Climate data induced uncertainty in model-based estimations of terrestrial primary productivity

We employ the Lund-Potsdam-Jena General Ecosystem Simulator (LPJ-GUESS; Smith et al 2001, Smith et al 2014) to estimate GPP. LPJ-GUESS is a DGVM which uniquely combines an individual-based representation of woody plant growth, demography and interspecific competition with process-based physiology and biogeochemistry. It employs gridded time series of climate data (air temperature, precipitation and incoming shortwave radiation) as forcing and simulates the effects of climate on vegetation structure and composition in terms of plant functional types (PFTs), soil hydrology and biogeochemistry. Atmospheric carbon dioxide (CO₂) concentrations, nitrogen (N) deposition rates and soil physical properties (fixed) provide additional inputs. The 11 PFTs adopted for the present study and their prescribed parameters are given in tables S1 and S2 available at stacks.iop.org/ERL/12/064013/mmedia. We employ LPJ-GUESS version 3.0 which incorporates ecosystem nitrogen cycling, and the CENTURY soil biogeochemistry scheme (Parton et al 1993). A full description of the LPJ-GUESS is given in Smith et al (2014) and references therein.

2.1.1. Simulations

We compare simulated GPP to the flux tower-based global estimation of GPP from Jung et al (2011) and Jung et al (2017) (herein after, FLUXCOM) and the EO-based MODIS (Moderate-resolution Imaging Spectroradiometer) GPP product (MOD17 v55) from (NTSG 2015), at global scale during 2000–2009. These 10 years mark the temporal common time period of the datasets used in table 1. The FLUXCOM here includes four GPP products derived from different machine learning methods: model trees ensemble (MTE, Jung et al 2009), Artificial Neural Networks (ANN, Papale and Valentini 2003), Multivariate Adaptive Regression Splines (MARS, Friedman 1991) and Random Forests (RF, Breiman 2001).

The land cover classification (figure 1) is used to aggregate the global land area into six land cover categories: tropical forest, extra-tropical forest (boreal and temperate), semi-arid ecosystems, tundra and arctic shrub land, grasslands and land under agriculture (crops, here combined), and areas classified as barren (sparsely vegetated). This classification is based on the MODIS land cover classification, MCD12C1 type3 (Friedl et al 2002) and the Köppen-Geiger climate classification system (Kottek et al 2006) following Ahlström et al (2015).

Zoom In Zoom Out Reset image size

We partition climate induced uncertainty in simulated GPP into temperature, precipitation and shortwave radiation. Six climate datasets are available in this study, which limits our ability to partition uncertainties to specific climatic drivers. Therefore, we combine climate variables of the different datasets in such a way that a given simulation can be forced by the temperature from one dataset, the precipitation from another and the radiation from a third dataset. This approach also allows us to investigate the potential maximum GPP uncertainty stemming from climate variables from different climate datasets. The maximum potential climate induced uncertainty is found by first calculating the apparent model sensitivities to each of the three climate drivers using multiple linear regressions on an ensemble of simulations where climate variables are combined from a subset (n = 3) of the six climate datasets. The maximum potential GPP uncertainty is then found by multiplying the apparent model sensitivities to the climate drivers with the range (maximum minus minimum) of each of the climate drivers at each location (grid cells).

We select three out of six datasets (CRU, NCEP and ECMWF) since their corresponding global GPP represent the minimum, median, and maximum (997, 1064, and 1089 g C m⁻² yr⁻¹, respectively) within the 6 datasets. We run the simulations with the full combination of these 3 dataset (27 simulations), plus 3 original simulations for the other climate datasets. With these 30 simulation results, we use a multiple linear regression approach to identify the relative importance of each climate variable to total climate induced uncertainty in GPP. The regression is performed separately for each grid cell, cross climate datasets or combinations of climate variable among datasets but not along the time series:

where GPP, R, P and T represent z-scores (number of standard deviations from the mean) of the annual mean values of GPP, shortwave radiation, precipitation and temperature, respectively, used to force the simulation under consideration. The regression parameters α, β and γ represent the apparent model sensitivities to each of the drivers while ε is the residual error term. We refer to these sensitivities as apparent model sensitivities because they may be influenced by co-variation between climate variables (e.g. Piao et al (2013)). To investigate the potential influence of co-variation between the predictors we performed principal components regression (PCR) for analysis of the relationship between GPP and climate. Both analyses yield very similar sensitivities (see supplement 2 and figures S1–4) which add confidence that the sensitivities presented here are good approximations of the model's sensitivities and not heavily influenced by predictor co-variation.

The true sensitivities should ideally also include interaction terms between the variables, and other drivers not explicit in (1), for instance atmospheric CO₂ concentration, may also contribute to the variability of GPP. However, for clarity we choose to illustrate an isolated effect of a driver on the climate induced uncertainty. This is motivated by the fact that interactions only explain a small part (∼5%) of the response variability in GPP according to ANOVA analysis (figure S5).

