Evaluation of forest carbon uptake in South Korea using the national flux tower network, remote sensing, and data-driven technology

Abstract

Forests provide most of the carbon sequestration of atmospheric carbon dioxide (CO₂); however, accurately quantifying the uptake amount over a region remains challenging. For reginal or national estimates, the forest productivity model and forest inventories are used which provide information for national greenhouse gas inventories. However, it has some limitations, such as not considering below-ground biomass, its lack of species-specific allometric models, the restrictions it places during fieldwork, and the long period it takes to complete the survey. In contrast to inventory-based biomass estimates, the eddy covariance (EC) method can assess net CO₂ exchange of a whole ecosystem continuously and automatically with a high temporal resolution. Since, these measurements only represent a site-level observation scale (∼ 1 km²), upscaling via linkages with observation data, remote sensing, and modeling methods has been used to estimate regional or national land-atmosphere carbon fluxes. In this study, we employ a data-driven method to estimate the national-scale gross primary production (GPP) and net ecosystem CO₂ exchange (NEE) by combining EC flux data from 10 sites in South Korea with remote sensing data through a machine learning algorithm based on support vector regression (SVR) for the period 2000–2018. Site-level evaluation of estimated GPP and NEE from the SVR-based model shows equivalent performance compared to other continental and global upscaled models. The mean estimated annual GPP and NEE of the South Korea forests region over the period 2000–2018 were 1465 ± 37 and −243 ± 32 g C m² year⁻¹, respectively. The SVR-based net primary production (NPP) was consistent with the biometric-based NPP (r² = 0.46, p < 0.05). This study shows that combining data from a national flux network and remote sensing using a data-driven approach can be used to estimate forest CO₂ fluxes on a national scale.

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 (R_a) (i.e., NPP = GPP – R_a) (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 (R_E) (i.e., NEE = GPP – R_E ≈ NPP – soil heterotrophic respiration (R_h)) (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 R_h 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?