A dynamic global vegetation model (DGVM) is applied in a probabilistic
framework and benchmarking system to constrain uncertain model parameters by
observations and to quantify carbon emissions from land-use and land-cover
change (LULCC). Processes featured in DGVMs include parameters which are
prone to substantial uncertainty. To cope with these uncertainties Latin
hypercube sampling (LHS) is used to create a 1000-member perturbed parameter
ensemble, which is then evaluated with a diverse set of global and
spatiotemporally resolved observational constraints. We discuss the
performance of the constrained ensemble and use it to formulate a new
best-guess version of the model (LPX-Bern v1.4). The observationally
constrained ensemble is used to investigate historical emissions due to LULCC
(
Due to constraining atmospheric CO
Amongst others, dynamic global vegetation models (DGVMs) are used to quantify
the contribution of LULCC to the terrestrial carbon budget
The assimilation of observations should be an integral part of model
development. Various approaches to incorporate constraining data exist, such
as variational approaches minimizing a cost function using the adjoint of the
model
While the land–atmosphere carbon flux can to some extent be constrained by
the other components of the global carbon cycle, the contribution of LULCC
and, in turn, the implied residual terrestrial carbon sink are highly
uncertain. Efforts to fill this knowledge gap have been made using
bookkeeping approaches
The Land surface Processes and eXchanges (LPX-Bern) model
Two different configurations are used to treat the transition between
different classes of land use. The simpler implementation adjusts the
fractional land-use cover at the end of each year such that the prescribed
area fractions are achieved; this computationally efficient configuration is
referred to as net land use. The more advanced gross land-use implementation
also includes effects of shifting cultivation and wood management by
prescribing all the transitions between different land-use classes and
harvested wood
The model is run on a 1
The model parameter space is sampled using Latin hypercube sampling
Description of sampling parameters with values for LPX v1.2 and the new best-guess version 1.4. If not otherwise indicated parameters are unitless.
Table
The prior distribution used for LHS was derived in multiple steps following
partly an explorative approach. An initial version of the ensemble with 1000
members was run using the 10 LPX parameters and distribution used by
Kernel density estimations of the prior probability distribution
(blue) and the posterior probability distribution (red) of the constrained
ensemble with net land use (
The performance of the final 1000-member model ensemble is evaluated using
the set of observational constraints listed in Table
The MSE
The so-called posterior distribution of a parameter or quantity of interest is obtained by using the skill score weighted normalized histogram, which can be interpreted as a probability density function, of the constrained ensemble. The skill weighted median and confidence interval of a given quantity is then determined by transforming the histogram to a discrete cumulative density function using a cumulative sum and approximating the desired quantiles by a first-order interpolation. Throughout this paper we report the skill weighted median of numerical results along with the 5 and 95 % quantiles, corresponding to the 90 % confidence interval, in parentheses.
Hierarchical weighting scheme to aggregate the relative mean squared error of individual observational constraints to a total error which is then mapped to a total skill score.
Observations used to constrain the model ensemble.
The calculation of the MSE
The modeled total and soil carbon distribution between 1982 and 2005 are
compared to a data set based on observations
For site level observed NPP (multi-biome NPP;
The TM2
The modeled mean annual evapotranspiration between 1989 and 2005 was compared to
the LandFlux-EVAL evapotranspiration data product
The global terrestrial carbon flux is constrained by a deconvolution, for
which the global atmospheric CO
The estimates of global soil and vegetation carbon as given by IPCC
To quantify emissions from LULCC, a second simulation featuring a
time-invariant pre-industrial land-cover distribution and nitrogen
fertilization is run for every ensemble member. In accordance with the TRENDY
model intercomparison (
For each of the parameter sets four transient simulations over the industrial
period are performed: (i) a simulation with prescribed net transitions
(
For the
The use of the ensemble framework allows us to quantify both the magnitude
and the uncertainty of land-use emissions in a model due to parameter spread.
Following the procedure outlined in the method section,
Skill weighted median net biome production (NBP)
Global aggregates of skill weighted median NBP,
Comparison of the skill weighted median emissions due to
land-use change in the two constrained LPX parameter ensembles (90 %
confidence intervals in brackets) to the bookkeeping method and DGVM model
ensemble of
In addition to the standard model configuration a second ensemble of a model
configuration
A third model configuration
Skill weighted median annual net biome production
(NBP)
In the following the ensemble version with gross land use and skill scores
from the net land-use ensemble
The land–atmosphere fluxes show large regional differences
(Fig.
The
Overview of 10 countries with the highest overall contribution to
emissions due to land-use change from 1901 to 2016 in the model ensemble
By using the Natural Earth Data administrative borders the
A mapping of the MSE
In this section, the performance of the net land-use ensemble members
(
As an illustration of the observational constraints, we consider the seasonal
cycle of atmospheric CO
For the scalar targets, the median values and range of the full and
constrained ensemble are compared in Fig.
Vegetation carbon inventory and spatial distribution are highly relevant for
We compare the total land–atmosphere exchange flux to the results of the
atmospheric CO
We investigate the dependency of the constrained ensemble on the choice of
the observational constraints by reevaluating the ensemble for a subset of
observations. We created 19 weighting schemes, each missing one of the
individual observational constraints (Fig.
