-level understanding can be effectively integrated into models. The forcings applied in both models and experiments need to be selected carefully in order to identify key processes and sensitivities that can then be compared. However, single model experiment comparisons may not be sufficient. Rather, the distribution of model results may need to be compared to the distribution of experimental results (Bridget Emmett, Centre for Ecology and Hydrology, Wallingford, UK), and crosscutting syntheses must be promoted. Finally, while modellers and experimental scientists must collaborate more closely to conduct the right experiments for the right models and vice versa, funding agencies must also be more open to jointly planned studies. This meeting was an excellent starting point to promote thinking beyond the usual thematic and methodological horizon. The outcomes will be summarized in a series of reviews planned by the three topical focus groups, and ways forward will be suggested in the form of emerging proposals.