Central to the development of Earth system models (ESMs) has been the coupling of previously separate model types, such as ocean, atmospheric, and vegetation models, to address interactive feedbacks between the system components. A modelling framework which combines a detailed representation of these components, including vegetation and other land surface processes, enables the study of land–atmosphere feedbacks under global climate change.
Here we present the initial steps of coupling LPJ-GUESS, a dynamic global vegetation model, to the atmospheric chemistry-enabled atmosphere–ocean general circulation model EMAC. The LPJ-GUESS framework is based on ecophysiological processes, such as photosynthesis; plant and soil respiration; and ecosystem carbon, nitrogen, and water cycling, and it includes a comparatively detailed individual-based representation of resource competition, plant growth, and vegetation dynamics as well as fire disturbance. Although not enabled here, the model framework also includes a crop and managed-land scheme, a representation of arctic methane and permafrost, and a choice of fire models; and hence it represents many important terrestrial biosphere processes and provides a wide range of prognostic trace-gas emissions from vegetation, soil, and fire.
We evaluated an online one-way-coupled model configuration (with climate variable being passed from EMAC to LPJ-GUESS but no return information flow) by conducting simulations at three spatial resolutions (T42, T63, and T85). These were compared to an expert-derived map of potential natural vegetation and four global gridded data products: tree cover, biomass, canopy height, and gross primary productivity (GPP). We also applied a post hoc land use correction to account for human land use. The simulations give a good description of the global potential natural vegetation distribution, although there are some regional discrepancies. In particular, at the lower spatial resolutions, a combination of low-temperature and low-radiation biases in the growing season of the EMAC climate at high latitudes causes an underestimation of vegetation extent.
Quantification of the agreement with the gridded datasets using the normalised mean error (NME) averaged over all datasets shows that increasing the spatial resolution from T42 to T63 improved the agreement by 10 %, and going from T63 to T85 improved the agreement by a further 4 %. The highest-resolution simulation gave NME scores of 0.63, 0.66, 0.84, and 0.53 for tree cover, biomass, canopy height, and GPP, respectively (after correcting tree cover and biomass for human-caused deforestation which was not present in the simulations). These scores are just 4 % worse on average than an offline LPJ-GUESS simulation using observed climate data and corrected for deforestation by the same method. However, it should be noted that the offline LPJ-GUESS simulation used a higher spatial resolution, which makes the evaluation more rigorous, and that excluding GPP from the datasets (which was anomalously better in the EMAC simulations) gave 10 % worse agreement for the EMAC simulation than the offline simulation. Gross primary productivity was best simulated by the coupled simulations, and canopy height was the worst. Based on this first evaluation, we conclude that the coupled model provides a suitable means to simulate dynamic vegetation processes in EMAC.
Simulation models are at the forefront of Earth systems research. Historically, such models were initially developed to simulate one component of the Earth system in isolation, such as ocean and atmospheric general circulation models (GCMs) or dynamic global vegetation models (DGVMs), with prescribed boundary conditions at the interfaces to other Earth system components. However, the interactions between Earth system components are dynamic, and representations of feedbacks are necessary to assess the functioning and response of the Earth system as a whole. To this end, models have increasingly been coupled to each other to provide dynamic multidirectional fluxes between models, as opposed to prescribing simple non-interacting boundary conditions. This approach has yielded atmosphere–ocean general circulation models (AOGCMs) which are utilised to understand the dynamics of the physical components of the climate system
Interactive carbon cycles and dynamically changing vegetation have been recognised as important processes in the Earth system
To take the first steps towards constructing an ESM with dynamic vegetation, anthropogenic influences, and fire, we have combined an atmospheric chemistry-enabled AOGCM, EMAC See
EMAC (ECHAM/MESSy Atmospheric Chemistry)
By bringing together these two modelling systems, our intent is to produce a fully featured ESM, which benefits from the continuous development of all components. We plan to follow a stepwise model integration roadmap, whereby the coupling between LPJ-GUESS and EMAC is tightened in well-defined consecutive steps, and processes (such as land use) are included or enabled in a consecutive manner. This will allow us to assess the effects of one model on the other and the effects of the inclusion of new processes in a stepwise and logical fashion. For our first step, we have chosen to simulate and evaluate the vegetation produced when LPJ-GUESS is forced by an EMAC-simulated climate, i.e. a one-way coupling without the feedback from the land surface to the atmosphere.
