A stand-alone tree demography and landscape structure module for Earth system models

Authors


Abstract

[1] We propose and demonstrate a new approach for the simulation of woody ecosystem stand dynamics, demography, and disturbance-mediated heterogeneity suitable for continental to global applications and designed for coupling to the terrestrial ecosystem component of any earth system model. The approach is encoded in a model called Populations-Order-Physiology (POP). We demonstrate the behavior and performance of POP coupled to the Community Atmosphere Biosphere Land Exchange model (CABLE) applied along the Northern Australian Tropical Transect, featuring gradients in rainfall and fire disturbance. The model is able to simultaneously reproduce observation-based estimates of key functional and structural variables along the transect, namely gross primary production, tree foliage projective cover, basal area, and maximum tree height. Prospects for the use of POP to address current vegetation dynamic deficiencies in earth system modeling are discussed.

1 Introduction

[2] Changes in biomass storage and structure of forest and savanna ecosystems are a significant driver of the current terrestrial carbon sink which removes around one quarter of all anthropogenic carbon dioxide emissions [e.g., Ahlström et al., 2012]. Dynamic Vegetation Models (DVMs) incorporated in Earth System Models (ESMs) attempt to describe changes in vegetation biomass compartments over time as the resultant of net primary production (NPP) and biomass turnover, the latter an outcome of phenological cycles of foliage and roots, mortality of plant individuals, and disturbances such as wildfires and storms.

[3] Current “first generation” DVMs, suitable for coupling to ESMs, employ large-area parameterizations of NPP allocation and biomass turnover designed for application on grid cells 10s to 100s of kms in size. Being computationally efficient, they reduce the combined effects of underlying population and community processes such as recruitment, mortality, and competition between individuals and species for limiting resources to a small number of static parameters or environmental dependencies [e.g., Sitch et al., 2003]. The generality of these dependencies across vegetation zones and robustness under a rapidly changing climate are often questionable [e.g., Fisher et al., 2010]. Further, the lack of mechanistic detail also means that such models cannot be directly calibrated or evaluated using forest inventory data on stand structure and development [Wolf et al., 2011]. Different DVMs simulate divergent time evolution of biomass pools, especially under future climate projections [e.g., Sitch et al., 2008], with some retaining a net biomass sink over the coming century, and others simulating a source or reduced sink by late 21st century [Ahlström et al., 2012]. In ESM simulations with an active carbon cycle feedback to climate, such differences translate into divergence in the simulated global climate [Friedlingstein, 2006]. Indeed, forest dynamics and its representation in global models may be one of the greatest sources of uncertainty in predicting the future climate [Purves and Pacala, 2008].

[4] Performance shortcomings of first-generation DVMs have been one motivation for the development of a second generation of DVMs that more explicitly represent individual and population-level processes governing ecosystem demography and competition. Such models adopt some generalization of the “forest gap” approach based on recruitment, mortality, and resource competition amongst trees and plant functional types (PFTs) co-occurring in patches [e.g., Moorcroft et al., 2001; Sato et al., 2007; Smith et al., 2001] or upscaling to grid cells, stochastic representations may be adopted for demographic processes, including disturbances, leading to high computational demands, and complicating the analysis of results because they are not strictly deterministic. In addition, it is technically challenging to couple these models with the existing land surface models (within ESMs) because of their intricate internal representation of stand structure and its integration with plant physiological processes such as carbon assimilation, allocation, and phenology.

[5] In this paper we propose and demonstrate a new approach for the simulation of woody ecosystem stand dynamics, demography, and disturbance-mediated heterogeneity suitable for continental to global applications and designed to be readily coupled to the terrestrial ecosystem component of any ESM. We demonstrate the behavior and performance of the model when coupled to a biogeochemical land surface model and applied along a rainfall and fire disturbance gradient in northern Australia.

