Evaluation of forest carbon uptake in South Korea using the national flux tower network, remote sensing, and data-driven technology
Introduction
Mitigating elevated concentrations of atmospheric carbon dioxide (CO2) requires an understanding of the carbon cycle, including quantification of terrestrial ecosystem CO2 fluxes at local, regional, and global scales. Forests play an important role as CO2 sinks; they cover ∼30% of the land surface and account for 80% of the total gross primary production (GPP) of terrestrial ecosystems globally (Beer et al., 2010). In South Korea, forests cover ∼64% of the land area and account for a CO2 uptake of ∼47.0 million tons (KFS, 2019). Considering the international obligations introduced by the Paris Agreement to reduce greenhouse gas (GHG) emissions (UNFCCC, 2015), it is important and necessary to estimate the CO2 absorption accurately and quickly by forests for assessing their GHG reduction and submit annual national GHG inventory reports.
Inventory-based change estimation of biomass has been generally used at a regional or national scales to estimate the CO2 absorbed in forests (net primary production (NPP)) through the difference between photosynthesis (GPP) and whole-plant autotrophic respiration (Ra) (i.e., NPP = GPP – Ra) (Fang et al., 2001; Guo et al., 2010; Jenkins et al., 2003; Zhou et al., 2016). However, it has some limitations, such as its exclusion of non-tree biomass (shrubs, herbs, climbers, and grass), its assumptions regarding below-ground biomass, its lack of species-specific allometric models, the restrictions it places on fieldwork, and the biomass survey taking many years to complete. These limitations may lead to inaccurate estimates of biomass and forests carbon cycle, which may, in turn, bias the carbon uptake estimation. In addition, the deadwood, litter, and soil carbon pools are highly variable over space and time (e.g., seasonal change in litter and sudden changes due to disturbances, Prevost-Boure et al., 2010; Zhang et al., 2020), making it difficult to accurately assess the inventory of dead organic matters. Despite its importance, it is assumed, due to a lack of data that the soil carbon and dead organic matters do not change according to the Tier 1 method based on Intergovernmental Panel on Climate Change (IPCC) guidelines (2006).
To address such limitations of the forest inventory method, the eddy covariance (EC) technique can be used instead to quantify terrestrial carbon exchange, which uses the difference between CO2 uptake minus ecosystem respiration (RE) (i.e., NEE = GPP – RE ≈ NPP – soil heterotrophic respiration (Rh)) (Baldocchi et al., 2001; Wofsy et al., 1993). The EC method is widely used for estimating diurnal/seasonal/inter-annual variations, since it directly measures turbulent transport of mass and energy between forest ecosystems and the atmosphere (Aubinet et al., 2000; Baldocchi et al., 1988; Foken and Wichura, 1996). Furthermore, the IPCC guidelines also recommend EC as one of the technologies for determining overall net carbon exchanges, which can help determine forest carbon budgets and verify studies on carbon balances using the stock change methods (IPCC, 2006). However, these measurements only represent a site-level observation (at < 1 km2 scales) called the flux footprint (Gockede et al., 2008; Schmid, 1994). In addition, it is costly and requires a high level of training. For these reasons, the number of observations is limited. Therefore, to assess terrestrial carbon cycling over regions, EC measurements can be upscaled to a large area in conjunction with data from a meteorological observation network, remote sensing, and modeling (Ichii et al., 2017; Jung et al., 2009; Maselli et al., 2006, 2008; Papale and Valentini, 2003; Sasai et al., 2011; Xiao et al., 2014; Yang et al., 2007). As the number of EC observations increases, the EC network covers various land covers, climate regions, and species (Baldocchi, 2008). Such better coverage has accelerated interoperability of EC flux data with satellite remote sensing and land surface modeling, to assess regional and global CO2 budgets (Friend et al., 2007; Ichii et al., 2017; Jung et al., 2009; Stockli et al., 2008; Xiao et al., 2014).