Upon completion of the full model integration process (including bidirectional coupling which is not presented here), the trace-gas emissions from LPJ-GUESS will form key inputs to the atmospheric chemistry representations in EMAC, allowing for bidirectional chemical interactions of the surface with the atmosphere. Then the full model should become a powerful tool for investigating land–atmosphere interactions including the methane cycle and lifetime; the atmospheric chemistry of other reduced carbon species; fire effects and feedbacks; future nitrogen deposition rates and fertilisation scenarios; ozone damage to plants; and the contribution of biogenic volatile organic compounds to aerosol load and, via cloud condensation nuclei activation, to cloud formation (e.g. precipitation cycles).
When evaluating the vegetation produced in the one-way coupled configuration employed here, there are potentially three sources of error that may contribute to data–model mismatch: poorly constrained parameter values and inadequate representation of the processes in LPJ-GUESS, biases in the climate produced by EMAC (which are expected to have some dependency on the spatial resolution; e.g. see
The ECHAM/MESSy Atmospheric Chemistry (EMAC) model is a numerical
chemistry and climate simulation system that includes sub-models
describing tropospheric and middle atmosphere processes and their
interaction with oceans, land, and human influences
See
The following two paragraphs are modified from a standard LPJ-GUESS model description template, which is freely available and copyright free.
Primary production and plant growth follow the approach of LPJ-DGVM
Photosynthesis, respiration, and hydrological processes operate on a daily time step and require daily temperature, precipitation, and incident shortwave radiation. However, monthly climate data may be provided, in which case the model interpolates daily values from the monthly values. In these circumstances, the number of precipitation days in the monthly periods may also be provided to disaggregate total precipitation into distinct rain events. In the case of unmanaged natural vegetation (as simulated here), vegetation dynamics (such as establishment and mortality), disturbance, turnover of plant tissues, turnover between litter pools, and allocation of carbon and nitrogen to plant organs all occur on an annual basis. Simulation of wildfire is included via GlobFIRM (Global FIRe Model)
All stochastic processes in LPJ-GUESS are implemented “semi-stochastically” using a random number generator with a starting seed. This means that for a fixed starting seed, model runs with identical settings produce identical results.
The model coupling strategy employed here was to modify LPJ-GUESS such that it provides its functionality via a new sub-model in the MESSy framework. An important design priority was to maintain the integrity of the LPJ-GUESS source code by performing only minimal modifications and additions in order to facilitate straightforward synchronising with the main LPJ-GUESS trunk version in the future. This approach was successful, with only minor changes made to LPJ-GUESS infrastructure code and no changes to the scientific modules. For more details, see Appendix
To provide appropriate climate forcing for LPJ-GUESS, EMAC calculates the daily-mean 2 m temperature, daily-mean net downwards shortwave radiation, and the total daily precipitation at the end of the simulation day; and it provides it to LPJ-GUESS. This is similar to the approach taken by others when coupling LPJ-GUESS to EC-Earth
In turn, LPJ-GUESS provides fractional vegetation cover, leaf area index, daily net primary productivity and average height of each PFT to EMAC. However, these values are not used by EMAC in the simulations presented here. Thus, we are demonstrating only a one-way coupling where the land surface state does not affect the atmospheric state. The boundary conditions for the atmospheric model (in particular the surface energy and water fluxes) come from the pre-existing land surface representation. For an overview of the processes and feedbacks enabled in the EMAC configuration used here, as well as those to be included in future versions, please see Fig.
The main processes and exchanges in the coupled model framework. Processes and exchanges with normal black text and black solid arrows are included in the framework and used in the simulations presented here; processes and exchanges with normal grey text and grey solid arrows are included in the framework but not used in the simulations presented here; and processes and exchanges with italic grey text and grey dotted arrows are not included in the framework but planned in future work. All exchanges happen on a daily basis, except for soil properties, which happen only during the initialisation phase.
