Models for reporting forest litter and soil C pools in national greenhouse gas inventories: methodological considerations and requirements

ABSTRACT

The compilation of GHG inventories has become a common practice to support environmental decision- and policy-making, for example in the context of the UNFCCC. To estimate GHG emissions, simulation models present viable alternatives to measured data. In order to make an informed decision on model selection, clear information on the purpose and applicability of a particular model is needed. This article discusses model requirements with respect to the suitability of estimating the carbon balance of dead wood and litter (dead organic matter; DOM) and soil in forests with a particular focus on policy needs under international processes such as the UNFCCC. Based on criteria established for GHG reporting under the UNFCCC including transparency, consistency, comparability, completeness and accuracy, this paper presents an approach to classify commonly used simulation models for estimating C budgets of DOM and soil in European forests. Among the six discussed models, the authors found a clear trend toward models for providing quantitatively precise, site-specific estimates. To meet reporting needs for national GHG inventories, the authors conclude that there is a need for models producing qualitative realistic and unbiased estimates at larger scales in a transparent and comparable manner.

KEYWORDS:

• accuracy
• comparability
• completeness
• consistency, transparency
• carbon balance

Introduction

The information compiled in greenhouse gas inventories (GHGI) represents the basis for climate-related policy-making. The United Nations Framework Convention on Climate Change (UNFCCC) and its Kyoto Protocol (KP) recognize the importance of the land-use sector as a potential way to mitigate climate change. In this context, industrialized countries (Annex I) prepare annual inventories of estimated greenhouse gas (GHG) emissions and removals from the “Land Use, Land Use Change and Forestry” (LULUCF) sector. GHGIs have to be prepared using the methodological guidance prepared by the Intergovernmental Panel on Climate Change [1]. The LULUCF sector has been traditionally very complex and controversial [2,3], also due to the uncertainties of the estimated emissions and removals.

The reporting for LULUCF is slightly different under the UNFCCC, which distinguishes land-use categories (e.g., forest land, cropland, grassland, etc.) following IPCC [1] and under KP, which distinguishes “activities” (e.g., forest management, afforestation/reforestation and deforestation) following IPCC [4]. In both cases, forests play a major role as carbon (C) sources and sinks [5,6]. Furthermore, emissions and removals need to be reported for five different C pools: above- and below-ground living biomass, litter and dead wood (collectively also called “dead organic matter”, DOM), and soil organic matter (when relevant disaggregated between mineral and organic soil).

An important data source for estimating emission and removals from forests are national forest inventories (NFIs), which have been established in many countries and provide accurate and representative information on forest attributes based on field measurements [7,8]. Since, traditionally, NFIs had a focus on the productive capacity of forests – that is, growing stock and increment of stemwood – estimates for the volume and biomass of living trees can be obtained with high confidence [9]. With the recognition of the importance of forests as reservoirs for biodiversity and as C stores, the scope of NFIs was enlarged and new variables such as dead wood and litter have been added [7], improving also the ability to quantify DOM abundance and volume. However, estimates of C stocks in DOM are associated with large uncertainty since the spatial distribution of DOM in forests is heterogeneous and the quantity is controlled by many factors including abiotic and biotic damages, forest management strategies and physical characteristics such as volume and wood density of dead wood change during decomposition [10]. At the same time, reliable estimates of C stocks and C stock change in forest soils are also often not available [11]. Regarding the scarcity of observed data on soil C stock changes, the use of higher tier methods including models may be used to estimate both C stocks and C stock changes in DOM and mineral soils.

Within the GHGIs submitted under the UNFCCC, the coverage of reporting of forest C pools varies considerably among countries. For example, only about 55% of Annex I countries report estimates for mineral soil, and only 70% report on DOM. However, under the KP, if a C pool is not estimated, evidence should be provided showing that this pool is not a source of C. During the review of Annex I GHGIs under the KP, providing evidence for forest mineral soil not being a source has been one of the most important challenges encountered by countries. To this regard, the need for additional tools supporting countries in reporting C stock changes in forest mineral soils has been recognized by the Joint Research Center (JRC) of the European Union, which provides support to EU member states in the reporting of their GHGI.

In this context, the use of models is considered an appropriate method to estimate emissions and removals under the UNFCCC/KP, or at least to provide evidence for soils being “not a source” under the KP. According to IPCC [12], models in the LULUCF sector may have several advantages for estimating emissions and removals as compared to other approaches (e.g., repeated sampling of C stocks), such as potentially greater accuracy, higher spatial and temporal resolution, better representation of complex dynamics (e.g., DOM and soil) and of climate/disturbances effects, and also cost efficiency. Furthermore, models may be used to consistently estimate past data and future projections, a feature which is of increasing interest, for example, for the Forest Management Reference Level [4]. Models used in GHGIs have to comply with the general UNFCCC reporting principles of transparency, consistency, comparability, completeness and accuracy [1,101]. For a model, the most challenging principle is probably transparency [see, e.g., 12]. To this regard, a model should be verifiable and should be able to realistically reproduce the main processes relevant for the estimation (e.g., the biomass–DOM–soil C dynamics). At the same time, a model should be implementable over a range of environmental conditions, and capable of accommodating differences in the available (data) resources.

