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.

森林提供了大气二氧化碳 (CO ₂ ) 的大部分碳固存；然而，准确量化一个区域的吸收量仍然具有挑战性。对于区域或国家估计，使用森林生产力模型和森林清单，为国家温室气体清单提供信息。然而，它也有一些局限性，例如没有考虑地下生物量、缺乏特定物种的异速生长模型、田间工作的限制以及完成调查需要很长时间。与基于清单的生物量估计相反，涡度协方差 (EC) 方法可以评估净 CO ₂以高时间分辨率连续自动交换整个生态系统。由于这些测量仅代表站点级观测尺度（~ 1 km ²），因此通过与观测数据、遥感和建模方法的联系进行了升级，已被用于估算区域或国家陆地-大气碳通量。在本研究中，我们采用数据驱动的方法来估计全国范围的初级生产总值 (GPP) 和净生态系统 CO ₂交换（NEE）通过基于支持向量回归（SVR）的机器学习算法将来自韩国 10 个站点的 EC 通量数据与遥感数据相结合，在 2000-2018 年期间。与其他大陆和全球升级模型相比，基于 SVR 的模型对估计的 GPP 和 NEE 进行的站点级评估显示出等效的性能。2000-2018 年期间韩国森林地区的平均估计年 GPP 和 NEE 分别为 1465 ± 37 和 -243 ± 32 g C m ² year ^-1。基于 SVR 的净初级生产 (NPP) 与基于生物特征的 NPP 一致 ( r ²  = 0.46, p < 0.05)。该研究表明，使用数据驱动方法将来自国家通量网络的数据与遥感数据相结合，可用于估算国家范围内的森林 CO ₂通量。