A comparison of approaches to quantify carbon for ecosystem service assessments through time

Field-based approaches are designed to take representative in-situ measurements of carbon in the field for a given study area (Table 1). Different field survey techniques are used to estimate carbon stocks in key pools (e.g., soils, aboveground and belowground vegetation) and fluxes between pools (e.g., soil respiration). Field-based monitoring requires a study design that considers appropriate levels of replication and statistical power for trend detection (Larsen et al. 2001; Lindenmayer and Likens 2010). Uncertainty in carbon estimates can be assessed through subsequent statistical analysis of the field data.

Direct field observations can have lower uncertainty compared to other approaches, although accuracy will depend on sampling intensity, especially in areas with high spatial variability of carbon (Eigenbrod et al. 2010). Therefore, it is important to consider whether a trend can be detected based on the sampling design that was carried out (Larsen et al. 2001). If the number of field survey samples is not large enough to accurately reflect the value of the population (e.g., actual carbon value across a study area) due to time or cost limitations in collecting data, then field surveys can be coupled with remotely sensed data for more comprehensive spatially continuous carbon estimates across larger areas and longer time periods; see Sections II, III, and IV (Houghton et al. 2001; Goetz et al. 2009). Of course, all methods would require enough field survey samples to capture carbon variability across key environmental gradients.

Below we focus on field measurements needed to assess (1) carbon stocks within key pools (e.g., soils and aboveground and belowground vegetation) and (2) carbon fluxes between pools (e.g., soil CO₂ efflux and net ecosystem CO₂ exchange).

The stratify and multiply approach relies on first stratifying the study area into multiple land cover or ecosystem types (Goetz et al. 2009; Table 1). A representative reference carbon stock value is assigned to each cover type. To estimate total carbon stocks, the area of each cover type is multiplied by the associated reference carbon stock value (Goetz et al. 2009). If reference values are only available for ecosystems, then the reference value can be multiplied by stock change factors, which are factors reflecting carbon changes due to land use (e.g., Bai et al. 2020) or management (e.g., VandenBygaart et al. 2004, 2008; IPCC 2006).

Carbon sequestration can be estimated in two ways, either by differencing two carbon storage maps built using LULC information from two time periods (stock-difference method, e.g., Sharp et al. 2020) or by using a gain-loss method to account for carbon losses through emission factors and carbon gains through removal factors (e.g., Harris et al. 2021).

It is relatively easy to obtain the required data for the stratify and multiply approach with low financial costs; however, this approach usually relies on simple land cover maps and therefore does not often account for carbon stored in different ecosystem cover types due to species composition, management, intact vs. degraded forests, successional differences (e.g., stand age), elevation, site history (e.g., legacy effects), and climatic zone (Sharp et al. 2020). This limitation can be overcome by using more detailed land cover maps that include additional classes (GFOI 2016; Sharp et al. 2020). The stratify and multiply approach also rarely accounts for global change drivers other than land cover change, the effect of time lags, and novel or mixed ecosystem types.

The largest sources of uncertainty for the stratify and multiply approach come from (1) land cover misclassification errors and (2) variability in reference carbon values. For example, when differencing two carbon storage maps built using this approach, the expected carbon stock changes may not be greater than the errors introduced from misclassifications in the LULC maps over time (Goetz et al. 2009). Also, there may be large spatial variability of carbon stocks within each land cover type; however, variability is typically reduced when using local site-specific field measurements to derive reference carbon stocks and stock change factors instead of default values from the IPCC (2006). Uncertainty in the carbon estimates can be calculated by combining uncertainties in the inputs (e.g., LULC data, reference carbon stocks, stock change factors) using error propagation (IPCC 2006). A Monte–Carlo analysis can also be used to assess uncertainty in carbon estimates when data are available to account for variability in each input value (e.g., VandenBygaart et al. 2004).

Many ecosystem service modeling platforms rely on the stratify and multiply approach, including the Integrated Valuation of Ecosystem Services and Tradeoffs, InVEST (Sharp et al. 2020), and the ARtificial Intelligence for Ecosystem Services, ARIES (Martínez-López et al. 2019). See supplementary material for examples of the stratify and multiply approach for monitoring carbon across ecosystems (Bai et al. 2020) and in agricultural (VandenBygaart et al. 2004, 2008), wetland (Ducks Unlimited Canada 2015; Euliss et al. 2006), and coastal systems (Filbee-Dexter and Wernberg 2020).