By definition, a model is a simplified representation of a real system. Following Levins [13], they thus present a compromise between realism, generality and precision. Although Levins’ theory has been critically discussed, there is consensus that trade-offs exist [cf., 14]. Realism, or the aim to obtain qualitative rather than quantitative (i.e., precise, sensu Levins) results, is probably the foremost objective of international GHG reporting as it ensures comparability. Depending on the model aims and structure, a model presents a trade-off among the three criteria [14]. It is therefore important that these trade-offs are considered in the model selection with respect to the objectives of the intended application [15]. For example, a model aiming for precise, site-specific estimates compromises on the generality of the predictions.

This article reviews six DOM and soil models which have been applied for estimating C balances of European forests with respect to the requirements for reporting GHGs under international agreements. The focus here is on change in soil-related C pools, whose reporting proved particularly challenging even for developed countries [cf., 16]. The application of a model is challenging for several reasons, including that a model must be developed or it must be selected from available models; criteria such as transparency, comparability and accuracy must be met, which is especially true for policy or reporting applications such as under the UNFCCC or the KP; and the model must be able to work within the constraints of available data. The aim of this paper is to identify criteria that need to be considered in the selection of a model, and to discuss available alternatives and approaches and the implications for a model application in a policy context. To this end, a framework is suggested that can be used in situations where it is not clear which approach to take.

To the authors’ knowledge, this work represents the first attempt to define a qualitative measure relating model attributes to general policy and reporting criteria at the national scale. For this purpose, the authors integrated information from several modeling studies in order to derive a robust and consistent evaluation of the examined models. The goal for a qualitative evaluation is particularly important as it allows assessing the utility of a modeling approach in cases of data constraints and inconsistencies at regional and national scales. Here, the authors expand on previous model evaluations which were generally restricted to the sub-national scale, and optimal conditions for model implementation.

Models for estimating C budgets of DOM and soil in forests

For research on biogeochemical processes in soils, a large number of models have been developed [17,18]. The majority of these models have been developed for cropland or grassland to study C and nutrient fluxes. Few of these models have been modified for application in forests [e.g., 19–21] and only a few models have been developed specifically for application in forests (e.g., CBM-CFS [22]; ROMUL [23]; and Yasso07; Tuomi et al. [24]). Almost all DOM and soil models simulate C stocks in mineral soils, but are typically not able to operate on organic soils with prevailing anaerobic conditions [25]. Based on literature and national GHGI submissions to the UNFCCC [26], the authors found the following models as candidates for estimating C budgets of DOM and mineral soils in European forests (the main model reference and most recent applications in European forest ecosystems, with comparisons between simulated and observed data, are cited):

• - CBM-CFS ([22,27–29,);

• - CoupModel [30–33];

• - Q [34–36];

• - ROMUL [23,37–38];

• - RothC [39–41];

• - Yasso07 [42,24,43,10].

All of these models have been applied in research in European forests, and all except CBM-CFS have been included in studies ceomparing model performance (Table 1). The relevant studies could rely on the availability of comprehensive data sets required for model implementation, which is not necessarily the case in regional or national assessments of the forest C balance [44]. Further models exist but have not been used to estimate the C balance of forests in Europe and are, thus, not further discussed here, for example, G'Day for eucalyptus plantations [19] or FullCAM for the C balance of Australian forests [45]. For applications on waterlogged soils, including peatlands, specialized models have been developed to account for the different GHGs found in such soils, including methane and nitrous oxide [see 46].

Table 1. Comparative studies of C-cycling models which also evaluated the models presented above for studying C-cycling in European forests. Models considered in this article in bold.

Models	Main findings	Ref.
Century, CoupModel, Forest-DNDC, Q, ROMUL, RothC, SoilCO2/RothC, Yasso07	The study recommended the use of relatively simple DOM and soil models with low data requirements for inventory-type, large-scale carbon assessments because of better data availability, easier assessment of uncertainties and higher degree of transparency.	Palosuo et al. [65]
ROMUL, Yasso, Yasso07	The model estimates were very similar for sites with low annual precipitation. The differences in model estimates at sites with high annual precipitation were likely due to the different time steps – that is, annual in Yasso07 and daily in ROMUL. The results from both models were reasonably close to measured values of the examined forest soils in Finland. The conclusion was based on the prevalent conditions in Finland, and the authors indicated possible shortcomings of ROMUL in regions with higher precipitation.	Rantakari et al. [61]
Century, CN-SIM, ICBM, Q, RothC, Yasso07	Field study data from agricultural soils in Sweden were used in different modelling studies to evaluate model performance, among others the Q-model [91] and RothC [49]. Karhu et al. [90] applied Yasso07 and synthesized the findings from previous studies: all studied models produced estimates that were in good agreement with measured data. In comparison with more complex models with detailed representations of multiple soil pools, Yasso07 was able to reproduce effects of N fertilization well without the complexity of explicit simulation of soil processes.	Karhu et al. [90]
CoupModel, Q	Qualitatively, the models behaved similarly. Due to the higher productivity of the forest simulated by the Q model, the model predicted larger losses of soil carbon.	Eliasson et al. [92]
Q, Yasso07	C stocks, C stock changes, inter-annual variations and uncertainties were of the same magnitude among measured data from the Swedish Forest Soil Inventory (SFSI), and simulated data from Q and Yasso07. Uncertainty in soil C stocks was driven primarily by litter inputs for Yasso07 and by model parameters for Q. Litter inputs and climate were the main source of uncertainty in soil C stock changes in both models.	Ortiz et al. [62]