Over the past two decades, data-driven upscaling technologies have been developed to estimate CO2 fluxes at larger scales using available regional EC tower observations (Ichii et al., 2017; Jung et al., 2009; Xiao et al., 2014, 2008; Yang et al., 2006; Zhang et al., 2011). In particular, a data-driven approach by machine learning (ML) (e.g., artificial neural networks (ANNs), random forests (RFs) along with other model tree ensembles (MTEs), and support vector regression (SVR)) is one of the most effective ways, because the crucial information in a large amount of data can be adaptively extracted, without any assumption for a given application. Thus, over the last decade, these techniques have been extensively used to extend information for a limited area to a wider space, based on the correlation between target flux data and its drivers. For example, Papale and Valentini (2003) estimated the European forest ecosystem CO2 fluxes by combining satellite data derived from an Advanced Very High Resolution Radiometer (AVHRR) with 16 European EC data and using an ANN. Yang et al. (2007) applied an SVR algorithm for the United States, thereby producing spatial and temporal variations in terrestrial GPP by combining Moderate Resolution Imaging Spectroradiometer (MODIS) and 33 EC sites across the US. Jung et al. (2009) proposed empirical upscaling using the MTE approach and global EC network data (i.e., FLUXNET), to perform predictions and introduce both a new model tree induction algorithm and a specific ensemble approach. In addition, upscaling approaches have been conducted for large regions in the United States (Xiao et al., 2008; Yang et al., 2007), Europe (Jung et al., 2008; Papale et al., 2015; Papale and Valentini, 2003), and Asia (Ichii et al., 2017; Yao et al., 2018). The ANNs can capture nonlinear complex relations from large-scale datasets, but they require large amount of training data and time (Papale and Valentini, 2003). The RFs are flexible enough to handle several input variables without variable selection and overfitting, but they require a substantial amount of time and computing resources to train and build numerous trees to combine their outputs (Bergen et al., 2019). Similarly, the MTEs use multiple tree models jointly applied for better performance, outperforming single trees including the reduction of extrapolation errors. However, model ensembles cost more to create, train, and are not always better. The effectiveness of an ensemble is correlated with the accuracy and diversity of the individual members that make up the trade-offs (Jung et al., 2009). As one of the ML methods, SVR has attracted significant attention due to its generalizability, ease of associated training, and versatility. Furthermore, SVR performs well on datasets with many attributes and little overfitting, even with limited amount of data for training (Burges, 1998; Mountrakis et al., 2011). Thus, it is more appropriate for application in Asia's coarse EC flux network areas, rather than in the United States or Europe with dense tower networks.
The Korea Forest Service has surveyed forest resources through the National Forest Inventory (NFI) to determine the forest carbon stock and forest resources trends, as well as to produce national forest statistics. Similarly, many countries regularly or irregularly estimate forest biomass using NFIs at the regional level. In South Korea, the NFI is conducted through field surveys of sampling plots (∼4000 sample points) deployed across the country in 5-year cycles. Although these methods may roughly represent the country-scale carbon stock, it is difficult for them to capture rapid changes in the ecosystems. To better understand these processes, the Korean regional network of EC tower sites (KoFlux) monitors the cycles of energy, water, and carbon dioxide between the atmosphere and the key terrestrial ecosystems over various timescales (e.g., from half-hours to years). In addition, the National Institute of Forest Science (NiFoS) has recently deployed six EC flux towers across the country's forests to improve the understanding of the forest climate and nutritional cycle, and to assess forest productivity. Therefore, this study focused on examining modeling capabilities using the national-based EC network data and one of the ML approaches which are emerging as highly efficient data-driven technologies.
In this study, we aim to evaluate the operability of the national EC flux observation network and its upscaling when combining it with remote sensing through SVR, to estimate forest CO2 fluxes on a national scale, as a complement and support to existing/traditional methods for estimating the national carbon inventory. This research on national carbon fluxes not only considered previously unused data at global and continental scales for data-driven approaches (Ichii et al., 2017; Kondo et al., 2017; Yao et al., 2018), but also addressed the lack of studies to estimate CO2 fluxes at a national scale using a dense domestic EC observation network. To address the discrepancy between up-scaled forest C fluxes and inventory-based biomass growth, we used the NPP by adding NEE to Rh datasets (Tang et al., 2020). We (1) investigated the representativeness of the EC flux observation data in South Korea from the Korea Flux Monitoring Network (KoFlux) and the National Institute of Forest Science (NiFoS); and (2) estimated CO2 flux by a data-driven approach based on SVR using 49 site-years of observation data and evaluated on site and national scale. Next, we discussed the following specific questions: (1) What is difference between the estimated forest CO2 fluxes of South Korea based on upscaling using data from national networks, compared to those previously upscaled based on data from continental and global networks (i.e., FLUXCOM (Jung et al., 2017; Tramontana et al., 2016), SVR (Ichii et al., 2017), MODIS (Zhao et al., 2005) and BESS (Jiang and Ryu, 2016; Ryu et al., 2011))?; and (2) What is the difference between the estimated CO2 fluxes obtained by the SVR model, compared with estimates obtained by the inventory method?
