Canada has extensive forests and peatlands that play key roles in global carbon cycle. Canadian soils and peatlands are assumed to store approximately 20% of the world’s soil carbon stock. However, the spatial and vertical distributions of the soil organic carbon (SOC) concentration in Canada have not been systematically investigated. SOC concentration, expressed in g C kg⁻¹ soil, affects the chemical and physical properties of the soil, such as water infiltration ability, moisture holding capacity, nutrient availability, and the biological activity of microorganisms. In this study, we tested a three dimensional (3D) machine learning approach and 40 spatial predictors derived from 20 years of optical and microwave satellite observations to estimate the spatial and vertical distributions of SOC concentration in Canada in six depth intervals up to 1 m. A quantile regression forest method was used to build an uncertainty map showing 80% of prediction intervals. Results showed that a random forest model associated with 25 covariates was successful in capturing 83% of spatial and vertical SOC variation over the country. Soil depth was the most important covariate to predict SOC concentration, followed by surface temperature and elevation. The SOC concentration in forested areas was highest in the top layers (0–5 cm), but soils located in peatlands showed higher C concentration in all soil depths. Among the terrestrial ecozones of Canada, Pacific Maritime and the Hudson Plain mostly covered by forest trees and peatlands, respectively, show highest SOC concentration, while the lowest concentration are observed in the Prairies and Mixed Wood Plain ecosystems that are the agricultural areas of the country. This study provides a deeper understanding of the major factors controlling SOC concentration in Canada and shows potential areas with high carbon reserves, which are crucial in view of the ongoing climate change. In addition, the presented methodological framework has great potential to be used in future soil carbon storage inventories using satellite observations. Mapping SOC concentration and associated uncertainties in Canada are fundamental to detect trends of gains or losses in SOC linked to recent and future national or global policy decisions related to soil quality and carbon sequestration.

加拿大拥有广阔的森林和泥炭地，在全球碳循环中发挥着关键作用。假定加拿大土壤和泥炭地储存了世界大约 20% 的土壤碳储量。然而，加拿大土壤有机碳 (SOC) 浓度的空间和垂直分布尚未得到系统研究。SOC 浓度，以 g C kg ^{-1 表示}土壤，影响土壤的化学和物理特性，如水分渗透能力、持水能力、养分有效性和微生物的生物活性。在这项研究中，我们测试了一种三维 (3D) 机器学习方法和 40 个来自 20 年光学和微波卫星观测的空间预测因子，以估计加拿大 6 个深度间隔内 SOC 浓度的空间和垂直分布，最远可达 1 m。分位数回归森林方法用于构建显示 80% 预测区间的不确定性图。结果表明，与 25 个协变量相关的随机森林模型成功捕获了全国 83% 的空间和垂直 SOC 变化。土壤深度是预测 SOC 浓度的最重要协变量，其次是地表温度和海拔。森林地区的 SOC 浓度在顶层（0-5 厘米）最高，但位于泥炭地的土壤在所有土壤深度都表现出较高的 C 浓度。在加拿大的陆地生态区中，主要被林木和泥炭地覆盖的太平洋海洋和哈德逊平原分别显示出最高的 SOC 浓度，而在该国农业区的草原和混合木材平原生态系统中观察到的浓度最低. 这项研究提供了对控制加拿大 SOC 浓度的主要因素的更深入的了解，并显示了具有高碳储量的潜在区域，鉴于持续的气候变化，这些区域至关重要。此外，提出的方法框架具有巨大的潜力，可用于使用卫星观测的未来土壤碳储存清单。绘制加拿大的 SOC 浓度和相关不确定性对于检测与近期和未来与土壤质量和碳固存相关的国家或全球政策决定相关的 SOC 收益或损失趋势至关重要。