The examined models differ in scope and complexity: Q, ROMUL, RothC and Yasso07 are DOM and soil models that describe C decomposition in woody and non-woody DOM and soil, and require external inputs of DOM production as a driving variable. These models can be run at individual sites. The ecosystem models CBM-CFS and CoupModel additionally simulate plant growth including DOM turnover and the interaction with soil. CBM-CFS was developed specifically to support GHG reporting applications. It has an empirical foundation, as the simulation of biomass change in the model relies on spatially referenced forest inventory data [22]. A more process-based approach is represented in CoupModel, where growth emerges from the simulated constraints of climate and soil and competition for light [31]. Decomposition of litter and soil processes are simulated in a similar fashion – that is, empirically based – in CBM-CFS, and process-based in CoupModel. The DOM and soil models are process-based but differ in the number and type of processes that are incorporated: Yasso07 and Q describe C decomposition based on litter properties, while ROMUL and RothC focus on soil properties such as soil texture, clay and moisture contents, and bulk density. The models also differ in the extent to which they consider soil depth and layers within the soil ranging from a fixed depth and single layers (e.g., 20 cm in ROMUL, 100 cm in Yasso07) to variable depth and multiple layers in CoupModel. A further property by which models can be distinguished is the temporal resolution. The CoupModel can be used with sub-daily time steps, while the temporal resolution in the other models ranges between 1 month (e.g., ROMUL, RothC) and 1 year (e.g., CBM-CFS, Yasso07).

Dead wood and also litter are important C sources that affect the soil organic carbon (SOC) pool [47,48]. Following the IPCC guidelines, dead wood, litter and SOC constitute separate C pools which need to be accounted for [1]. The examined models differ in their ability to separate these three C pools. CBM-CFS explicitly accounts for all pools defined in the guidelines [22] – however, with the need to derive all required parameters. Other models except CoupModel accept inputs separated by source (i.e., dead wood and litter), or the pools may be distinguished conceptually (e.g., fast and slow pools in RothC; cf., [49]). This allows those models to account for the difference in decay according to the C source. With the exception of CBM-CFS and ROMUL, the models do not separate C from the inputs that are not emitted to the atmosphere into distinct pools, and consider only one common C pool as default output.

In all discussed models temperature is a main driver of decomposition, which is consistent with the current understanding of the decay process [50,51]. The models differ regarding the representation of the effect of moisture in decomposition which has been established at very different levels of detail. The varied implementation of the decomposition dependency on moisture reflects the underlying approach in the development of a model, and the only moderate understanding of the effect of moisture on decomposition [52,51]. The degree of accounting for the effects of moisture ranges from no (CBM-CFS, Q) or readily available annual precipitation sums (Yasso07) in the models aimed for general application to sub-annual estimates of soil moisture as a function of evapotranspiration (ROMUL, RothC) to daily data on precipitation, wind speed, humidity and radiation (CoupModel) in the more complex models aiming for site-specific precision.

Some of the models (e.g., ROMUL and Yasso07) were developed specifically for application in forests and have been improved to reproduce decomposition of the highly variable litter types in forests [54] compared to grasslands and croplands. This includes dead wood of different sizes and dead roots as well as fallen leaves and needles [55] with different decomposition characteristics depending particularly on their size [e.g., 56,57], and chemical composition including lignin content [e.g., 58]. The models work well with average values for these input parameters, but where available local or regional values can be used to improve the accuracy of the results [e.g., 10].

A consequence of the outlined differences in model structure and simulation approach is a varying degree of data needs (Table 2). Models aiming for quantitatively precise C stock estimates typically require more data than simpler models with the aim of producing qualitative realistic results do. Depending on the model and the aim of the user, the process of deriving initial conditions for a simulation may require additional data. All discussed models can be started using measured values or following an initialization procedure (spin-up), which generates conditions where C pools are in an equilibrium with assumed historical C inputs. Only CBM-CFS explicitly considers the dominant historical disturbance, and the spin-up produces a quasi-steady state [22].

Table 2. Data requirements for a model application to estimate soil related C stocks (mineral soil, litter, dead wood). Required input data are ranked based on their availability^‡ (both spatially and temporally) and accuracy: “+” indicates generally available data with high accuracy; “o” some limitations to data availability and accuracy (e.g., downscaling or temporal interpolation necessary); “-” availability is spatially limited (e.g., only at local scale and not consistently across space and time but default values provide good surrogates); “--” availability and accuracy are generally limited and defaults are not applicable for use across space and time. The criteria for ranking availability and accuracy of input data were based on the need to apply methods or models to obtain estimates, which may be labor intensive (e.g., to obtain daily estimates further processing is required, reducing availability and accuracy). The table presents the minimum input requirements as indicated in relevant publications; additional optional inputs to replace defaults can contribute to improving accuracy and reducing uncertainty in the estimates. Where applicable, the required temporal resolution is provided. The ecosystem models CBM-CFS and CoupModel simulate the C inputs (i.e., needles/leaves, branches, stems, coarse roots, fine roots) to the embedded soil module; the respective input requirements to simulate the dynamics of the forest ecosystem are generally available from regular forest inventories and are not included here.