2 Methods

2.1 Study Region

[6] The North Australian Tropical Transect (NATT) (see e.g., [Hutley et al., 2011]) is a 1000 km transect that shows a systematic decline in mean annual precipitation (MAP) with distance (~ 1 mm per km) from the northern coast of the Northern Territory, Australia (Figure 1i). Here we represent the gradients in rainfall and fire regime of the NATT transect by selecting 1000 random 0.05° × 0.05° grid cells from an area bounded by 19.95°S, 11.4°S, 130.0°E, 134.5°E. The NATT is characterized by largely intact savanna and arid vegetation, with tree leaf area index (LAI) declining from about 0.75 at the northern end of the transect to 0.25 at the southern end. In the north of the region (MAP > 600 mm), the dominant vegetation is tropical savanna (overstorey of evergreen Eucalyptus and Corymbia tree species, and an understorey dominated by C4 grasses), while Acacia woodlands and shrublands and hummock grasslands become increasingly prominent as MAP declines [Hutley et al., 2011]. The vegetation is subjected to fire regularly (once every 3 to 7 years, Figure 1ii, data derived from Craig et al. [2002]). The fire frequency peaks at latitudes of −13 to −14°, declining with decreasing latitude due to decreasing fuel loads. The fraction of early dry-season (pre-August) fires follows a similar latitudinal pattern to the fire frequency, which is an effect of fire management. Fire timing is a predictor of fire intensity, with late-season fires generally being significantly more intense as fuels cure and weather becomes more extreme.

Figure 1.

Latitudinal variation of mean annual precipitation, tree leaf area index (LAI), fire frequency, and fraction of fires occurring in the early part of the dry season (pre-August). Each point represents a spatial average across ~65 points lying within a latitude bin of width 0.57°, with error bars representing one standard deviation.

2.1.1 The POP Model

[7] The proposed tree demography and landscape structure model is called Populations-Order-Physiology (POP, a palindrome whose meaning is also valid when read from right to left). POP is designed to be modular, deterministic, computationally efficient, and based on defensible ecological principles. Parameterizations of tree growth and allometry, recruitment, and mortality are broadly based on the approach of the LPJ-GUESS DVM [Smith et al., 2001].

[8] Input variables are stem biomass increment and mean return times for two classes of disturbance: (i) “catastrophic” disturbance, which kills all individuals (cohorts) and removes all biomass in a given patch, (ii) “partial” disturbances, in the present study representing fire, which result in the loss of a size-dependent fraction of individuals and biomass, preferentially affecting smaller (younger) cohorts. Intensity of partial disturbance is also an input for the present study.

[9] State variables include the density and biomass of tree stems partitioned among age/size classes (cohorts) of trees and representative neighborhoods (patches) of different age-since-last-disturbance across a simulated landscape or grid cell. Output variables such as tree height and diameter distribution, basal area, stem density are derived from these state variables. Stem biomass turnover due to mortality and disturbances is an additional output (Figure 2). The time step is 1 year.

Figure 2.

Input and output variables for each of the CABLE and POP models, including variables which are exchanged between the models when they are coupled.

[10] POP simulates allometric growth of cohorts of trees that compete for light and soil resources within a patch. A new cohort is created each year. The annual stem biomass increment, in this study a biome-specific fixed proportion of net primary production (NPP), is partitioned among cohorts according to a power function of their current aggregate stem biomass (size), on the assumption that larger individuals preempt resources owing to a larger surface area and exploration volume of their resource uptake surfaces (leaves and fine roots), and due to the advantage conferred on taller individuals by the shading of shorter ones in crowded stands [Westoby, 1984]. A cohort's share of the total annual biomass increment is divided equally among individuals and results in height and stem diameter growth following a prescribed relationship between stem height and diameter, in this study also affected by rainfall (Appendix 2).

[11] Recruitment and mortality govern the population dynamics of each patch. A new cohort is born each year with a stem density that declines rapidly as biomass accumulates and crowding increases, depleting the resource supply for establishment of new individuals. Mortality is affected as a proportional annual reduction in cohort density, arising primarily from resource limitations which reduce vigor and interfere with defense and other stress response mechanisms, leading to an enhanced risk of death. Resource limitation may arise due to abiotic factors such as drought or biotic factors such as shading and other forms of resource preemption by neighbors. Growth efficiency, characterized by the annual biomass increment for a cohort normalized by current biomass, is used as a proxy for resource availability (the inverse of resource limitation). Mortality is assumed to rapidly increase as current growth efficiency falls below a prescribed threshold, which is a calibration parameter in the model.

