Representing subgrid-scale edaphic heterogeneity in a large-scale ecosystem model: A case study in the circumpolar boreal regions



[1] Strong feedback interactions occur between terrestrial ecological processes and climate. However, vegetation dynamics have not been adequately reproduced by many large-scale ecosystem models and coupled climate models because local environmental variations are not sufficiently considered. By estimating subgrid-scale ecosystem productivity potentials within large grid cells, we explicitly incorporated local edaphic heterogeneity into the Spatially Explicit Individual-Based Dynamic Global Vegetation Model (SEIB-DGVM). We assumed that high-resolution land cover data reflected the underlying ecosystem productivity potentials originating from edaphic variations. We categorized high-resolution Global Land Cover 2000 database (GLC2000) pixels into high, intermediate, and low productivity potentials and obtained fractional cover of these potentials on a DGVM grid. Then we performed transient dynamic vegetation simulations on circumpolar boreal regions with high, intermediate, and low photosynthetic capacities. Using the productivity potentials as weight, we estimated total leaf area index (LAI) that integrates the smaller-scale variation of each DGVM grid. Ground-based observations and a remote sensing product were used to evaluate the model performance. The regional pattern of simulated LAI was significantly improved, especially for the North American boreal regions. Improvement was minor for northern Eurasia, suggesting the qualitative classification by GLC2000 was the critical control on our approach. The results suggested that even a rather simplistic consideration of subgrid-scale heterogeneity could significantly improve large-scale simulations.

1. Introduction

[2] Ecosystem functions on land arising from ecological dynamics are important controlling factors of regional and global climate, and the representation of terrestrial land-surface processes in climate models is now one of the most urgent requirements for a better understanding of future climates [Moorcroft, 2006; Purves and Pacala, 2008]. For large-scale terrestrial ecosystems, the conventional modeling method is the “big-leaf” approach, in which the functions of gas exchange and element cycling occur over the homogeneous space of a given grid (“big leaf”). However, because ecological processes are inherently operating on much smaller scales than the sizes of typical climatological cells, and also because of the nonlinear nature of vegetation populations and community dynamics, it is not sufficient to represent land surfaces as homogeneous big leaves [Levin et al., 1997]. The simplistic big-leaf approach implicitly assumes monopoly by a single plant functional type (PFT) on a grid. This lack of functional diversity (i.e., redundancy) could be responsible for some overly drastic projections from big-leaf simulations, such as the prediction of a massive die-off of the Amazonian rainforest in this century [Cox et al., 2000].

[3] Small-scale environmental heterogeneity due to edaphic factors significantly affects vegetation [Nichols et al., 1998]. Local variations in physical characteristics often result in a mosaic-like landscape of discrete vegetation types. Good examples are distinct upland/lowland ecosystems in the boreal landscape, variations in community composition and physiognomy due to slope and aspect, and vegetation patchiness due to soil chemistry [Turetsky et al., 2005]. However, these edaphic variations are not fully represented in the conventional DGVMs [e.g., Sitch et al., 2003]. Moreover, the big-leaf approach is often prone to random and/or human biases—if the environmental conditions assigned for a particular grid by the model are not appropriate to represent the “average” site, the simulation for the entire grid will have systematic biases.

[4] A recent advancement for the treatment of biotic diversity is represented in SEIB-DGVM [Sato et al., 2007]. Structured on an individual-based forest model [Pacala and Deutschman, 1995], SEIB-DGVM simulates dynamics of population (e.g., recruitment and mortality) and community (e.g., competition and succession) according to the ecophysiology of several PFTs in a spatially-explicit representative stand. By incorporating biotic heterogeneity within grid, this model allows researchers to study nonlinear ecological responses under environmental changes. The timing and transient dynamics of vegetation changes such as colonization, die back, and biome shift and the resultant terrestrial feedback to the atmosphere can be estimated by the individual-based approach [Purves and Pacala, 2008].

