Understanding the sequestration of organic carbon (C) in agroecosystems is of primary importance for greenhouse gas (GHG) accounting in managed lands, to reduce the environmental footprint of land use, and inform crediting programs. However, a broader application of precise C accounting is currently constrained by a limited number of direct flux measurements. Aside well-studied ecosystems via the eddy covariance technique (EC), many still bear significant uncertainty. In this study, we propose and evaluate a method for estimating accumulated C stocks in agricultural sites, by assessing the plant aboveground carbon (AGC) throughout two growing seasons using unstaffed aerial vehicles (UAV) and machine learning (ML) regression methods. Then, we used these estimates to assess total plant C, and benchmarked it with CO fluxes derived from the eddy covariance method from the ICOS DK-Vng site in Denmark. We utilized a light detection and ranging (LiDAR) sensor onboard a UAV to derive the structural characteristics of crops, and we conducted in parallel destructive field-based measurements of AGC. Then, we designed a ML pipeline to provide estimates of AGC as a supervised regression problem, using the LiDAR-derived point cloud data to extract predictive features and the AGC labels as ground-truth target values. The best performing ML model attained predictions of = 0.71 and = 0.93 at spatial resolutions of 1 m and 2 m, respectively. The C content in the aboveground plant components was assessed via laboratory analysis (46.6 ± 0.3% of C-to-biomass in barley and 47.7 ± 0.3% in wheat), while the belowground components (root allocation and rhizodeposition) were estimated based on a phenology-dependent allometric ratio. The cumulative value of C uptake along the growing season (i.e. net primary productivity) was compared with the difference of C predictions between UAV-LiDAR survey dates, finding an optimal disagreement between methods below 9% in two different cereal crops. The plant carbon budget in croplands, determined through UAV-LiDAR and machine learning regression, aligns with the carbon ecosystem uptake estimated through the eddy covariance technique, showcasing comparable results. Thereby, the proposed method also demonstrates the potential to estimate cumulative CO fluxes in areas lacking direct eddy covariance measurements. Various experimental setups are evaluated as well as the sources of uncertainty resulting from the sampling design.

中文翻译：

---

农田净碳吸收的独立估计：无人机-激光雷达和机器学习与涡流协方差


了解农业生态系统中有机碳 (C) 的封存对于管理土地中的温室气体 (GHG) 核算、减少土地使用的环境足迹并为信贷计划提供信息至关重要。然而，精确碳核算的更广泛应用目前受到直接通量测量数量有限的限制。除了通过涡流协方差技术（EC）对生态系统进行充分研究外，许多生态系统仍然存在很大的不确定性。在这项研究中，我们提出并评估了一种估算农业场所累积碳储量的方法，通过使用无人值守飞行器（UAV）和机器学习（ML）回归方法评估两个生长季节的植物地上碳（AGC）。然后，我们使用这些估计值来评估总工厂 C，并将其与丹麦 ICOS DK-Vng 站点的涡流协方差方法得出的 CO 通量进行基准比较。我们利用无人机上的光探测和测距 (LiDAR) 传感器来获取农作物的结构特征，并并行进行 AGC 破坏性现场测量。然后，我们设计了一个 ML 管道来提供 AGC 的估计作为监督回归问题，使用 LiDAR 派生的点云数据来提取预测特征和 AGC 标签作为地面实况目标值。性能最佳的 ML 模型在 1 m 和 2 m 的空间分辨率下分别获得了 = 0.71 和 = 0.93 的预测。地上植物成分中的碳含量通过实验室分析进行评估（大麦中碳生物量的 46.6 ± 0.3%，小麦中碳生物量的 47.7 ± 0.3%），而地下成分（根分配和根际沉积）则根据物候依赖的异速生长比率。整个生长季节（即生长季节）吸收碳的累积值 将净初级生产力）与 UAV-LiDAR 调查日期之间的 C 预测差异进行比较，发现两种不同谷类作物的方法之间的最佳分歧低于 9％。通过无人机激光雷达和机器学习回归确定的农田植物碳预算与通过涡流协方差技术估计的碳生态系统吸收量一致，显示出可比较的结果。因此，所提出的方法还证明了在缺乏直接涡流协方差测量的区域估计累积二氧化碳通量的潜力。评估各种实验设置以及抽样设计产生的不确定性来源。