In the coupled model, the vegetation produced by LPJ-GUESS within EMAC will be directly sensitive to biases in the climate produced by EMAC. It is well known that these biases are dependent on spatial resolution; see
The
As the simulations conducted here utilise only a one-way coupling, EMAC uses its standard land surface scheme, which is taken from the ECHAM5 model and is described in detail by
To aid the interpretation of the EMAC simulations, we also performed an “offline” LPJ-GUESS simulation using observed climate data from the CRUNCEP bias-corrected reanalysis dataset
In all model simulations a 500-year spin-up phase was used to allow the LPJ-GUESS vegetation to reach approximate equilibrium (confirmed by checking that net ecosystem exchange shows no systematic deviations from zero; see Appendix
For the
Stand-alone LPJ-GUESS has a long history of development and has been evaluated in detail in previous work. (Some recent examples include modelled potential natural vegetation and forest stand structure and development,
At this stage of model development we do not seek to precisely simulate the vegetation state of a particular year or exact period. Our atmospheric simulations are not nudged by meteorological data but rather an unconstrained simulation based on climatological SSTs and SIC, so they do not correspond to a particular calendar period. Furthermore, we prescribe a fixed atmospheric
Knowledge of EMAC biases is very useful for disentangling the causes of model–data disagreement in the simulated vegetation. To this end, we include bias plots of seasonal and annual biases in surface temperature, precipitation, and net (plant-available) shortwave radiation of the EMAC
To provide a visual assessment of the structure and functioning of the vegetation cover at a level of detail relevant for studying interactions between the land surface and the atmosphere, we categorised the simulated vegetation into eight “megabiome” types and compared them to an expert-derived PNV map with equivalent categories. The classification of the simulated vegetation was based on leaf area index (LAI) following
For quantitative evaluation of the simulated vegetation, we chose four vegetation state variables which are informative when evaluating ESMs or DGVMs, particularly in regards to the biophysical coupling and carbon flux between the land surface and the atmosphere: fractional coverage of trees, standing biomass, canopy height, and gross primary productivity (GPP). Fractional coverage of trees is relevant both for the evaluation of stand-alone DGVMs (as it is a result of both overall productivity and vegetation dynamics such as tree–grass competition and disturbance regimes) and for land surface schemes (as forested areas have different biophysical properties as non-forested areas). To evaluate tree cover, collection 6 of the MOD44B MODIS tree cover dataset
Standing biomass is a key state variable in ESM and DGVMs as it is connected to productivity, carbon sequestration, evapotranspiration, vegetation cover, canopy height, and other critical processes and variables. As such, it is a useful quantity for evaluating DGVM and ESM performance. We produced a near-global map of standing biomass by combining two biomass datasets, one tropical
Canopy height is highly relevant in a land–atmosphere context as it has a direct effect on atmospheric circulation through surface roughness length. To evaluate simulated canopy height, a 1 km tree canopy height map
GPP is the critical quantity in Earth system modelling, both in terms of the planetary
Taken together, these four quantities or datasets capture many of the key features of vegetation structure and functioning which affect biophysical land–atmosphere exchanges. The datasets were regridded from their intermediate resolutions to the simulation resolution using second-order conservative remapping
To provide an overall summary metric of data–model agreement across the relevant spatial domain, the normalised mean error (NME) is presented following the prescription and recommendations in
Distribution of PNV megabiomes simulated by LPJ-GUESS within EMAC (
As NME quantifies the absolute error in the model as compared to the data, the relative difference in the values for two models (compared to the same dataset) can be considered the relative improvement of one over the other. For example, if one model yields a score of 0.8 and a second yields a score of 0.6, the second one can be said to be 25 % better than the first, since
It should be noted that the NME is rather different from a coefficient of correlation or a coefficient of determination. It does not attempt to derive a correlation but instead sums the differences between the model and the observation. It can be thought of as quantifying the deviation from the one-to-one line of perfect data–model agreement, rather than the deviation from a line of best fit. This means that it is a rather direct and unforgiving metric, since every deviation of the model from the data is penalised (uncertainty is not considered), and there is no possibility for the line of best fit to move to compensate for systematic biases. It also means that the values are interpreted in the opposite direction to a correlation coefficient; an NME score of zero implies perfect agreement between observation and model, whereas an
Comparison of GPP from the
The simulations reproduced the global patterns of vegetation type well (Fig.
Comparison of tree cover from the
The extent of the temperate forest vegetation zones of the east coasts of the USA and China was underestimated in the EMAC simulations (Fig.