Model	Data requirements for model application
CBM-CFS (ecosystem model)	- mean annual temperature (+) - annual C inputs from simulated living biomass development (+)
CoupModel (ecosystem model)	- mean daily temperature (-) - daily precipitation sums (-) - mean daily radiation (-) - mean daily air humidity (–) - mean daily wind speed (–) - daily N deposition^† (o) - C inputs from simulated living biomass development (+) - clay, silt, sand fraction (default available for Sweden)^† (o) - soil C content^† (o) - soil N content^† (o)
Q	- mean annual temperature (+) - annual C inputs (+) - fractions of chemical compounds of fresh litter^† (+)
ROMUL	- mean daily temperature (-) - daily precipitation sums (-) - daily C inputs (-) - N and ash content of fresh dead wood and litter (+) - soil water holding capacity (o) - saturated water content (o) - soil temperature (-) - soil N content^† (o)
RothC	- mean monthly temperature (+) - monthly precipitation sums (+) - monthly evaporation (o) - monthly C inputs (o) - fraction of decomposable and resistant material in fresh dead wood and litter^† (+) - clay content ^† (o)
Yasso07	- mean annual temperature (+) - annual precipitation sums (+) - annual temperature amplitude (+) - annual C inputs (+) - fractions of chemical compounds of fresh litter^† (+) - mean diameter of dead wood (+)

^†default values are provided and alternative estimates can be used.

^‡Data availability (see also the in-text discussion):

- Meteorological data: monthly and annual are readily available for common variables such as temperature and precipitation; further processing (e.g., [81]) is required for less common variables such as evaporation and radiation and for daily data.

- C inputs: annual data can be derived from national forest inventories; sub-annual data are typically not available. Data on dead wood and litter quality is available from literature.

- Soil data: information is often not available consistently across space and limited regarding completeness of measured attributes.

The trade-off against local precision that simpler models make can be addressed by concurrent applications of different model types. For this, it is necessary that all required inputs are available (cf., Table 2). Further, it would be valuable to harmonize model interfaces to simplify the concurrent application of more than one model. The SoilR modelling framework [59,60] presents a first step in this direction. In such a way, concurrent model application can be used to corroborate the estimated C dynamics and induce more confidence in the results. To the same end, it may also be possible to develop alternative parameter sets for models with a focus on generality. These parameter sets can be based on subsets of the calibration data in order to obtain a “local” implementation of a model (e.g., Yasso07 for Nordic countries; [61]). Such exercises with parallel model application can help to identify knowledge gaps and uncertainties.

A concurrent application of models or a use of alternative parameterizations can also help with the understanding and interpretation of simulated decomposition dynamics particularly with regards to the realism under environmental conditions within a case study region. Estimates of C stocks in particular may differ depending on the model [61,62] or the parameterization. Figure 1 shows the effect of two parameterizations of Yasso07 on the decomposition rate of humus, highlighting the importance of applying a model or a parameterization only at the scale for which it was designed [cf., 18]. The figure demonstrates that for a given combination of annual mean temperature and precipitation, the emerging humus decomposition rate can differ strongly as illustrated by the example of data from Hernández et al. [63] for three climatically different countries. Information of this kind is valuable to increase transparency and to inform potential model users of the suitability of a model for a particular application.

Figure 1. Humus decomposition rate (kH) across a gradient of mean annual temperature and precipitation sum based on two parameterizations; left: Tuomi et al. [68] using a global data set of litter mass loss measurements (Table 1 in Tuomi et al. [68]) with additional data on SOC accumulation from a soil chronosequence in southern Finland (Liski et al. [89]); right: Rantakari et al. [61] using a subset of the data used by Tuomi et al. [68] which was restricted to Scandinavian sites. Data from Hernández et al. [63]. The estimates for kH do not present actual decomposition rates since these may vary due to covariation between parameters within one set of parameters, i.e., Tuomi et al. [68] and Rantakari et al. [61], respectively. Note the difference in the range of the y-axis values in the two panels.

Based on its data needs and modelling approach, a model can be classified with respect to the trade-off among precision (producing quantitatively precise estimates), realism (producing qualitative realistic estimates) and generality (representing a broad range of conditions without model modifications), following Levins [13]. Table 3 and Figure 2 present, respectively, a quantitative and a qualitative attempt to organize the discussed models with respect to an application for estimating national C balances of forest dead wood, litter and soil.

Table 3. Model suitability to estimate national-scale C balances of forest dead wood, litter and soil with respect to the trade-off among precision (producing quantitatively precise estimates), realism (producing qualitative realistic estimates) and generality (representing a broad range of conditions without model modifications) following Levins [13]. 5: high, 4: intermediate-high, 3: intermediate, 2: intermediate-low, 1: poor. The interpretation was based on the need for model (and parameter) modifications for applying a model at different sites, the number of input data, the ease of availability of input data (cf., Table 2) including temporal resolution and for multiple locations within one region, and studies of model comparison (cf., Table 1). The ecosystem models CBM-CFS and CoupModel were classified considering the additional data requirements for simulating forest growth. Note that this classification is subjective and presents an estimation that is valid within the scope of the applications for national scale forest C balance reporting.