The direct remote sensing approach relies on the integration of remotely sensed observations with field measurements of carbon, linked through a statistical or machine learning algorithm (Table 1). For example, field-based measurements are used to calibrate (e.g., train) a model that relates the field measurements of biomass or soil carbon to environmental conditions captured by remotely sensed imagery. Once these relationships have been determined, the model can then be used to predict biomass or soil carbon in areas with remotely sensed imagery, but without field observations. The direct remote sensing approach is useful to obtain fine-scale (e.g., 30 m, 250 m) spatially continuous carbon stock maps through time. Using a stock-difference method, these carbon storage maps can be differenced to obtain carbon sequestration (e.g., Xu et al. 2021). Unlike the stratify and multiply approach, statistical models can explicitly account for uncertainty in carbon estimates (e.g., through prediction intervals).

Compared to the stratify and multiply approach, the direct remote sensing approach can capture the spatial heterogeneity of carbon within land cover classes and in ecosystems that are novel, mixed, or in transition, which can reduce uncertainty (Goetz et al. 2009). The Direct Remote Sensing Approach can also estimate pixel level and fine-scale carbon stock changes (e.g., due to degradation or afforestation, Goetz et al. 2009), whereas the stratify and multiply approach typically only captures carbon changes due to land cover change (United Nations et al. 2022). Using local site-specific field measurements for calibration and validation will improve accuracy and lower uncertainty (Sothe et al. 2022).

A key limitation of this approach occurs when models are calibrated using sparse or clustered field sampling data that do not capture the spatial variability of carbon across key environmental gradients (Ploton et al. 2020; Ludwig et al. 2023). In these cases, if random subsets are selected when performing cross-validation (e.g., nonspatial cross-validation), then overfitting may occur, and prediction accuracy may be overly optimistic because spatial autocorrelation is ignored (e.g., neighboring sampling plots are not independent and yet are included in the withheld testing subsets; Ploton et al. 2020). Alternatively, spatial cross-validation can reduce overfitting and lead to more reliable accuracy assessments by selecting folds in a way where folds are spatially distant at a level that reflects the distance between field sampling locations and prediction locations across the whole study area (Schwantes et al. 2016; Ludwig et al. 2023). Also, predictions should be limited to areas that are well represented by the field data's environmental variable space. For example, an area of applicability can be determined by only predicting within the range of the observed environmental variable space of the field data or by using a method that also accounts for gaps within the environmental variable space of the field data (e.g., Ludwig et al. 2023).

In comparison to optical sensors, active sensors such as light detection and ranging (LiDAR) and synthetic aperture radar (SAR) are more sensitive to vegetation structure and biomass. LiDAR emits laser pulses typically in the near-infrared range to estimate vegetation structure and height with high accuracy; however, due to cost constraints, wall-to-wall coverage across large areas is often not possible, and as such, LiDAR data are most useful for extending field data (Hyde et al. 2007; Goetz and Dubayah 2011). For example, data from space-based LiDAR can systematically detect forest structure across a broader range of forest types to fill in gaps created by sparse or clustered field data (Goetz and Dubayah 2011). LiDAR can also accurately detect changes in elevation; for example, multi-temporal LiDAR estimates of elevation change following a wildfire show promising results in capturing soil carbon loss, as demonstrated for a peatland in the United States (Reddy et al. 2015). Another active sensor, SAR, transmits short microwave pulses to measure vegetation structure even in cloudy conditions and is often available across broader regions compared to LiDAR (Flores-Anderson et al. 2019).

Each sensor has its own strengths and limitations, and as such, integration of multi-sensor data (e.g., optical, SAR, LiDAR) often improves accuracy (Hyde et al. 2007; De Sy et al. 2012). When relying on optical imagery or SAR, saturation can occur where models under-predict biomass in areas of high biomass (Gibbs et al. 2007). Also, maps derived using the direct remote sensing approach are often only available for snapshots in time (e.g., Hengl et al. 2017; Spawn et al. 2020; Sothe et al. 2022). However, advances in analysis-ready data, cloud computing, and automated algorithms have led to this approach being increasingly implemented over time (e.g., annual biomass maps from 1984 to 2016: Matasci et al. 2018). See supplementary materials for examples of the direct remote sensing approach for monitoring carbon across ecosystems (Hengl et al. 2017; Spawn et al. 2020; Sothe et al. 2022) and in agricultural systems (Sanderman et al. 2017; Gholizadeh et al. 2018), peatlands (Reddy et al. 2015), and forests (Beaudoin et al. 2018; Matasci et al. 2018).

