In ashiklom/fortebaseline: Forest Resilience and Threshold Experiment (FoRTE) ED2 baseline runs source code and manuscript

Journal ideas

• Global Change Biology -- 8000 word limit may be OK, especially if many methodological details are punted to the appendices.
• Ecological Monographs -- To do this topic justice may require a lot of words and figures. It may be hard to squeeze all of this information into a regular research paper.

Introduction

Earth's climate is changing, and its terrestrial ecosystems are likely to be affected (REF: IPCC). Many of these ecosystems will be affected through persistent changes in their environment, for instance through elevated atmospheric CO~2~ concentrations and temperature, changing precipitation regimes (REF). At the same time, climate change is likely change the frequency and severity of disturbances [@dale_2001_climate], and the impacts of these changes may be more acute.

In trying to anticipate future changes to terrestrial ecosystems, ecologists have four general techniques. One approach is to look to the past to see how ecosystems responded to similar changes (REFs -- e.g. PALEON?). This is challenging for two reasons: For one, we often have to look very far back to find similar climate changes, and the techniques we have for determining the state of ecosystems in the distant past have limited detail (REF). More importantly, there are few (if any) times where the climate has changed as rapidly as it is now, and it is unclear whether slower changes in the past will trigger responses similar to the faster changes happening now (REF). A second approach is space-for-time substitution, whereby we use existing climatic gradients through space (e.g. with latitude or elevation) to infer the corresponding changes through time (REF). However, many processes that drive spatial variation in ecosystems play out over hundreds or thousands of years (e.g. evolution), far longer than policy-relevant timescales of years to decades and far slower than the current pace of climate change (REF). A third approach is to conduct experimental manipulations, where individuals are subjected to controlled changes in environmental conditions [@medlyn_2015_using]. Because they are able to isolate specific changes by controlling for most other factors, experiments are very useful for testing mechanisms. Unfortunately, resource constraints limit the spatial and temporal extent of experiments, and emerging evidence from (the relatively few) long-term experiments suggests that short-term results cannot always be extrapolated from longer-term results [@reich_2018_unexpected].

The fourth approach is to use computer models based on existing understanding of ecosystem function. Models unconstrained by data are little more than thought experiments with limited value to decision-making. However, models provide a useful tool for confronting hypotheses with observations from a variety of sources, including the three described above [@dietze_2013_improving].

Model design involves an invariable trade-off between complexity and predictive uncertainty. The more processes are represented in a model (and the more parameters are used to represent those processes), the higher the uncertainty in the "true" value of each parameter, and the higher the uncertainty in model predictions (especially out of sample) (REF -- information theory, e.g. AIC?). On the other hand, simplistic models often fail to match important properties of the observations. A key challenge in model development is to find the balance between these two extremes.

Big leaf models (e.g. CLM) fail to capture vegetation dynamics (REF). This is due to their inability to represent demographic processes, which are important (REF). An emerging class of models can simulate vegetation demographic processes, including competition for resources within a patch, recruitment, and mortality [@fisher_2017_vegetation]. Individually-based models are one possibility (REFs -- Shugart, Ensheng Wang, etc.). However, they are complicated (REF). An alternative are models of intermediate complexity ("middle ground"), which use approximations. One such approximation is the "Ecosystem Demography" approach, whereby groups of individuals ("cohorts") of similar species, size, and structure are modeled together, and the statistical properties of these cohorts are modeled rather than the characteristics of each individual [@moorcroft_2001_method].

One realization of this is the Ecosystem Demography (v2) model [@moorcroft_2001_method; @medvigy_2011_predicting; @medvigy_2009_mechanistic]. This model has been used to simulate (examples...). Previous uncertainty analyses with ED: [@dietze_2014_quantitative]... More recently, [@raczka_2018_what]...

[@rollinson_2017_emergent] -- ED2 paleoclimate [@miller_2015_alteration] -- ED2 at Duke FACE [@dekauwe_2013_forest] -- ED2 at Duke and Oak Ridge FACE

[@longo_2019_ed1] provide a description of the most recent version of ED (2.2), and [@longo_2019_ed2] provide an evaluation. However, these studies only consider the default configuration of ED2. In fact, ED2 provides several alternative structures for some of its submodels.