Model	Generality	Realism	Precision
CBM-CFS (ecosystem model)	3: With only temperature as input in addition to C inputs in various pools, the generality of the litter and soil module is limited only by the number of C pools (11) with respective data and parameter needs.	3: Limitations with respect to temperature as the sole driver of C dynamics.	3–4: The model simulates a total of 21 C pools including 11 litter and soil pools.
CoupModel (ecosystem model)	1: Limitations due to high data requirements at fine temporal resolution also for the soil module.	2: The great detail of data requirements and of implemented processes limits the verifiability due to the scarcity of relevant data.	4–5: The decomposition of C in the soil is simulated with great detail, and multiple soil layers can be defined.
Q	4: Some limitations due to the detailed C input requirements with respective data and parameter needs.	3: Limitations with respect to temperature as the sole driver of C dynamics.	2: Limitations with respect to site-specific application for soil C, as soil properties are not represented.
ROMUL	2: Limitations due to the detail and temporal resolution of the input data, and restricted range of parameter validity.	3: Limitations with respect to model validity outside the range where it was calibrated.	4: The model uses a daily time step, and the soil moisture regime is simulated with a separate module.
RothC	3: Restricted by the need for monthly data. Limitations with respect to validity of model parameters for woody litter.	3–4: Important drivers except dead wood dimension are used.	3: Monthly data and clay content as proxy for soil properties achieve a degree of precision.
Yasso07	4–5: Model parameters are derived from global data sets and model inputs requirements are low.	4: Important drivers of dead wood and litter decomposition (dimensions, chemical composition, temperature, moisture) are used.	2: Limitations with respect to site-specific application for soil C, as soil properties are not represented.

Figure 2. Following Levins [13], models present a trade-off among precision (producing quantitatively precise estimates), realism (producing qualitative realistic estimates) and generality (representing a broad range of conditions without model modifications). Circles indicate a likely range estimated based on Tables 1–3.

Criteria for policy application of models

The UNFCCC identified five important criteria that an inventory of GHGs should meet [101]: transparency, consistency, comparability, completeness and accuracy (TCCCA). Additional elements such as uncertainty estimation, quality assurance and control (QA/QC), and verification apply as well [12]. Due to the international focus that guided the development of the IPCC criteria, individual criteria may or may not apply in applications at national or smaller scales.

Transparency was identified as the central concern in the use of models in GHGIs [12]. The IPCC guidelines define transparency as the requirement for a clear explanation of assumptions and technologies to facilitate replication and assessment of the inventory estimates [1]. A transparent model would thus have a simple structure which is easy to understand [12]. However, simple models of C cycling are criticized as being overly simple as they do not include soil properties such as clay content and N, which have been found to affect C sequestration in soils [e.g., 64]. While this criticism may be correct, simple models have the advantage of being transparent and easy to implement at different sites due to the low input requirements [65]. Comparisons including models of different complexity have not shown that simple models are less accurate than more complex models (e.g., [61,65,66]; Table 1). Helfrich et al. [66] argued that because there is still a lack of data for applying and validating complex models, simple models suffice for describing C dynamics in the soil.

Among other things, a clear documentation of model assumptions, domain of application, and data sources used for model parametrization, evaluation and implementation improves transparency [12]. These aspects can be covered more concisely for simpler models, resulting in a better understanding of a model, especially for non-modelers, which may increase the use of models in GHGIs. This is desirable as it provides the means for improving the overall quality of the inventory [12].

Consistency refers to the need to apply the same methodology over time and space, for example to national territory or homogenous strata, including consistent data sets. Ensuring consistency in the data sets over time becomes easier the fewer data are needed. The more data sets are used, the more difficult it is to ensure that they match regarding the spatial and temporal resolution, and quality [cf., 65]. Inconsistencies between data sets introduce a source of uncertainty in the estimates which is very difficult to quantify [67], partly due to their correlation.

Comparability requests the use of common methodologies and formats by Annex I Parties to the UNFCCC to ensure that estimates of the C balance reported in their GHGIs are comparable among the parties. This calls for approaches which are applicable across a range of different conditions, including environment and data constraints. Regarding Levins’ approach to characterize models, a high degree of generality is required. Generality implies that a model can be applied in different regions without modifications to the model structure itself including parameter values, which also ensures reproducibility.

Comparative studies that included two or more models in this article (Table 1) demonstrate the difficulty of model comparisons. For example, Palosuo et al. [65] found large differences in the estimated C stock dynamics that were due to different model input requirements and assumptions. Large-scale comparability may come at the price of reduced accuracy at finer scales. Figure 1 illustrates the challenge of ensuring comparability, based on data from a study with Yasso7 along an environmental gradient in Europe [63]. For two different model parameterizations based on a global data set [68] and a subset restricted to Scandinavian countries [61], the figure shows the range of temperature and precipitation for which the individual parameters may be valid. Such information provides a straightforward option to compare the effect of climate on a particular variable of interest. As an example, Figure 1 demonstrates this for three locations along the examined environmental gradient with respect to the humus decomposition rate.

To achieve completeness of a GHGI, separate estimates for C stock changes in litter, dead wood and soil in forests are required for reporting under the UNFCCC and KP [1]. A key requirement of reporting is to avoid double counting or omissions of changes in C pools. This is a particularly important concern for soil related pools and subpools, where transfers of organic matter are difficult to follow. The temporal variability in C stocks of these pools can be considerable due to differences in the production of litter and dead wood, and in the speed of decomposition and volatility of C in litter, dead wood and soil [cf., 10]. In order to represent the differential decay rates of these pools, they must be accounted for in a model. The decay of litter and dead wood depends particularly on the dimension [e.g., 56,57], chemical composition including lignin content [e.g., 58] and climate (temperature and moisture; [10]). The models included in this study differ in their approach to reproduce these aspects, which affects their realism (e.g., [62]; see also Table 3).