[5] In this study, we introduced the subgrid-scale edaphic heterogeneity to SEIB-DGVM, which already has biotic heterogeneity, to realistically simulate large-scale ecosystem conditions. Specifically, edaphic heterogeneity includes abiotic environmental factors such as the physical and chemical characteristics originated from geomorphological, geochemical, and topographical features that we assumed to be stable for the time frame of the DGVMs. These edaphic variations affect vegetation through nutrient, water, substrate stability, and microclimatic conditions [Nichols et al., 1998]. By explicitly treating both biotic and edaphic heterogeneity simultaneously, this simulation exercise could provide a powerful framework to appropriately simulate ecosystem structure and function in DGVMs and coupled climate models.

2. Model Framework

[6] SEIB-DGVM simulates forest dynamics on spatially explicit 30 × 30-m stands and extrapolates the outcomes from the representative stands over the entire DGVM grid. This approach allows researchers to examine the simulation against field data with coexisting PFTs. SEIB-DGVM will be coupled to CCSR-FRCGC GCM (Center for Climate Systems Research – Frontier Research Center for Global Change GCM [Kawamiya et al., 2005]) for IPCC AR5 simulations. For this study, simulations were run at a grid resolution of 1° × 1°, and the subgrid-scale edaphic variations were treated with the method described below.

2.1. Study Area and Regional Datasets

[7] Simulations were implemented on the circumpolar boreal regions. The grid-based spatial extent was defined using the biome classification of the Integrated Biosphere Simulator (IBIS) [Foley et al., 1996]. Since the IBIS dataset has a resolution of 0.5° × 0.5°, we reduced the resolution to 1° × 1° to conform to our grid resolution and defined a cell as “boreal” if two or more smaller cells nested within the larger cell were boreal.

[8] To run simulations, air temperature at 2 m, soil temperature at 5–10 cm depth, precipitation, cloudiness, wind speed, and specific humidity were obtained from the NCEP/NCAR reanalysis [Kalnay et al., 1996] for 1948–2004. These 57-year daily meteorological data were sequentially and repeatedly used to run the model for a simulated period of 200 years with the assumption that no significant edaphic changes occurred during this time. In addition, the CRU monthly mean climatology of 1961–1990 [New et al., 2000] was used to adjust the air temperature and precipitation averages. Monthly mean air temperature and precipitation of the NCEP/NCAR dataset were calculated for each grid and compared to the CRU climatology, and the excursion from the mean climatology was corrected additively for air temperature and multiplicatively for precipitation.

2.2. Edaphic Variations

[9] GLC2000 (2003) is a SPOT4-VEGETATION derived land cover classification dataset at ∼1-km resolution. Using multi-temporal satellite observations, characteristic behavior such as phenology was detected to classify land cover. This classification scheme is qualitative and regionally different because each region (e.g., North America, northern Eurasia, etc.) was mapped by local experts. We assigned productivity potentials to GLC2000 classes based on categorical descriptions. For North America, there are 29 classes [Latifovic et al., 2003], and closed-canopy forests were assigned to the high productivity potential, open woodlands and closed shrublands were assigned to the intermediate productivity potential, and other less-dense vegetation types were assigned to the low productivity potential (Table S1 of the auxiliary material). The Northern Eurasia dataset has 28 classified cover types [Bartalev et al., 2003]. We assigned various closed forest types to high, patchy forests to intermediate, and less densely vegetated areas to low productivity potential (Table S2). Fractions of a DGVM grid at latitude i longitude j covered by high, intermediate, and low productivity potentials (Figure 1) are:

equation image
equation image
equation image

where Nhighi,j, Nintermediatei,j, and Nlowi,j are numbers of GLC2000 cells (∼1-km resolution) of high, intermediate, and low productivity potentials, respectively, within a DGVM cell (1° resolution). Since there are ∼10,000 GLC2000 cells in a DGVM grid, we assumed that the calculated fractions appropriately represent the relative productivity potentials of the grid.

Figure 1.

Productivity potentials of the boreal regions based on the GLC2000 [IES Global Environment Monitoring Unit, 2003] classification. The productivity potential shown here is a weighted average of high, intermediate, and low productivity from the fraction of the total grid area covered.