Once the regression coefficients (α, β, γ) are determined, the partial GPP uncertainty caused by the corresponding variables (shortwave radiation, precipitation, temperature) is estimated by:

where GPP_{R_unc} is the partial GPP uncertainty caused by shortwave radiation (2), and α presents an apparent GPP uncertainty sensitivity to shortwave radiation range, R_{_range}, which is the difference between the minimum and maximum shortwave radiation value across forcing datasets. Equations (3) and (4) are analogous cases for precipitation and temperature, and the total GPP uncertainty (GPP_unc) is the sum of the three partial GPP uncertainties.

We also assess how the uncertainty in GPP arises from a combination of climate data uncertainty and apparent model sensitivity by decomposing the partial GPP uncertainty into components associated with the climate data range and simulated apparent sensitivity of ecosystem GPP to the respective driver. Figure 2 shows an example of two possible outcomes with the same partial GPP uncertainty caused by shortwave radiation but with different contributions from climate data range and apparent model sensitivity. For Case 1, apparent model sensitivity contributes more than the climate data range, while the climate data range has higher influence in Case 2. The regression slopes provide their relative contributions to the partial GPP uncertainty (being larger or smaller than 1). Therefore, the fractions of apparent model sensitivity (f_sen) and climate data range (f_range) that contribute to a climate driver (e.g. shortwave radiation) induced GPP uncertainty can be estimated by:

where α is the regression coefficient of multiple regression for shortwave radiation, ∣α∣ stand for the absolute value of α.

Zoom In Zoom Out Reset image size

Global GPP simulated by LPJ-GUESS when forced by the six climate datasets without any mixing of climate variables between the datasets shows agreement with FLUXCOM and MODIS GPP (figure 3(a)). However, when stratified by land cover classes, the simulated GPP of the tropical forest (figure 3(e)), shows markedly lower values than both FLUXCOM and MODIS GPP, while for grasslands, croplands (figure 3(d)) and extra-tropical forest (figure 3(f)), simulated GPP is higher than both FLUXCOM and MODIS. These biome-specific discrepancies tend to cancel out to produce a simulated global GPP estimate closer to the remote sensing and flux tower-based estimates (Piao et al 2013). Aggregated to the global land surface area, the range among simulations amounts to 11 Pg C yr⁻¹ or 9% of mean GPP.

Zoom In Zoom Out Reset image size

Here we show inferred apparent model sensitivities (section 3.2.1) and the maximum potential climate induced GPP uncertainty (section 3.2.2) as inferred by applying the regression-based partitioning on an ensemble of (n = 30) simulations. The simulations differ only in forcing climate, where climate variables from n = 3 climate datasets have been mixed to new and unique combinations. This method allowed us to better investigate the apparent model sensitivity and to investigate the maximum potential climate induced GPP uncertainty. Section 3.2.3 describes the results of partitioning of GPP uncertainty into apparent model sensitivity and climate data range.

3.2.1. Local apparent model sensitivities.

3.2.2. Potential maximum climate induced GPP uncertainty

3.2.3. Partitioning of GPP uncertainty into apparent model sensitivity and climate data range

By using only a single model (LPJ-GUESS), our analysis excludes quantification of uncertainty stemming from differences in the structure and parameterization of alternative carbon cycle models. However, LPJ-GUESS is a well-established DGVM that has been evaluated and applied in a wide range of published studies, and has also been found to show relatively similar predictive skills and response to climate variations compared to other global ecosystem models (McGuire et al 2012, Murray-Tortarolo et al 2013, Piao et al 2013, Sitch et al 2015). Most models simulate similar GPP inter-annual variations at global scale, and these have been traced to large-scale variations in climate, particularly linked to global circulation phenomena such ENSO (Ahlström et al 2015). While our study does not address variation among models in sensitivity to climate forcing, we believe that our findings are likely to be representative for global ecosystem models as a class. LPJ-GUESS may thus tentatively be considered representative for how other DGVMs and carbon cycle models behave in response to uncertainties induced by climate forcing data. Although FLUXCOM and MODIS are considered as evaluation data in this study, both datasets involve considerable components of modeling (additionally to the measurement uncertainty) and are therefore also subject to some limitations and uncertainty (Zhao et al 2006, Jung et al 2011, Lin et al 2011). Further, both empirical datasets depend on climate data to extrapolate site measurements to gridded estimates of GPP. Previous studies have evaluated sources of uncertainty in the FLUXCOM (Jung et al 2009, Beer et al 2010, Jung et al 2011), and MODIS GPP products (Running et al 2004, Zhao et al 2005, Zhao and Running 2010).

The six climate datasets analysed are derived from either quasi-point-based measurements or climate model-based reanalysis. Station networks, however, vary greatly in density across the globe, with sparse coverage over certain areas, such as the high latitudes (Jones and Harris 2013). Hence the data represent a limited sampling fraction and include an unknown error. Comparing the three main climatic variables of these datasets (figures S10–S12) reveals no strong overall difference in temperature over global and regional scales while a clear difference can be observed for precipitation and shortwave radiation, especially over tropical regions. Our results bring forward that these differences translate into large differences in simulated GPP in our model, and likely in other carbon cycle models.