The Mass Balance Approach relies on models that account for all key carbon pools and flows between them over time (GFOI 2016; Table 1). These models can account for carbon flows due to natural processes (e.g., growth, decay) and anthropogenic drivers of change (e.g., urbanization, climate change, management, LULC change, harvest, wildfire). These models often require user expertise, can be computationally intensive, and are not always spatially explicit, which can limit their use when fine-scale spatially explicit carbon estimates are required. For example, output carbon estimates can either be spatially explicit where values are known at exact locations (e.g., pixels, stands) or spatially referenced where values are unknown at exact locations but rather aggregated to coarser scale analysis units (e.g., ecoregions, administrative regions) (GFOI 2016). Mass balance (e.g., bookkeeping, conservation of mass) models can be (1) empirical (e.g., inventory-based), (2) process-based, or (3) hybrid models incorporating empirical and process-based components (GFOI 2016). Process-based carbon models can more accurately project changes in carbon under novel conditions, but they have high data requirements for calibration, are computationally intensive, and can be overly general (Sleeter et al. 2022). In contrast, inventory-based models (e.g., empirical models) usually rely on field measurements to derive carbon flux rates across pools over time (Stinson et al. 2011; Sleeter et al. 2022). As such, inventory-based models are less computationally intensive and can generate spatially explicit fine-scale estimates of carbon (Shaw et al. 2021; Sleeter et al. 2022).

For all model types, using local site-specific field measurements for model calibration and validation will reduce uncertainty. Furthermore, recent studies have also improved on the integration of remote sensing and forest carbon modeling (Zhou et al. 2021). For example, the generic carbon budget model (GCBM) and the land use and carbon simulator (LUCAS) use spatially explicit input data, including time series of forest disturbances derived from remote sensing, for more accurate accounts of the effects of disturbance (e.g., wildfire, harvest) and LULC change on carbon (e.g., Bona et al. 2020; Shaw et al. 2021; Zhou et al. 2021; Sleeter et al. 2022). Multiple types of uncertainty can be accounted for, including overall uncertainty in the carbon estimates due to uncertainties in the input parameters, using a Monte–Carlo approach (e.g., Hirsch et al. 2004).

Different models may include different drivers. For example, to assess changes in boreal peatland carbon stocks in Canada, modeling approaches include (1) the Canadian model for peatlands, which accounts for wildfire disturbances but not permafrost thaw (Bona et al. 2020), (2) dynamic global ecosystem models that incorporate permafrost dynamics (Chaudhary et al. 2017), and (3) inventory-based models (Hugelius et al. 2020) that can also account for abrupt permafrost thaw (Turetsky et al. 2020). Also, most modeling studies focus on one ecosystem (e.g., forests, peatlands); however, modeling across ecosystems can account for flows between ecosystems and avoid double counting (Daniel et al. 2018; Sleeter et al. 2019). See supplementary materials for examples of the mass balance approach for monitoring carbon across ecosystems (Daniel et al. 2018; Sleeter et al. 2019), and in agricultural systems (Smith et al. 2000), peatlands (Chaudhary et al. 2017; Bona et al. 2020; Hugelius et al. 2020; Turetsky et al. 2020) and forests (Fig. 1; Kurz et al. 2009; Stinson et al. 2011; Shaw et al. 2021).

A researcher's goals for carbon monitoring will shape the selection of an approach, metric, and level of assessment. Figure 3 synthesizes key questions relevant for selecting and employing the most useful approach towards monitoring carbon for a given question or decision. Table S1 in the supplementary material provides example studies for each monitoring question in Fig. 3. Field-based sampling may be sufficient if a researcher or decision-maker does not require spatially continuous carbon estimates over large regions. However, if time or cost constraints limit field data collection, then linking field-based measurements with remote sensing or land cover data can expand carbon monitoring across larger regions and longer time periods. Methods often focus on one ecosystem; however, methods accounting for the flow of carbon across ecosystems will avoid double counting. It is also important to select an approach that captures all relevant drivers for a study area and time period. For example, LULC might be the dominant driver of change at short time scales; however, climate change may become a relevant driver at longer time scales.