[10] We then ran three independent SEIB-DGVM simulations, varying the maximum photosynthetic rates to represent the intrinsic differences in productivity due to edaphic variations. We defined the default configuration of SEIB-DGVM as the high productivity simulation in closed forests with optimal photosynthetic capacity. Estimated from a chronosequence study in the BOREAS Study Areas (stand ages ranging 20–131 years [Bond-Lamberty et al., 2004]), Canada, for the low productivity simulation, the photosynthetic capacity was reduced to 60%. For the intermediate productivity simulation, the photosynthetic capacity of coniferous trees was reduced to 80% and that of broadleaf trees was reduced to 60%, based on a theoretical framework of the PFT-specific tolerance to nutrient limitation (Text S1). The overall simulated leaf area index (LAItotal) of the grid at latitude i longitude j is a weighted average of the three simulated LAIs (LAIhigh, LAIintermediate, and LAIlow):

equation image

where LAIhigh, LAIintermediate, and LAIlow are LAI outputs from high, intermediate, and low productivity simulations, respectively.

2.3. Validation

[11] The model's dynamics were first compared to field observations made at the BOREAS sites. PFT compositions as functions of stand age were obtained from a biometry report (TE-13) [Apps and Halliwell, 1999]. The high productivity simulation was compared to field sites in mesic broadleaf or mixed stands. The intermediate productivity simulation was compared to sandy upland stands dominated primarily by jack pine (Pinus banksiana). The low productivity simulation was compared to black spruce (Picea mariana) stands on organic-rich continental bogs. The transient dynamics in the forest LAI after a stand-replacing disturbance were then compared to chronosequence datasets [Bond-Lamberty et al., 2004] and the TE-13 biometry. The upland and lowland chronosequence data were plotted with the high and low productivity simulation, respectively. Jack pine stands in TE-13 were plotted with the intermediate productivity simulation.

[12] The regional simulations of the circumpolar boreal regions were then compared to the LAI product derived from the MODIS Terra MOD15A2 C4.1 [Yang et al., 2006]. For the dataset spanning 2000–2006, we took a mean of the MODIS LAI product of July of each year to produce the average summertime LAI.

3. Results

[13] The model simulations were significantly different among the three productivity potentials at BOREAS sites (Figure 2). In the high-productivity stand, the forest in the early stage was dominated by the broadleaf deciduous PFT due to its fast growth and intensive resource use. As the stand gradually became mature, inter- and intra-PFT competition was intensified, and the stand was gradually replaced by the coniferous PFT, the late-successional, shade-tolerant group. In the intermediate-productivity stand, the total LAI at maturity was similar to that of the high-productivity stand, but here the broadleaf PFT was more suppressed by the coniferous group. This pattern of PFT composition was also observed in field data from sites with intermediate productivity; broadleaf trees were only occasionally observed in young and intermediate sites and almost excluded in mature sites [Apps and Halliwell, 1999]. In the poor stand, the total LAI was suppressed due to resource limitations. Since the canopy of coniferous PFTs did not completely close even at maturity, the broadleaf PFT persisted as a minor component.

Figure 2.

Difference in productivity and successional patterns at 55.5°N, 98.5°W (the grid containing BOREAS Study Areas). Simulations with (a) high, (b) intermediate, and (c) low productivity potentials. In Figures 2a and 2c, the LAI dynamics data from a chronosequence study (X) were superimposed on the simulated dynamics. The field data from upland forests were compared to the high productivity simulation, and those from lowland boggy forests were compared to the low productivity simulation. In Figure 2b, stand ages and site LAIs estimated from Apps and Halliwell [1999] were shown with ±1 SD.

[14] The simulation results were compared to observed LAIs. The high and low productivity simulations were plotted against the upland and lowland chronosequence datasets, respectively. The intermediate case was compared to the TE-13 biometry. In general, SEIB-DGVM reproduced the stand development trajectory and the equilibrium LAI within the observational variabilities. On each edaphic type, the simulation produced dynamic equilibria of LAI before a stand age of 50 years. This simulated result agreed adequately with the lowland site, but at the upland sites, LAI maturity lagged behind that of the simulations. The trajectory of the intermediate simulation was in the large variability of the field observation.

