TL;DR
This study uses field-based tree mortality data to constrain projections of future forest carbon sinks from global vegetation models. The findings suggest that these models often overestimate tree mortality, leading to an overestimation of tropical forest carbon uptake and an underestimation in North America.
What Problem Does This Paper Address?
Accurately projecting the future capacity of forests to sequester atmospheric CO2 is crucial for understanding and mitigating climate change. However, significant uncertainties and debates exist regarding how forest carbon sinks will respond to future climate conditions. Dynamic Global Vegetation Models (DGVMs) are key tools for these projections, but they often struggle to accurately represent complex forest processes like tree mortality. This can lead to biased estimates of net primary productivity (NPP) and heterotrophic respiration (HR), ultimately affecting the projected net carbon sink. This paper aims to reduce these uncertainties by using empirical, field-based data on tree mortality to constrain DGVM outputs.
Methods
The researchers employed an "emergent constraint" (EC) approach, which leverages observed relationships between variables to improve model projections. Specifically, they:
- Generated a spatially explicit map of tree mortality (LOSS): They used a large dataset of field-based tree mortality measurements from unmanaged forest plots and a Random Forest machine learning algorithm to upscale these observations, incorporating 57 environmental covariates.
- Compared DGVMs to observations: They evaluated how well six different DGVMs simulated historical LOSS, finding that most models overestimated it, particularly in tropical regions and eastern North America.
- Constrained DGVM projections: They used two main EC approaches:
- Conventional EC: This involved establishing a statistical (linear) relationship between historical LOSS (from plots or the upscaled map) and projected NPP and HR from the DGVM ensemble.
- Machine Learning (ML) EC: This approach used a Random Forest model to capture potentially non-linear relationships between historical LOSS and projected NPP/HR at a grid-cell level. This method was favored for its ability to incorporate spatial patterns and non-linearities.
Datasets
- Forest Plot Data: A dataset of 2676 largely unmanaged forest plots with long-term observations (1951-2018) was used to derive tree mortality (LOSS) estimates. These plots were selected based on criteria ensuring they represented natural, mature forests in a quasi-steady state.
- Environmental Covariates: 57 variables including climate (temperature, precipitation, aridity), vegetation properties (tree density, basal area), and soil conditions were used for geospatial modeling.
- Dynamic Global Vegetation Model (DGVM) Outputs: Data from six DGVMs (ORCHIDEE, CABLE-POP, JULES, LPJ-GUESS, LPJmL, and SEIB-DGVM) were used for projected NPP and HR (2015-2099) and historical mortality flux (1961-2014).
Key Insights and Ideas
- Tree Mortality is a Critical, Underestimated Factor: The study highlights that tree mortality is a significant driver of forest carbon dynamics, and its accurate representation in models is crucial.
- DGVMs Overestimate Tree Mortality: A key finding is that most DGVMs tend to overestimate historical tree mortality, especially in tropical forests. This suggests a potential oversimplification of mortality processes in these models.
- Field Data Can Effectively Constrain Models: The research demonstrates the power of using empirical, field-based demographic data to constrain complex ecosystem models, leading to more robust projections.
- Tropical Forest Carbon Sink May Be Overestimated: By constraining DGVMs with observed mortality, the study projects a weaker future carbon sink in tropical regions than previously estimated by unconstrained models.
- North American Carbon Sink May Be Underestimated: Conversely, the constrained models suggest a potentially larger carbon sink in North America.
- Spatial Scale Matters: The effectiveness of the EC approach varied with spatial scale, with the ML approach proving more adept at capturing fine-scale relationships and generating spatially explicit constraints.
Key Results
- DGVMs Overestimated LOSS: Most DGVMs overestimated historical LOSS by as much as 5 Mg ha⁻¹ y⁻¹ in tropical regions and eastern North America.
- Constrained Tropical Sink Reduction: The projected tropical forest carbon sink by the end of the century was reduced from 2 to 1.4 PgC y⁻¹ in South America, and similarly for Africa and Asia & Australia.
- Constrained North American Sink Increase: The North American forest carbon sink was projected to increase from 0.8 to 1.1 PgC y⁻¹ after applying the constraint.
- Reduced Model Uncertainty: The ML constraint approach significantly reduced the spread of projected NPP and HR uncertainties, particularly in tropical regions.
Limitations and Future Work
- Model Simplifications: The authors acknowledge that DGVMs often use simplified representations of tree mortality (e.g., a fraction of standing stock) and may not fully capture complex processes like competition or climate-induced disturbances.
- Resource Use Efficiency: The study notes that changes in resource availability (water, nutrients) and their usage efficiency could alter the couplings between growth, mortality, and respiration, which were not fully accounted for.
- CO2 Fertilization: The approach did not explicitly account for the effects of atmospheric CO2 concentrations on carbon fluxes, which is a known driver of forest productivity.
- Future Work: The authors suggest that future DGVM development should incorporate more detailed demographic processes and competition dynamics. They also highlight the importance of considering spatial trade-offs in ML constraint approaches and the potential for climate-induced disturbances to further influence carbon sinks.
Summary updated