To demonstrate the process of selecting a monitoring approach, we consider the goal of monitoring soil organic carbon storage in an agriculture system. We synthesize the questions a researcher or decision-maker might face in selecting a monitoring approach using a decision tree (Fig. 4). Decisions for each question depend on many factors, including dominant drivers of change, spatial extent, time period, data availability, and user expertise. Field surveys and soil experiments can be designed to capture all relevant drivers of change; however, some approaches for scaling up to a landscape/region are not well-suited to account for all relevant drivers. For example, a monitoring method designed to capture changes in carbon related to land use change will not capture changes in carbon related to management or climate change. Therefore, it is important to assess which drivers of change are most relevant for a given study area and time period, because a driver can be context- and scale-dependent. For example, in urban and peri-urban areas at short time scales, LULC may be the dominant driver. The next decision will depend on the availability of data. If a researcher only has a few field measurements for each land cover type, then a stratify and multiply approach could be selected (e.g., Sharp et al. 2020). However, when more field measurements are available within and across cover types and replicated across key environmental gradients, then a direct remote sensing approach could be selected (e.g., Hengl et al. 2017; Sanderman et al. 2017). For direct remote sensing approaches (see Section III), accuracy should be assessed using spatial cross-validation, and prediction should only occur in areas within the range of the observed environmental variable space of the field measurements used to calibrate the model. If agricultural management is identified as a key driver, then the monitoring approach could be adapted to capture carbon changes related to management (e.g., VandenBygaart et al. 2004, 2008). At longer time scales, climate change may become a relevant driver, which may require a process-based model through a mass balance approach (e.g., CENTURY, Smith et al. 2000) as well as validation and calibration data from long-term soil-warming experiments to better account for the complexity of how soil carbon will respond to climate change (Richter et al. 2007).

No matter the selected approach, all studies should consider the stability, longevity, and resilience of the ecosystem storing carbon (Keith et al. 2021). Carbon storage resilience could be maintained through a combination of (1) rapid recovery of carbon to pre-disturbance levels and (2) resistance of carbon stocks against future disturbances. Recovery following a disturbance will depend on many factors (e.g., disturbance type, severity, and frequency; ecosystem type; species composition; species life-history traits; local climate; and soil properties; Bartels et al. 2016). For example, grassland carbon stocks following drought took only a few years to recover to a pre-disturbance state (Fu et al. 2017), boreal forest carbon stocks following wildfire took ≥100 years to recover (Palviainen et al. 2020), whereas disturbance of peatland carbon stocks would take centuries to recover, and as such are often considered irrecoverable at timescales needed to avoid climate impacts (Harris et al. 2022).

An area of active research concerns whether the management and conservation of resilient and diverse ecosystems could aid in maintaining stable rates of carbon stock increases and stable carbon stocks resilient to disturbances. Forests with greater tree species richness tend to have both higher productivity (Liang et al. 2016; Mori et al. 2021) and stability of productivity over time (Jucker et al. 2014). For example, the temporal stability of aboveground wood production was greater in mixed-species forests relative to monocultures (Jucker et al. 2014). Furthermore, Osuri et al. (2020) found that the temporal stability of a vegetation index was greater in species-rich natural forests compared to monodominant plantations, especially during drought years. Ecosystems with greater diversity often have higher resilience to disturbances, including insects (Jactel and Brockerhoff 2007) and drought (Anderegg et al. 2018). Therefore, accounting for biodiversity in addition to carbon storage and sequestration will support the resilience of carbon stocks and stock increases even in the face of increased disturbances under continuing climate change. The Essential Biodiversity Variables (EBVs) developed by the Group on Earth Observations Biodiversity Observation Network (GEO BON) could provide valuable information on the state of biodiversity (Navarro et al. 2017), including species or functional richness, which could be a useful predictor for forest resilience (Messier et al. 2019; Aquilué et al. 2020), and in turn, forest carbon storage and sequestration stability.

After selecting and applying an approach, researchers should follow the Findable, Accessible, Interoperable, and Reusable (FAIR) guiding principles (Wilkinson et al. 2016) to ensure transparency, reproducibility, and reusability of carbon monitoring products (e.g., data, models, tools, and algorithms). Hence, interoperability can be improved by facilitating information exchange through a high-level synthesis of the strengths, limitations, and best practices of four key carbon monitoring methods. Our synthesis could help researchers and decision-makers understand the key metadata and methods behind open-source carbon datasets and the appropriate conditions for the re-use of open-source models and algorithms. However, this is only a first step; enhancing interoperability would require many additional initiatives, including fostering a culture of data sharing, data transparency, reproducibility, and collaborations for sharing technology, protocols, and expertise (Vargas et al. 2017; Powers and Hampton 2019).