The potential maximum climate induced uncertainty, if we combine climate variables from different datasets, is found to be 32% of mean GPP. Although the potential maximum uncertainty appears large, Barman et al (2014) found that the climate induced uncertainty could reach 20%–30% for simulations of savanna, grassland, and shrubland vegetation types. However, the way the maximum potential uncertainties are calculated here assumes that the most extreme values of each climate variable at each location are compared to create the simulated maximum and minimum GPP. This maximum uncertainty is therefore rather an illustration of the importance of climate drivers and to a lesser extent a representation of actual and realistic uncertainties; in most simulations cancellation effects will reduce the aggregated uncertainties presented here.

The tropical region shows disproportionate large climate induced uncertainty (figure 5) and empirical uncertainty based on observations (figure 3(e)). Earlier studies have likewise identified the tropics as a region of high spread in estimated GPP depending on forcing (Zhao et al 2006, Poulter et al 2011, Ahlström et al 2012b, Anav et al 2015). Data limitations are likely to be an important source of a high model spread in GPP. Meteorological station networks are generally sparse across the tropics (Medany et al 2006). Moreover, characterization of the climate, as based on measurements and modeling, is challenging for tropical regions due e.g. to the influence of extreme climate events (Trewin 2014, Wentz 2015) and the impact of heavy and extended cloudiness on remote sensing measurement (Fensholt et al 2007). Furthermore, the results show that the large uncertainty is mainly due to precipitation and radiation, which coincides with findings at site level of Barman et al (2014). Jung et al (2007) suggested that cloud and aerosol physics (which govern precipitation and radiation transfer) are most likely the principal causes of differences in precipitation and radiation estimates between datasets, since cloud and aerosol properties are implemented differently in different meteorological reanalyses. Our finding as to the global distribution of climate-induced uncertainty in GPP (figure 4) agrees with a recent empirically based study of the sensitivity of global terrestrial ecosystem to climate variables (Seddon et al 2016), who found ecologically sensitive regions with amplified responses to climate variability over the world. Moreover, the relative contributions of drivers show similarities in spatial patterns compared to the study by Nemani et al (2003) mapping the primary climate constraints to plant growth.

Overall, climate data range contributes more uncertainty to simulated global GPP than the sensitivity of the simulated ecosystem processes to climate forcing. This implies that errors in climate datasets play an important role in model-based carbon cycle estimations, and may exceed the importance of shortcomings in ecosystem model structure or parameterisation. We found that GPP uncertainties in simulations by our model are associated strongly to precipitation in large spatial extent (figure 4). In contrast to the general case described above, precipitation induced uncertainty in dry and tropical ecosystem is also relatively strong associated to apparent model sensitivity to the climate forcing (figures 6(c) and (7)). In other words, LPJ-GUESS simulated GPP has a strong sensitivity to precipitation at the global scale, though a model intercomparison study suggests DGVMs in general may be even more sensitive to precipitation than LPJ-GUESS (Piao et al 2013). Recent studies emphasize the importance of precipitation variation in controlling vegetation growth, driving inter-annual variability and dominating uncertainty in predictions of future plant production (Beer et al 2010, Ahlström et al 2012a, Ahlström et al 2015).

Aside from limitations, discussed in section 4.1, to the generality of our findings due to the choice of a single (though arguably representative) carbon cycle model, this study is limited by the availability of independent climate datasets. The six climate datasets used here are relatively few to accurately estimate ensemble summary statistics, therefore we decide to combine climate variables of the different datasets to assess the potential maximum climate induced uncertainty. Current (Model Intercomparison Projects) MIPs do use climate forcing that are combinations of other datasets (e.g. Séférian et al (2015)), which is valuable to investigate the influence of differences between climate datasets. We apply the multiple linear regression method mainly because the different datasets vary only to a limited extent and therefore the assumption of a linear response of GPP is reasonable. If larger differences between the datasets would have been found in the screening of the data, a non-linear parameterization of the relationship between GPP and the climate variables would have been required (e.g. GPP is positively related to temperature over a certain part of the temperature gradient but too high temperature can cause drought stress). The partitioning between uncertainty stemming from apparent model sensitivity and data uncertainty should therefore be considered indicative and exact numbers are not reported here. Moreover, previous studies (Piao et al 2013, Seddon et al 2016) have addressed empirically linear relationship between vegetation productivity and climate.

We acknowledge that dependencies between the three climate variables exist, but since we are interested in the response of the model to differences in the climate data we consider these potential inter-dependences as less important for our purpose. Two-way or three-way interactions among climate variables can be substantial for both carbon and water processes. These multiple forcing interactions are not included in this study due to a small contribution to GPP variability (figure S5). It will be a future research challenge to adequately quantify interactive climatic uncertainty and corresponding apparent model sensitivity and interactive climatic data range.