Following the IPCC guidelines, accuracy is obtained by ensuring that estimates have no systematic error and neither under- nor overestimate true C emissions or removals – that is, estimates are unbiased. This objective corresponds to Levins’ interpretation of sacrificing precision to realism and generality – that is, to emphasize the qualitative rather than the quantitative dimension of the result [13]. Evans [14] considers this approach a compromise between an exact solution and a distribution of likely values. Considering the many sources of uncertainties that model estimates of C cycling are confronted with [62], including uncertainties due to parameter, input and climate variability, obtaining exact (or very precise) estimates is highly unlikely. This is recognized in the IPCC guidelines, which acknowledge that uncertainty exists and encourage reducing it as much as possible.

Accounting for uncertainty can be achieved with techniques such as error propagation and Monte Carlo simulations [e.g., 62]. More challenging is the estimation of parameter uncertainty where different methods exist for different types of model complexity [cf., 69], since with increasing complexity more parameter and interrelated processes need to be considered.

The IPCC guidelines stipulate that all soil-related C pools are separated into C stored in soil organic matter, litter and dead wood. As discussed in the section on models, this differentiation is not made by all models, and parties to the UNFCCC and the KP use individual approaches to address the requirement by, for example, subsuming all pools in the soil C pool [e.g., 70] or using the source of input for separation [e.g., 71]. Whether these pools are reported as one or separately, it is paramount that a model is based on current scientific understanding of the processes involved [72]. The scientific soundness of a model is demonstrated by documenting and evaluating (including verification and validation; cf., [12]) a model.

Important considerations in GHG reporting under the UNFCCC and the KP are the effect of land-use change and of disturbance (natural or anthropogenic), which can result in changes to biomass production. In the ecosystem models CBM-CFS and CoupModel, the impacts on the living crops need to be accounted for by the model. For DOM and soil models, the C inputs need to reflect the changes occurring in the above-ground living trees. With respect to the C pools in dead wood, litter and soil, all models need to be able to account for the difference in the quality and quantity of C inputs resulting from land-use change and disturbance – for example, increased coarse woody debris production after windthrow. Furthermore, in a model the conservation of mass must be ensured – that is, that the fluxes equal the stock changes in the individual C pools. Among the examined models, only CBM-CFS includes a mechanism that at each time step controls for conservation of mass [73], which is particularly important for an ecosystem model with a large number of pools. These elements of QA/QC should be supported by a model.

Data constraints for model application and validation

Considering the application for obtaining C stock inventories at national or regional scales, the ease of model implementation on a (very) large number of sites is important to capture the variability and reduce uncertainty in the estimates. Hence, the selection of a model needs to consider data input requirements. The two primary considerations with regards to model application and validation are data availability (including temporal and spatial resolution, and consistency) and completeness (including data for all C pools, and additional information such as climate).

Although many countries have implemented NFIs or similar programs (e.g., the International Co-operative Programme on Assessment and Monitoring of Air Pollution Effects on Forests, or ICP-Forests; [74]) to estimate and monitor their forest resources, the number of repeated inventories varies considerably [8], resulting in differences in the time series with available data. The design of the inventory, including whether permanent or temporary plots are used, and grid spacing, affects the accuracy and the precision of the inventory-based estimates. Inconsistencies in an inventory, such as the switch from a stand-level to a plot-based design as in former Eastern Bloc nations, need to be considered for a model application.

The number of forest attributes which can be measured during an inventory is limited, and not all relevant C pools (i.e., all tree compartments including coarse and fine roots, stem, branches and foliage) are covered. In order to obtain a complete C budget of a forest, estimates for all tree compartments are needed and models may require relevant data for initializing the model to specific conditions, and to provide the amount of inputs per unit of time over the simulation length.

The volume or biomass compartments of trees and stands can be estimated based on measured tree attributes such as diameter at breast height (DBH). The accuracy of these estimates depends on the methodology applied to obtain values for the tree or stand biomass compartments – that is, biomass equations (BE) or biomass (expansion) factors (BF or BEF; [75]), and also on the availability and representativeness of BEs or BFs for a particular study region. Uncertainty associated with the estimates of the size of tree compartments propagates through the model and affects the model results [76]. This effect may be minimized by selecting an appropriate method [75] for the study region or, alternatively, from existing collections [e.g., 77,78]. In the case of ecosystem models such as CBM-CFS and CoupModel, where the inputs of litter and dead wood are emerging from the simulated forest development, the limitation is to ensure that the simulated production of dead wood and litter is realistic. If tree growth in ecosystem models is simulated based on yield-table data, which reflect growth conditions in the past, there is a risk that the estimates of dead wood and litter production do not match with reality, leading to biased C stock change estimates.