Comparison of biomass from the
The high productivity of the tropical rainforests was strongly underestimated by the EMAC simulations and to a lesser degree by the
The coupled model showed a tendency to overestimate GPP and, to some extent, biomass in the arid continental interiors in central Asia and central North America (Figs.
Comparison of canopy height from the
NME scores for the vegetation produced by the
The canopy height data were produced in such a way that no land use correction is necessary, and the land use cannot be meaningfully applied to the modelled GPP.
Considering tree cover specifically, the combined model produced reasonable global tree cover patterns (Fig.
The NME scores for all simulations and all evaluation variables are presented in Table
Applying the LUC has a marked improvement on the tree cover NME scores (in terms of percentage reduction of error: 16 % for
For the coupled simulations, increasing spatial resolution improved the agreement between simulations and observations for all variables, with the exception of biomass at the higher resolutions. The GPP agreement improved consistently by 2 % with increasing spatial resolution. The canopy height NME improved by 13 % from
In light of these results, we would recommend against using the T42 resolution given the high-latitude growing-season temperature and radiation biases and its effect on the vegetation and, potentially, the resulting feedback to the atmosphere (in the case of the fully coupled model). These effects can be mitigated to a large extent by using the T63 resolution without incurring too much additional computation cost. Stepping up to T85 resolution or higher may or may not be beneficial, depending on the details of the simulation and the study.
The work and simulations presented here are only the first milestone on a planned model integration roadmap. Figure
Following this, the next critical step will be to enable land use and agriculture in LPJ-GUESS within EMAC. This will be beneficial not just in terms of improving the representation of the land surface as a boundary condition to the atmospheric circulation model, but also because model evaluation and benchmarking will become easier to perform and interpret. However, this step will involve a significant amount of development work to modify the LPJ-GUESS code to receive land cover, state transition, and management data from EMAC rather than through the existing channels in LPJ-GUESS. In contrast, the calculation of biogenic volatile organic compounds in LPJ-GUESS will be fairly simple as the only additional variables that are required are daily maximum and minimum temperatures
Another further developmental step is to improve the representation of fire and associated emissions in LPJ-GUESS. The GlobFIRM fire model
Initially, LPJ-GUESS was developed as a stand-alone DGVM featuring biogeochemical cycling and vegetation dynamics. It was not designed as a land surface scheme and so some physical properties of the vegetation, such as canopy height, were not high priorities during development. Furthermore, many remotely sensed datasets, such as the canopy height data used here, were not available during the model's initial development and calibration. It is therefore not surprising that in this study we found that GPP was the best-simulated quantity, and canopy height was the least well simulated. Given the direct effect of canopy height on the atmosphere via roughness length, it may be appropriate to adjust the parameterisation in LPJ-GUESS to improve the simulation of canopy height. Candidate parameters include PFT-specific coefficients in the allometric equations
Longer-term and more ambitious goals of the roadmap are to fully replace the soil–vegetation part of the hydrological cycle in EMAC with that of LPJ-GUESS and to use LPJ-GUESS to close the land surface energy balance. Such developments may benefit from synergies with other ongoing coupling work in the LPJ-GUESS community. When completed, these developments will extend the EMAC model into a full Earth system model including atmosphere (ECHAM5) with full chemistry
Here we have reported the first steps towards producing a new atmospheric chemistry-enabled ESM by combining an atmospheric chemistry-enabled AOGCM with a DGVM. The technical coupling work is now complete and has been achieved in a manner which respects both the integrity and philosophy of the two modelling frameworks, and this will therefore allow for relatively straightforward updates to both components.
Results from one-way coupled simulations (in which climate information generated by EMAC is used to force LPJ-GUESS, but no land surface information is relayed back to EMAC) showed that the vegetation patterns produced from the EMAC climate are reasonable on a global scale. However, some regional deviations from the observed vegetation are apparent. Some of these are due to the simple fact that in this configuration LPJ-GUESS produces PNV (potential natural vegetation with no human impacts), while the observed vegetation implicitly includes human impact. This effect was confirmed by performing a correction to account for human land use, which improved agreement between simulation and observation. Human land use can be included in future model versions by utilising the recently developed crop and managed land module in LPJ-GUESS
A second class of deviations is due to biases in the simulated climate, particularly precipitation biases. This is a more difficult problem to solve; improving climate simulations is the subject of much ongoing research. However, it is clear that using higher spatial resolution mitigates climate biases, which results in tangible improvements in the simulated vegetation. Based on the three spatial resolutions, we recommend using the T63 resolution as a minimum due to climate biases in the high latitudes in the T42 simulation, which resulted in insufficient growth of vegetation. However, using dynamically simulated land surface boundary conditions (in this case from LPJ-GUESS) in a bidirectionally coupled model will alter the atmospheric state and therefore the climate biases. This will be the subject of future studies.