Section snippets
Target region and flux observations
We used the EC observation data from 10 forest flux-tower sites in South Korea that were available from 2006 to 2018 (Fig. 1a). The network across South Korea covers a variety of climatic and topographical characteristics that arise from a complex mountainous terrain. The network also spans a wide variety of plant functional types (PFTs) according to the land cover classification of the International Geosphere Biosphere Program (IGBP), such as evergreen needleleaf forest (ENF), evergreen
Site-level model evaluation against observation
The GPP and NEE predictions at the 8-day temporal scale showed that all data-driven SVR models we set up performed reasonably well at the site level (supplementary Table S2). For GPP, the r2 values ranged from 0.75 to 0.80, while the RMSE values ranged from 1.11 to 1.43 g C m−2 day−1. For NEE, the r2 values ranged from 0.65 to 0.69, while the RMSE value ranged from 0.78 to 0.84 g C m−2 day−1. The slopes of the regression between the predicted and observed values of GPP ranged from 0.89 to 1.06,
Differences between SVR and other estimates
In this study, we used a data-driven approach based on local flux network data to estimate GPP and NEE of South Korean forests, and found a good consistency with the other models from the previous study in terms of temporal and spatial variations (Figs. 5 and 6). The mean annual GPP of SVR is lower for SVRASIA (∼2%) and higher than for FLUXCOM (∼9%). This seems to be largely caused by different EC data sets being used for training models. FLUXCOM included only two South Korean sites of the 224
Conclusions
In this study, using the 10 sites EC flux observations data over 2006–2018 in South Korea (KoFlux and NiFoS) and a data-driven approach, we upscaled the GPP and NEE at 8-day temporal resolution from site-level to regional scale (whole nation) based on an SVR model. The geographical location of EC flux sites was skewed to the north and did not exist the inland, but was uniformly distributed across the PFTs and climate zones. The SVR-estimated CO2 fluxes not only show high accuracy when validated
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgements
This work was supported by the R&D Program for Forest Science Technology [Project No. 2017099A00-1919-BB01] provided by Korea Forest Service (Korea Forestry Promotion Institute), and is a Joint Research Program of CEReS, Chiba University (2018). Our thanks go out to all the members of KoFlux who have committed and dedicated for continuous data collection and site managements.
References (136)
- et al.
Tree spatial patterns and stand attributes in temperate forests: the importance of plot size, sampling design, and null model
For. Ecol. Manage.
(2018) - et al.
The role of the Baekdudaegan (Korean Peninsula) as a major glacial refugium for plant species: a priority for conservation
Biol. Conserv.
(2017) The application of a weighted infrared-red vegetation index for estimating leaf-area index by correcting for soil-moisture
Remote Sens. Environ.
(1989)- et al.
Quantifying allometric model uncertainty for plot-level live tree biomass stocks with a data-driven, hierarchical framework
For. Ecol. Manage.
(2016) Cross-site evaluation of eddy covariance GPP and RE decomposition techniques
Agric. For. Meterol.
(2008)- et al.
Tools for quality assessment of surface-based flux measurements
Agric. For. Meterol.
(1996) - et al.
Relative humidity effects on water vapour fluxes measured with closed-path eddy-covariance systems with short sampling lines
Agric. For. Meterol.
(2012) NDWI - A normalized difference water index for remote sensing of vegetation liquid water from space
Remote Sens. Environ.
(1996)- et al.
Multi-year convergence of biometric and meteorological estimates of forest carbon storage
Agric. For. Meterol.
(2008) - et al.
Inventory-based estimates of forest biomass carbon stocks in China: a comparison of three methods
For. Ecol. Manage.
(2010)
Addressing a systematic bias in carbon dioxide flux measurements with the EC150 and the IRGASON open-path gas analyzers
Agric. For. Meterol.
A Cubesat enabled Spatio-Temporal Enhancement Method (CESTEM) utilizing Planet, Landsat and MODIS data
Remote Sens. Environ.
Overview of the radiometric and biophysical performance of the MODIS vegetation indices
Remote Sens. Environ.
Geographical variations in gross primary production and evapotranspiration of paddy rice in the Korean Peninsula
Sci. Total Environ.