Light availability is a first order control on photosynthesis, and competition for light is a key process in driving competitive dynamics and community succession [@bazzaz_1979_physiological]. How light is distributed through a canopy is therefore critically important to accurately simulating community dynamics. Radiative transfer models are used to simulate light proliferation through a canopy. Detailed, 3D representations [@widlowski_2014_abstract; @widlowski_2007_third; @malenovsky_2008_influence] have prohibitively high computational costs for use in dynamic vegetation models. Therefore, these models have had to rely on computationally-efficient approximations. One common approach is the "two-stream approximation", originally developed for atmospheric radiative transfer [@meador_1980_two] and later adapted for plant canopies [@dickinson_1983_land; @sellers_1985_canopy]. Although popular, the two-stream approach has some important limitations (REFS, details...). An alternative approach is the "multiple scattering" model of @zhao_2005_multiple. Accounting for multiple scattering enhances canopy light absorption, particularly in near-infrared wavelengths [@zhao_2005_multiple], therefore reducing the amount of light that makes it into the understory (TODO confirm).

Competition is also strongly determined by canopy structure and how it is represented (REF). In other words, how much does canopy height matter compared to gappiness and horizontal competition? C.f. [@hogan_2018_fast; @dietze_2008_changing].

Trait plasticity with light level also plays a role -- i.e. shaded leaves adapt to lower light levels. C.f. [@keenan_2016_global; @lloyd_2010_optimisation; @niinemets_2016_within].

This manuscript addresses the following research questions: (Q1) To what extent can the ED2 model reproduce 100 years of observed succession (or observed anything) at UMBS starting from bare ground? (Q2) How does the inclusion of various processes affect the accuracy and predictive uncertainty of ED2 predictions? (Q3) What drives the predictive uncertainty of the ED2 model at UMBS, and what additional observations are necessary to improve predictions / constrain uncertainty?

Based on previous results [@fisher_2015_taking], we predict ED2 will over-predict dominance of early-successional canopy species (and under-predict understory productivity) under its default configuration. Enabling dynamic N cycling and a finite canopy radius model will alleviate these issues, but will significantly increase the parameteric uncertainty due to strong sensitivity to N-cycling and canopy gap fraction parameters, respectively. Constraining these parameters should be a high priority...

Methods

Photosynthesis / respiration ratio (i.e. C balance) in light and shade can be used as an indication of succession -- expect early successional species to have higher ratio in light and late successional species to have higher ratio in shade [@bazzaz_1979_physiological].

Discussion

Comparison against other studies

In two tropical sites, ED2 did a good job simulating observed top-of-canopy outgoing shortwave radiation, as well as radiation profiles [@longo_2019_ed2]. Completely closed canopy (i.e. no finite canopy radius) did not lead to an excessively dark understory; in fact, understory was slightly brighter than observed.

Future directions: Disturbance response modeling

Net ecosystem productivity (NEP) declines immediately after disturbance, but recovers in 10-20 years [@amiro_2010_ecosystem]. Extent of recovery depends on balance of gross (aboveground?) primary productivity (GPP) and respiration; GPP increase during recovery, while respiration may stay constant [@amiro_2010_ecosystem].

Old discussion notes

• Parameter uncertainty is the dominant driver of model predictive uncertainty.

• ...but the nature of this parameter uncertainty depends on model configuration.

• Ecological implications of structural uncertainty

  • Predictions about ecological succession are very sensitive to the understory light environment.
    • Need a good representation of canopy structure (finite canopy radius)
    • How you model canopy radiative transfer matters! Small differences in the model can give significant results.
  • But, trait plasticity can provide an alternative mechanism for success in the understory.
    • Important modeling implication for all the emerging research on intra-specific trait variability.
    • This combination of discrete functional types with plastic traits is closer to reality than purely trait-based ("eco-evolutionary?") models or completely discretized models with static traits on the other.
    • Important caveat: Current implementation is based entirely on tropical data, which has more extreme canopy shading than temperate forests (REF), so tropical species may have evolved stronger trait plasticity than temperate species (REF).
    • Perhaps note "ecological equifinality" here: there are different ways to arrive at similar outcomes in the model; this is also true in nature?
• Implications of uncertainty analysis for future data collection and model development
  • Looking beyond aggregate variables (e.g. net primary productivity, total LAI) is essential to model evaluation and calibration. It's possible to generate a relatively?
  • (Expected) Subsurface parameters (e.g. leaf-root ratios) have high sensitivity, but are poorly constrained.
  • Trait covariance may alleviate some data limitation (Shiklomanov et al. 2018 [preprint])
• Comparisons against other studies
  • Parameter sensitivity
    • [@pandit_2018_optimizing] -- Most important parameters for shrubs were SLA, Vcmax, stomatal slope, root:leaf ratio, and root turonver rate (slightly less important: leaf turnover rate, growth respiration, cuticular conductance; much less important: water conductance; basically irrelevant: storage turnover, leaf width)
    • Compare against @dietze_2014_quantitative and @raczka_2018_what
• Other relevant references:
• @gough_2019_high -- Canopy structural complexity is the strongest predictor of productivity across NEON sites. Possible future direction?

ashiklom/fortebaseline documentation built on May 9, 2020, 1:56 a.m.