Finally, there are some discrepancies arising as an inevitable consequence of the approximations, missing processes, and parameter uncertainties inherent in a process-based model such as LPJ-GUESS. These may be reduced by ongoing improvements occurring as LPJ-GUESS is further developed and refined. Given the rather rigorous requirements placed on a biosphere model when bidirectionally coupled to an atmospheric model, it may also be necessary to perform some focused model development work with the goal of improving vegetation functioning and structure so that key biophysical quantities (such as albedo and roughness length) are better simulated. Of the variables evaluated here, canopy height was found to be the least well simulated, suggesting that retuning tree height in LPJ-GUESS might be an important step to ensure good performance of the fully coupled model.
Whilst further work remains before the full ESM is completed, we have demonstrated that coupling LPJ-GUESS to the EMAC/MESSy modelling framework has been accomplished, and LPJ-GUESS provides a suitable basis for an improved and dynamic representation of the land surface in EMAC. Future development should focus on completing the two-way model coupling and investigate the effects of the atmosphere. Once the full coupling has been enabled and calibrated, the resulting model will be a powerful tool for investigating atmosphere–biosphere interactions. In addition to the broad range of applications possible for any ESM, the particular strength of EMAC with LPJ-GUESS vegetation will be applications studying interactions and feedbacks at the atmosphere–biosphere boundary, e.g. the nitrogen cycle; trace-gas emissions from fire; the atmospheric dynamics of reduced carbon, including biogenic volatile organic compound emissions from vegetation and methane from fires; ozone dynamics and the resulting damage to vegetation; and the effects of a wide spectrum of terrestrially emitted trace gases on cloud and aerosol formation and dynamics.
One of the main priorities during the coupling implementation was to change the LPJ-GUESS source code as little as possible. As such, only the following modifications were made to the LPJ-GUESS code:
Three new functions were created, which are called externally by the MESSy framework. The first function initialises an LPJ-GUESS simulation (or restarts from a saved state if appropriate); the second function performs 1 d of LPJ-GUESS simulation given 1 d of EMAC climate data; and the final function saves the LPJ-GUESS state to disc. These key functions encapsulate the interactions between MESSy and LPJ-GUESS. A new input module was created (an instantiation of the LPJ-GUESS C++ class One additional internal function was created to calculate the daily values to be handed back to EMAC (such as vegetation cover for a particular PFT). An additional output module was included to save model output that is useful for benchmarking. Minor modifications to the standard output module were made such that the MPI (Message Passing Interface) rank number of each process is added to the file output names, allowing the output from each process to be stored in the same directory. Minor modifications to the standard LPJ-GUESS restart code were made to allow the MESSy restart cycle number to be added to the names of the state files to be saved or read by LPJ-GUESS. Some of the code for the LPJ-GUESS real-time visualisations, which is incompatible with the MESSy framework, was removed.
No changes to the scientific modules were made, and the directory structure and compilation machinery were untouched. Wherever new code conflicted with the standard offline version, a preprocessor directive was used to ensure that the model still compiled in the standard way outside the MESSy framework. Thus, the integrity of LPJ-GUESS was maintained so that updates from the LPJ-GUESS trunk version can be applied relatively easily, and the code can still be compiled and run offline.
On the MESSy side, the The current development version of EMAC including LPJ-GUESS is equipped with a new “Makefile” for a standard Linux “make (gmake)” with the same functionality as the original “cmake” compilation; the updated compilation process does not require an up-to-date installation of cmake.