Multi-scale evaluation of global gross primary productivity and evapotranspiration products derived from Breathing Earth System Simulator (BESS)
Remote Sens. Environ.
Modification of the moving point test method for nighttime eddy CO2 flux filtering on hilly and complex terrains
MethodsX
Comprehensive synthesis of spatial variability in carbon flux across monsoon Asian forests
Agric. For. Meterol.
Use of remotely sensed and ancillary data for estimating forest gross primary productivity in Italy
Remote Sens. Environ.
Support vector machines in remote sensing: a review
Isprs. J. Photogramm.
Global products of vegetation leaf area and fraction absorbed PAR from year one of MODIS data
Remote Sens. Environ.
Comparison of the eddy covariance and automated closed chamber methods for evaluating nocturnal CO2 exchange in a Japanese cypress forest
Agric. For. Meterol.
Testing the generality of below-ground biomass allometry across plant functional types
For. Ecol. Manage.
Evaluation of land surface radiation balance derived from moderate resolution imaging spectroradiometer (MODIS) over complex terrain and heterogeneous landscape on clear sky days
Agric. For. Meterol.
Satellite-driven estimation of terrestrial carbon flux over Far East Asia with 1 km grid resolution
Remote Sens. Environ.
Ecosystem photosynthesis inferred from measurements of carbonyl sulphide flux
Nat. Geosci.
Estimates of the annual net carbon and water exchange of forests: the EUROFLUX methodology
Adv. Ecol. Res.
Canopy near-infrared reflectance and terrestrial photosynthesis
Sci. Adv.
Breathing of the terrestrial biosphere: lessons learned from a global network of carbon dioxide flux measurement systems
Aust. J. Bot.
FLUXNET: a new tool to study the temporal and spatial variability of ecosystem-scale carbon dioxide, water vapor, and energy flux densities
Bull. Am. Meteorol. Soc.
Measuring biosphere-atmosphere exchanges of biologically related gases with micrometeorological methods
Ecology
Factors controlling long- and short-term sequestration of atmospheric CO2 in a mid-latitude forest
Science
Present and future Koppen-Geiger climate classification maps at 1-km resolution
Sci. Data
Terrestrial gross carbon dioxide uptake: global distribution and covariation with climate
Science
Machine learning for data-driven discovery in solid Earth geoscience
Science
A global database of soil respiration data
Biogeosciences
Addressing the influence of instrument surface heat exchange on the measurements of CO2 flux from open-path gas analyzers
Global Change Biol.
A tutorial on Support Vector Machines for pattern recognition
Data Min. Knowl. Disc.
Age, allocation and availability of nonstructural carbon in mature red maple trees
New Phytol.
The ecology and economics of storage in plants
Annu. Rev. Ecol. Syst.
Towards a worldwide wood economics spectrum
Ecol Lett
Updated generalized biomass equations for North American tree species
Forestry
Seasonal fluxes of carbonyl sulfide in a midlatitude forest
Proc. Natl. Acad. Sci. USA,
Biases in open-path carbon dioxide flux measurements: roles of instrument surface heat exchange and analyzer temperature sensitivity
Agric For Meteorol
Nonstructural carbon in woody plants
Implications of allometric model selection for county-level biomass mapping
Carbon Balance Manag.
Changes in forest biomass carbon storage in China between 1949 and 1998
Science
FLUXNET and modelling the global carbon cycle
Global Change Biol.
Estimating photosynthetically available radiation at the ocean surface from ADEOS-II Global Imager data
J. Oceanogr.
On the blending of the Landsat and MODIS surface reflectance: predicting daily Landsat surface reflectance
IEEE Trans.Geosci. Remote Sens.
Phenological response to climate change in China: a meta-analysis
Global Change Biol.
Cited by (18)
Improving farmers’ livelihoods through the eco-compensation of forest carbon sinks
2024, Renewable and Sustainable Energy ReviewsChina's National Reserve Forest Project contribution to carbon neutrality and path to profitability
2024, Forest Policy and EconomicsTwo decades of carbon dynamics in an actively-managed, naturally-regenerated longleaf/slash pine forest
2023, Forest Ecology and ManagementRevisiting vegetation activity of Mongolian Plateau using multiple remote sensing datasets
2023, Agricultural and Forest MeteorologySuper resolution of historic Landsat imagery using a dual generative adversarial network (GAN) model with CubeSat constellation imagery for spatially enhanced long-term vegetation monitoring
2023, ISPRS Journal of Photogrammetry and Remote Sensing