Given the recent efforts to harmonize inventories (e.g., COST FP0803 “Belowground carbon turnover in European forests”, COST E43 “Harmonisation of National Forest Inventories in Europe: Techniques for Common Reporting” [cf., 8] and COST FP1001 “Improving Data and Information on the Potential Supply of Wood Resources”), more efforts are necessary to harmonize model technics as well. With the efforts to harmonize forest inventory data, it can be expected that improved and comparable estimates of litter and dead wood inputs for model applications will become available. Improved observational data can, in turn, result in more accurate model prediction and allow for a comprehensive model validation. This development may be of particular advantage for countries with limited temporal or spatial data availability (which is, for example, the case in many former eastern European countries) as it would allow historical time series to be constructed. To capitalize on the harmonization and increased availability of measured data, guidance on the selection of a suitable modelling tool is valuable, especially in cases where no previous experience exists.

Forest inventories typically also do not provide data on soil conditions. Where soil inventories are available, estimates on soil C are associated with large uncertainties resulting from measurement limitations and small-scale spatial variability [e.g., 79]. The scarcity of data on soil C stocks generally presents limits for model development and verification. In cases of no or limited data availability, or inconclusive results (i.e., whether a pool is a sink or a source), a modelling approach can be used to support that a pool is not a source of C during KP reporting [4].

Besides information on C pools, models require additional data (Table 2), introducing further constraints especially with respect to attributes which are not as commonly available, such as soil properties or climate data at fine temporal resolution (days to months). For some attributes, default values are distributed with the model (cf., Table 2). Where available, defaults can be replaced with alternative data for the study region, which may improve the accuracy of the resulting estimates and may also reduce their uncertainty. These modifications do not compromise transparency and comparability as the model structure and parameters remain unchanged.

Data limitations may affect model applications in various ways depending on the particular model. Peltoniemi et al. [44] investigated this aspect in detail and concluded that for regional- or national-scale applications, data limitations are particularly problematic for models with high data requirements. The authors related this finding to spatial data constraints as some information may be available only locally. Although data may be processed (e.g., upscaling or interpolating) to fill existing gaps, this results in reduced accuracy and increased uncertainty in the model predictions [44].

Table 2 evaluates constraints for a model application that result from data limitations. The authors identified three broad data categories (i.e., meteorological information, C inputs and soil attributes) for which different considerations apply. As discussed by Peltoniemi et al. [44], data may not be available consistently in space or time. Expanding on the analysis by these authors, the temporal resolution of each model with regards to data availability was also considered. Meteorological information at annual and monthly resolution for common variables such as temperature and precipitation is readily available for most regions from national or international meteorological institutes including estimates of uncertainty (e.g., [80] for a global data set). Additional attributes such as evaporation and radiation as well as daily data may be derived from further processing [e.g., 81]. C inputs are derived based on living stock and application of biomass functions and turnover rates which are readily available [e.g., 78]. Data on living stock are generally available from NFIs [e.g., 8] or from simulations, as in the case of the examined ecosystem models. Sub-annual data are typically not available and need to be derived from NFI data, applying interpolation procedures. Data on dead wood and litter quality such as lignin content are available from literature [e.g., 41,82]. Soil-related data are often not available consistently across space and are limited regarding the completeness of measured attributes [e.g., 83,84].

Generally, the model simulation may be implemented to function with data constraints – however, at the cost of losses in accuracy, precision and comparability of C stock estimates. Different approaches to account for missing or poor-quality data may be used, such as aggregating plot data to small regions [e.g., 70], applying model or IPCC default values [1] or, in the case of DOM and soil models, the model may be linked to a forest growth model to replace missing time series, or extend the time series, of dead wood and litter input data. Such a model link is also used for forward projections in time for constructing the Forest Management Reference Level (FMRL; (cf., [4] and, e.g., [85]).

Implications for using models in GHG reporting

Based on the documentation and application of the selected models described in the literature and the evaluation of TCCCA reporting guidelines above, the authors attempted to place each model within Levins’ triangle (Figure 2; see also Table 3 for details). Figure 2 is the result of a qualitative assessment with reference to the discussed criteria for policy application (Table 3). The evaluation also considered data requirements as discussed previously (Table 2), and results of comparative studies (Table 1). Fitting the models in Figure 2 was based as much as possible on objective criteria from relevant studies and model documentations (Table 2 and section on model introduction), but it can only present an approximate range along each axis. The figure presents a conceptual approach to elucidate important considerations for model selection for policy application.

It was not surprising that models differed since they were all implemented to answer specific research questions. It was, however, discouraging that the majority of models leaned toward emphasizing precision and also realism over generality. The strength of such models, which are generally more complex, is that they can provide precise answers. On the other hand, these models cannot easily be applied across a range of environmental conditions as they require re-parameterization [49,65]. The spatial extent of a model is limited by parameter requirements, which generally increases with complexity. This limits the generality and derivation of estimates at larger scales.

Comparability and transparency, which are probably the most important criteria for policy applications [12], are more readily achieved with simpler and more generalist models. Based on the current evaluation, the minority of the candidate models focus on these aspects. Yasso07 and Q from among the group of DOM and soil models and the ecosystem model CBM-CFS address these particular criteria. These models, particularly the DOM and soil models Yasso07 and Q, are simpler with respect to their structure and require only few and readily available input data (Table 2). Besides a wider range of applicability, the models and the estimates can be more easily verified. This comes at the cost of precision at the local scale. The application of these models to a large number of sites, and aggregating results at regional or national scales, allows the distribution of C stocks and spatial uncertainty to be estimated. But this is only possible because the models are valid over a wide range of sites without the need for local parameterization. This approach avoids spatial bias [cf., 86] and achieves general good accuracy [69,65] – that is, unbiased estimates near the true value. For example, Yasso07 can be applied using generic parameter sets that have been published and are valid across a large environmental range (Figure 1).