In the initialisation phase, the grid from EMAC is transferred into LPJ-GUESS. Note that currently there is only a geographic decomposition induced by EMAC, which could lead to some processors not
having a single land box and causing idle time for that specific CPU. In future, an additional, individual decomposition of the land grid cells to optimise CPU balance is desired, which could make use of the
In its interface layer, the VEG sub-model accumulates the required input fields (daily temperature, precipitation, incoming solar radiation, and atmospheric
The combined model uses the pre-existing restart facilities of the LPJ-GUESS code. When EMAC triggers a restart, a restart is triggered for LPJ-GUESS; when a simulation is continuing from a restart, a flag is sent to the LPJ-GUESS code and the restart files of the LPJ-GUESS state are read-in, allowing a seamless, continuous simulation. This feature may also be used to start a simulation with already well-established vegetation from LPJ-GUESS restart (state) files, potentially saving significant amounts of CPU time that would otherwise be required for vegetation spin-up (typically of the order of 500 simulation years).
The net ecosystem change plots shown in Fig.
Net ecosystem exchange (NEE) for all EMAC simulations
The
The
The
Canopy height of a patch was calculated from individual tree cohort heights by a simple algorithm that attempts to reconstruct top-of-canopy height as it would be viewed from above, for example by a satellite. It utilises the modelled quantity foliar projective cover (FPC), which is the ground area covered by the crowns of trees of a cohort expressed as a fraction of the patch area. LPJ-GUESS allows for limited overlapping of trees, and hence the sum of tree cohort FPC can be greater than unity. In this case cohorts are selected in descending order of height until the sum of their FPC reaches one, i.e. smaller cohorts are assumed to be under the taller cohorts and so do not contribute to top-of-canopy height. Cohorts smaller than 5 m do not contribute to canopy height as the remotely sensed dataset does not include canopies lower than 5 m. Having selected the contributing tree cohorts, the canopy height is simply the FPC-weighted sum of the contributing cohort heights.
In order to correct the model output for “missing” tree cover and biomass due to human land cover modification, a simple correction was derived from the Globcover2009 land cover product 40 closed-to-open broadleaved evergreen or semi-deciduous forest, 50 closed broadleaved deciduous forest, 60 open broadleaved deciduous forest/woodland, 70 closed needleleaf evergreen forest, 90 open needleleaf deciduous or evergreen forest, 100 closed-to-open mixed broadleaved and needleleaf forest, 110 mosaic forest or shrubland/grassland, 120 mosaic grassland/forest or shrubland, 130 closed-to-open (broadleaved or needleleaf, evergreen or deciduous) shrubland, 140 closed-to-open herbaceous vegetation (grassland, savannas or lichens/mosses).
The Modular Earth Submodel System (MESSy) is continuously developed and applied by a consortium of institutions. The usage of MESSy and access to the source code is licensed to all affiliates of institutions which are members of the MESSy Consortium. Institutions can become a member of the MESSy Consortium by signing the MESSy Memorandum of Understanding. More information can be found on the MESSy Consortium website (
LPJ-GUESS is used and developed world wide, but development is managed and the code maintained at the Department of Physical Geography and Ecosystem Science, Lund University, Sweden. Model code can be made available to collaborators on entering into a collaboration agreement with the acceptance of certain conditions. The MESSy-coupled version of LPJ-GUESS will be maintained as a derivative of LPJ-GUESS. Because access to LPJ-GUESS is also restricted, no DOI can be assigned to LPJ-GUESS versions. The specific code version used here to enable the MESSy coupling, the LPJ-GUESS code in EMAC, is archived on the LPJ-GUESS subversion server with tag “_publications/MESSY_1.0_20180108” in the catalogue “MESSy”. For more details and contact information, please see the LPJ-GUESS website (
For review purposes, the code used here is available to the editor and reviewers via a password-protected link on condition that the code is for review purposes only, it cannot be used for any other purposes and must be deleted afterwards.
HT and MF performed the model coupling. MF performed the simulations and analysis. All authors contributed to the overall model coupling strategy and to the article.
The authors declare that they have no conflict of interest.
Parts of this research were conducted using
the supercomputer Mogon and/or advisory services offered by Johannes Gutenberg University Mainz (
The authors acknowledge the support of the MESSy core development team and are grateful for hints and discussions. Similarly, the authors recognise and appreciate the many improvements to LPJ-GUESS by the LPJ-GUESS development team which made this work possible, and they thank the team (particularly Johan Nord) for their support. We also thank and acknowledge Allan Spessa for his contributions at the conception stage of this project.
This paper was edited by Gerd A. Folberth and reviewed by Douglas I. Kelley and two anonymous referees.