[15] We obtained the 2000–2006 mean July LAI from the MODIS dataset (Figure 3a) to analyze the regional simulation. Applying the productivity index from GLC2000, we took weighted means of three productivity simulations due to subgrid-scale heterogeneity (Figure 3b). The simulation with edaphic heterogeneity was generally more consistent with observations than the default SEIB-DGVM simulation (Figure 3c). In North America, the representation of edaphic heterogeneity successfully improved the regional simulation. However, even with edaphic heterogeneity, the simulation consistently overestimated LAI in most of western Eurasia. The possible cause of this bias was the productivity potentials constructed from GLC2000 (Figure 1). In North America, the productivity map showed gradual variations in productivity potential. However, in Eurasia, with the classification scheme that differed from that of North America, most of the areas had high productivity, there were no classes such as open woodland that potentially has intermediate productivity, and the resultant LAI was close to the default SEIB-DGVM. There was a systematic underestimation in eastern Siberia, although the productivity potential was very high in this region. This was probably caused by the inadequacy in ecophysiological representation for the deciduous conifer PFT, the dominant vegetation of this region.

Figure 3.

Regional LAI estimations from remote sensing observation and simulation. (a) From MODIS. (b) SEIB-DGVM simulation with subgrid-scale heterogeneity. (c) Default SEIB-DGVM simulation without subgrid-scale heterogeneity.

[16] To visualize the performance of this approach, latitudinal gradients of simulated LAI were compared to the observations (Figures S1 and S2). In eastern Siberia (50–66°N, 110–140°E), the model reproduced the higher productivities in flat terrains of 50–55°N and showed a steep decline in mountainous regions of 55–57°N. However, the rise in LAI around 60° N was not reproduced, and the model underestimated LAI in this region. In North America (48–62°N, 54–115°W), the simulated and observed LAI showed a relatively constant decline along the latitude. The overall mean LAI of the simulation (3.12) was very close to the observation (3.10). However, the latitudinal slope of the LAI decline of the simulation (−0.073, p = 0.019) was shallower than that of the observation (−0.160, p = 0.001).

4. Discussion

[17] Using the categorical land cover classification by GLC2000 with high resolution and semi-continental coverage, we created fractional maps consisting of areas with high, intermediate, and low productivity potentials for the circumpolar boreal regions. The DGVM predictions of boreal LAI made from a weighted average in productivity were significantly improved from the default simulation when appropriate fractions were provided. Our study indicated that the explicit consideration of subgrid-scale heterogeneity that is inherent to ecological processes is one of the areas for the improvement of global-scale predictions in plant production, biomass, and biogeochemical cycling [Moorcroft, 2006].

[18] The improvement in prediction due to the treatment of local edaphic variations was prominent in North America, but the change was minor in northern Eurasia. As seen in the productivity potential map (Figure 1), the high-resolution GLC2000 cover types of North America exhibited a gradual gradient in edaphic heterogeneity. However, in northern Eurasia, most of the boreal region was classified as high productivity (i.e., closed forests), and the weight for high-productivity simulation (the default SEIB-DGVM) was close to unity. Although GLC2000 has global coverage, the analysis and classification varies from region to region. In this study, the classification scheme used for North America was particularly suited for our purpose.

[19] Although the GLC2000 dataset is a snapshot of land cover in the year 2000, we assume that the vegetation cover in the dataset reflects the underlying edaphic heterogeneity. This is an important assumption in temporal variability on the subgrid scale. For example, if a particular pixel experienced a major disturbance just prior to 2000, the pixel is likely to be classified as low productivity, and its actual suitability for plant production becomes irrelevant. Such temporal factors cannot be ignored if the high-resolution edaphic variation is investigated in a spatially explicit manner. However, in this study, our focus was to estimate only the fractional coverage of edaphic types in large DGVM grids. Due to the significant scale difference between GLC2000 (∼1 km) and SEIB-DGVM (1°), the larger cell contains ∼10,000 smaller cells, and the overall fraction is likely to converge into a realistic value. Moreover, compared to other biotic indices such as biomass and basal area, LAI tends to attain mature, equilibrial values relatively quickly after disturbances. Thus, we believe that the bias should remain as an insignificant underestimation in productivity. As the next step, the mean stand age and/or mean disturbance interval of a grid, together with the estimate of the timing of LAI maturation, can be used as adjustment factors for this underestimation.


[20] T. I. thanks R. Suzuki and H. Kobayashi for instructive discussions on the usage of remote sensing products. This study was funded by Innovative Program of Climate Change Projection for the 21st Century of the Ministry of Education, Culture, Sports, Science and Technology (MEXT).