Soil organic carbon stock and uncertainties, 1m depth, at 250m spatial resolution in Canada

This project aimed to produce the first wall-to-wall estimate of C stocks in plants and soils of Canada at 250 m spatial resolution. This dataset contains the map with the soil organic carbon (SOC) in kg/m² for entire Canada in 1m depth, and the uncertainty in SOC predictions. The SOC stock map was produced using 6,533 ground soil samples, long-term climate data, remote sensing observations and a machine learning model. The soil samples containing the x and y coordinates, depth and SOC (in g/kg) information were overlaid with the stacked covariates (soil forming factors) to compose the regression matrix. Random forest models were trained using a recursive feature elimination scheme and a cross-validation assessment. The best model was used for spatial prediction of SOC over Canada in intermediate depths between 0 and 1 m. Afterwards, the SOC content maps were corrected with bulk density and coarse fragment information to compute the total carbon stock for each horizon. The horizons have been added to compose the 0-1m depth interval multiplied by root depths fraction to discount shallow soils. Water and ice/snow areas were removed using a mask based on the Land Cover of Canada map. The SOC stock uncertainty map was built using the first and third quantiles of RF quantile regression approach.<br>

本项目旨在以250米空间分辨率，生成加拿大境内植物与土壤碳储量的首份全覆盖估算成果。本数据集包含加拿大全境1米深度土层的土壤有机碳（Soil Organic Carbon, SOC）储量（单位：kg/m²）分布图，以及SOC预测结果的不确定性信息。该SOC储量分布图依托6533份野外土壤样本、长期气候数据、遥感观测数据与机器学习模型生成。携带有X/Y坐标、土层深度及SOC含量（单位：g/kg）信息的土壤样本，与堆叠后的协变量（即成土因子）叠加后构建回归矩阵。研究采用递归特征消除策略与交叉验证评估方法训练随机森林模型，选取最优模型对加拿大境内0至1米区间内的各中间土层开展SOC空间预测。随后，通过土壤容重与粗碎屑组分信息对SOC含量分布图进行校正，以计算各土层的总碳储量；将各土层数据整合得到0-1米深度段的总储量，并结合根系深度占比对浅土层进行折算修正。基于加拿大土地覆盖图生成掩膜，去除水体与冰雪覆盖区域。本研究基于随机森林分位数回归方法的第一与第三分位数，构建SOC储量不确定性分布图。