The value and uncertainty of the scalar targets (red) compared to an
unweighted histogram of the full (blue) and constrained (green) ensemble
The skill weighted median
The unweighted kernel density estimates of the prior (full ensemble) and
posterior (constrained ensemble) parameter distributions are shown in
Fig.
Cumulative net biome production (NBP) of the unconstrained (blue)
and constrained (green) ensemble with 90 % confidence interval shaded;
LPX version 1.2 (orange) and the new reference model version 1.4 (red). The
result of a so-called “single deconvolution” is shown by the black line and
grey range. In this deconvolution, the change in the land inventory is
inferred from the records of atmospheric CO
We use the constrained ensemble to establish a new reference model version, featuring a set of optimized parameters. The reference version will be used for model simulations where the use of an ensemble is not appropriate or required.
The skill weighted median parameter values of the constrained ensemble are
used as a reference model and its parameter values are shown in
Table
Overall, the updated parametrization shows a well-balanced performance in the
spatial benchmarks shown in Fig.
The choice of using the skill weighted median parameters of the constrained ensemble instead of simply using the best-performing parameter set for the reference version is motivated by its robustness and representativeness of the ensemble. While the best-performing model member certainly possesses a higher skill score, its parameter values can depend strongly on the choice and weighting of the observational targets, whereas the median parameter values depend less on individual metrics.
The simultaneous assimilation of multiple observational constraints yields
soil and vegetation stocks and distributions which are consistent with
observations. The total land–atmosphere carbon flux is reproduced relatively
well in the model configuration using net land-use
The observed uncertainties of
We investigated the magnitude and spatial distribution of
A good correspondence between simulated fluxes and the estimates of
A recent study by
Overall, the ensemble approach produces
A hierarchical weighting scheme to compare a diverse set of constraints was
employed, following earlier work
In addition to the weighting of model results with the global skill score, we
employed a minimum skill criterion, discarding runs with very bad performance
in a singular metric. This approach is somewhat comparable to pre-calibration
methods, where implausible parameter spaces are also ruled out
While the uptake of carbon by the terrestrial biosphere in the model ensemble
is significantly larger than earlier versions of LPX, it is still in the
lower range of estimates. A direct way of increasing the magnitude of change
in land carbon is to change pool sizes, which is here restricted by other
observational constraints. The inclusion of more processes, such as natural
and human-induced erosion
Fossil carbon emissions and thus the net biome production and the carbon sink
inferred from the deconvolution may be biased high for the most recent
decades. The fossil emissions are estimated from fossil-fuel production data,
which include the fraction used for non-combustion purposes such as the
production of plastics and asphalt.
The release of both spatially and temporally resolved carbon flux
observations by using remote sensing, such as the Carbon Monitoring System
Flux Pilot (CMS) project, featuring not only net fluxes but also gross
production and respiration, which is a very promising candidate for constraining
the parameter space further. The spatial structure might restrict the
apparent degree of freedom in partitioning the terrestrial sink in
Another avenue of increasing model performance is to introduce spatially
explicit parametrization, as used in multi-model averaging studies
The simultaneous assimilation of multiple observational constraints allowed us to formulate a well-rounded best-guess version of the model. While this parameter version does not necessarily excel at every single benchmark, it shows a consistent performance amongst all different targets. This behavior leads us to believe that the best-guess version is well suited for simulations spanning long time spans, both for paleo- and future research questions, where the use of a full parameter ensemble is not feasible. Furthermore, it can also be used in model intercomparison studies, where single realizations of different models are compared.
We successfully applied a multi-purpose model benchmark to a perturbed
parameter ensemble of a dynamic global vegetation model (DGVM). Specifically,
we developed a best-guess model version and constrained the residual
carbon sink flux and carbon emissions from anthropogenic land use
(
A new reference version of the LPX-Bern (v1.4) DGVM was established. We were
able to show that the constrained ensemble, as well as a resulting best-guess
version, performs consistently well under a range of benchmarks
(Table
Many previous studies have investigated inherent uncertainties in
The observation-constrained DGVM ensemble and best-guess version established
in this work are ready for use in model intercomparison studies
(
Model output is available upon request to the corresponding author (lienert@climate.unibe.ch).
The authors declare that they have no conflict of interest.
This article is part of the special issue “The 10th International Carbon Dioxide Conference (ICDC10) and the 19th WMO/IAEA Meeting on Carbon Dioxide, other Greenhouse Gases and Related Measurement Techniques (GGMT-2017) (AMT/ACP/BG/CP/ESD inter-journal SI)”. It is a result of the 10th International Carbon Dioxide Conference, Interlaken, Switzerland, 21–25 August 2017.
We thank Gianna Battaglia for supplying the Bern3D model output and Marko Scholze for providing the TM2 transport matrices. We would like to thank the data community for their efforts in providing high-quality data sets. This work was supported by the Swiss National Science Foundation (#200020_172476). Edited by: Christoph Heinze Reviewed by: Sönke Zaehle and Jean-François Exbrayat