[12] Replicate patches representing stands of differing age since-last-disturbance are simulated for each grid cell. It is assumed that each grid cell is large enough to accommodate a landscape in which the frequency of patches of different ages follows a negative exponential distribution with an expectation related to the current disturbance interval. This assumption is valid if grid cells are large relative to the average area affected by a single disturbance event and disturbances are a Poisson process, occurring randomly with the same expectation at any point across the landscape, independent of previous disturbance events. To account for disturbances and the resulting landscape structure, state variables of patches of different ages are averaged, and weighted by probability intervals from the negative exponential distribution. The resultant weighted average of, for example, total stem biomass or annual stem biomass turnover, is taken to be representative for the grid cell as a whole.

[13] A detailed description of the model is provided in Appendix 1.

2.1.2 Integration With CABLE: CABLE-POP

[14] Figure 2 indicates key input and output variables for each of the Community Atmosphere Biosphere Land Exchange (CABLE) and POP models, including variables that are exchanged between the models when they are coupled.

[15] POP was coupled to the CABLE land-surface scheme [Wang et al., 2011], as implemented in the BIOS2 modeling environment and described in Appendix 4, and in further detail by Haverd et al. [2013b].

[16] Coupling between CABLE and POP is achieved by exchange of variables as illustrated in Figure 2. Primarily, CABLE supplies annual stem biomass increment to POP and POP returns an annual stem biomass turnover to CABLE. To convert between stem biomass (POP) and tree biomass (CABLE), we assume a ratio of 0.6, based on values tabulated by Berry et al. [2010]. The resulting tree biomass turnover is applied as an annual decrease in the CABLE tree biomass pool, and replaces the default fixed biomass turnover rate. CABLE also pipes the tree LAI to POP, as an input for the foliage projective cover diagnostic. Annual fine structural litter and grass leaf carbon pools are used as inputs to the fire intensity variable (not shown).

[17] Additional key input variables to POP (Figure 2) are the return times for catastrophic and partial disturbances. For the former, we assume a fixed value of 200 years, corresponding for example to a cyclone return interval [Cook and Goyens, 2008]. For partial disturbance, we adopt return times based on mean fire frequencies as retrieved from burned area data and fire intensities (required to evaluate size-dependent mortality due to partial fire disturbance), as described in Appendix 3. Annual rainfall is also an input, required in this study for the NATT-specific rainfall-dependent tree height-diameter relation (see Appendix 2).

3 Results

[18] Figure 3 illustrates the model's ability to simultaneously predict vegetation function and structure. Variation of key modeled functional and structural variables with mean annual precipitation compares well with observation-based estimates. Observation-based estimates of gross primary production (GPP) were derived by Kanniah et al. [2011], using a model combining field-based light use efficiency, meteorology, and remotely sensed FPAR, which was evaluated against flux data. Observation-based tree height, basal area, and projected foliage cover were calculated from predictive empirical models developed by Williams et al. [1996], which describe the decline of these variables with rainfall, based on a data set of ~1000 quadrats (each 20 m × 20 m) lying north of 18°S within the Northern Territory.

Figure 3.

Time-averaged (1991–2011) CABLE-POP output variables: variation with rainfall and comparison with observation-based estimates. (i) Gross primary production (tree and grass components) and comparison with Kanniah et al. [2011]; (ii) tree foliage projected cover and comparison with Williams et al. [1996]; (iii) tree basal area and comparison with Williams et al. [1996]; (iv) maximum tree height and comparison with Williams et al. [1996].

[19] Each modeled variable is the outcome of multiple responses:

  1. [20] Gross primary production (combined tree and grass components, Figure 3i) is attributable the responses of GPP to LAI and soil moisture deficit.

  2. [21] Tree foliage projected cover (Figure 3ii) is the outcome of tree LAI and the clumping of tree leaf area into crowns, which depends on number density and size-dependent crown dimensions (Appendix 1).

  3. [22] Basal area (Figure 3iii) is the outcome of number density, tree biomass distribution, and allometry.