Model parameters themselves present a source of uncertainty [e.g., 87] as they are derived by calibration based on fitting a model output to data. To estimate the uncertainty of parameter values, different approaches are used by the candidate models including Bayesian statistics and Markov chain Monte Carlo simulation (e.g., Yasso07), and generalized likelihood uncertainty estimation (e.g., Q). When parameters are derived in a consistent manner, a robust estimation of uncertainty related to model parameters is possible. The effect of parameter uncertainty on model results can be estimated using techniques such as Monte Carlo simulations. This requires that the uncertainty in the value of individual parameters and correlations between parameters are known and available for evaluation in the model simulation procedure.

Due to the trade-offs regarding precision, realism and generality that are inherent to a model [13], the application of a model is restricted and may be highly specialized. Such models aim for local accuracy (i.e., unbiased and near the true value) and are thus less suitable for applications with policy- and decision-making objectives. For such application, transparency and comparability are the principal criteria [12,88]. For a transparent approach to model selection and application, the model inputs and its range for application should be known. Hernández et al. [63] provided information for a first assessment of the applicability of the Yasso07 model along an extended environmental gradient from boreal forests in Finland to Mediterranean forests in Spain (cf. Figure 1). Such evaluations of models can improve the confidence in reported C balances and help countries to comply with reporting obligations under the UNFCCC and the KP. However, they do not eliminate the need for a case-specific assessment and verification since the suitability of a model for consistent implementation across space and time may be limited by a lack of data.

The aspect of comparability and validity of models is gaining more importance for reporting of emissions and removals from Forest Management in the Second Commitment Period of the KP, especially when considering model-based approaches for constructing the FMRL and the subsequent accounting. The model based approach that many parties chose (cf., Box 2.7.3 in [4]) calls for further harmonizing efforts. Palosuo et al. [65] identified a lack of research to improve models for more transparent and comparable applications, which the present authors can confirm based on the current evaluation. The authors surmise that few efforts have been made to evaluate and improve model applications, and that the speed and ease of applying simpler models outweigh their limitations regarding precision. The example of Yasso07 demonstrates that simpler models can ensure comparability (few and readily available data, large number of applications and references from published papers), transparency (documentation), completeness (accounting for dimension-dependent decomposition of dead wood; cf., [10]), consistency (time series) and accuracy (including estimates of uncertainty).

Conclusions

Dead wood and litter are volatile C pools releasing C to the atmosphere as well as to the soil, and together create the largest C store in forests. Especially at larger spatial scales like national inventories, field data are typically not collected at annual intervals. Regarding the increasing need for repeated and timely data on environmental attributes such as, for example, for the reporting under UNFCCC and KP, models have been recognized as alternatives to measured data. The utility of a model depends strongly on the requirements of the application such as producing qualitative realistic and comparable estimates at larger scales, or quantitative precise site-specific estimates. The selection of a model which is suitable for the research question at hand is not straightforward as the potential model user is faced with a number of possible models. In order to make an informed decision on model selection, clear information on the purpose and applicability of a particular model is missing. With respect to trade-offs among generality, realism and precision [13], an approach was discussed to classify several models with respect to their suitability for estimating the C balance in dead wood, litter and soils in forests, and to meet policy requirements.

For GHG reporting, models that are driven by the litter input from NFIs have the advantage that trends in litter are based on measured biomass stocks of trees. Assuming that the estimates of litter inputs are reliable, the combination of NFI data and C decomposition models can be expected to provide very accurate estimates of C fluxes in dead wood, litter and mineral soil. On the other hand, for future predictions ecosystem models are also needed that concurrently estimate biomass- and litter production.

In view of the requirements for model applications for C reporting, and the resulting challenges, the authors suggest that model developers prioritize the generality of a model rather than making their tools more complex. Such an approach would result in greater transparency and comparability, which are both important considerations for GHG reporting under UNFCCC and KP. Current computing resources may present little or no constraint for increasing model complexity, but scarcity of data limits the applicability of complex models. Although a user may be intrigued by the precision of complex models, the model selection should be based on the research question and the principle of parsimony – that is, selecting the simplest models from among the suitable alternatives.

With regards to monitoring and reporting C stocks, models should not be used as an alternative to measurements, but should be considered complementary. Since it may not be feasible to collect measurements at the national scale annually, inventories at coarser spatial and temporal resolution enhanced by models may present a compromise to achieve high levels of transparency, consistency, comparability, completeness and accuracy.

Acknowledgements

This article is an outcome of a project under the task “Input to improving the comparability in MRV across EU MS” within the LULUCF MRV project: “Analysis of and proposals for enhancing Monitoring, Reporting and Verification (MRV) of land use, land use change and forestry (LULUCF) in the EU” funded by the European Commission. L. Saint-André (INRA) was also supported by a grant overseen by the French National Research Agency (ANR) as part of the “Investissements d’Avenir” program (ANR-11-LABX-0002-01, Lab of Excellence ARBRE) – QLSPIMS project. M. Didion was funded by the Swiss Federal Office for the Environment. L. Hernández was funded by the EG-13-072 agreement between MAGRAMA and INIA.

Disclosure statement

No potential conflict of interest was reported by the authors.