  4. [23] Maximum height, equated here with the tree height of the 95th percentile, (Figure 3iv) is the outcome of size-dependent growth, mortality, and allometry, and particularly how these processes affect the upper end of the age-class distribution.

[24] Results were obtained with minimal parameter tuning within POP. All parameters were set to their default literature-based values as described in Appendix 1, except for the stress mortality threshold (GEmin) (equation 1.10 in Appendix 1), which was manually tuned to a value of 0.015 based on model performance against observation-based basal area along the NATT (Figure 3iii).

[25] Figure 4 shows stand development over a 400 year period, illustrating the dynamic behavior of CABLE-POP at extreme ends of the NATT transect. At the low rainfall end, which is characterized by infrequent fire and periods of drought stress, biomass is highly variable at the decadal timescale (Figure 4i), largely due to variation in rainfall (and hence soil moisture and stem increment), which causes fluctuations in the nondisturbance component of mortality (Figure 4v). Centennial-scale cyclicity apparent under low rainfall is due to the repetition of the 100 year forcing time series. In contrast, biomass at the wet end of the transect is less variable (Figure 4ii), with a much higher proportion of biomass loss attributable to disturbance (Figure 4vi). Stronger density dependence (i.e., crowding leading to resource limitation and death of suppressed individuals) results in a sparser stand with fewer, taller individuals in the high rainfall case (Figures 4ii and 4iv) compared with low rainfall (Figures 4i and 4iii).

Figure 4.

Development of tree stands at (left panels) low and (right panels) high rainfall ends of the NATT transect: (i, ii) stem biomass by height class; (iii, iv) stem number density by height class; (v, vi) rate of biomass lost due to disturbance (partial and catastrophic) and nondisturbance-related mortality. All quantities are averaged over the grid cells with mean annual precipitation (1912–2011) in the ranges of (left panels) 357–447 mm yr−1 and (right panels) 1607–1696 mm yr−1. All quantities are time resolved annually, except for disturbance-related biomass loss, which is represented as a 5 year running mean. Meteorological forcing is a fourfold repetition of the 1912–2011 time series.

4 Discussion and Conclusions

[26] Our goal in developing POP was to fill a gap between the large-area approaches of first generation DVMs and the detailed patch-based approaches of most second-generation DVMs, providing a model with sufficient ecological realism to prognose standing biomass, demography, and disturbance-mediated heterogeneity of woody vegetation at grid scale with sufficient accuracy and robustness for continental to global applications. POP is highly modular (Figure 2) with few, well-defined input and output variables, allowing it to be readily coupled to any model that, like CABLE, can provide grid-level LAI and stem biomass production on an annual time step. POP may also be forced with observations, allowing it to be calibrated to biomass or stand demography data independent of bias propagating from any forcing model.

[27] The development of POP represents an advance toward quantifying the impact of disturbances in land surface models at regional to global scales. Of particular interest to the global models is to be able to model fires, which usually are partial disturbance events, with tree mortality depending on size (see e.g., Appendix 3). Therefore, it is important that tree biomass be disaggregated into size classes for the purpose of modeling the impact of fire on ecosystem function and carbon stocks. An example of a prospective application is to the Australian continental carbon budget. In a recent assessment [Haverd et al., 2013a], it was estimated that fires contributed significantly to gross C-CO2 emissions with 104 TgCyr−1, a quantity greater than the territorial fossil fuel emissions for the 1990–2011 period. However, the net impact of wildfire on the Australian carbon budget is currently unknown. The net effect could be estimated using a modeling approach including CABLE-POP, in which biomass pools respond to time-varying fire frequency and intensity, with these variables either being prescribed or prognosed using an internal wildfire module.

Acknowledgments

[28] Vanessa Haverd, Peter Briggs, and Josep Canadell acknowledge the Australian Climate Change Science Program for enabling their contributions to this work. Benjamin Smith acknowledges funding as an Ernest Frohlich Visiting Fellow to the CSIRO Division of Marine and Atmospheric Research, Canberra. This study is a contribution to the Strategic Research Area Modelling the Regional and Global Earth System (MERGE).

[29] The Editor thanks two anonymous reviewers for their assistance in evaluating this paper.

Ancillary

Advertisement