Abstract. Spatially explicit prediction of soil organic carbon (SOC) serves as a crucial foundation for effective land management strategies aimed at mitigating soil degradation and assessing carbon sequestration potential. Here, using more than 1000 in situ observations, we trained two machine learning models (a random forest model and a k-means coupled with multiple linear regression model) and one process-based model (the vertically resolved MIcrobial-MIneral Carbon Stabilization, MIMICS, model) to predict the SOC stocks of the top 30 cm of soil in Australia. Parameters of MIMICS were optimised for different site groupings using two distinct approaches: plant functional types (MIMICS-PFT) and the most influential environmental factors (MIMICS-ENV). All models showed good performance with respect to SOC predictions, with an R2 value greater than 0.8 during out-of-sample validation, with random forest being the most accurate; moreover, it was found that SOC in forests is more predictable than that in non-forest soils excluding croplands. The performance of continental-scale SOC predictions by MIMICS-ENV is better than that by MIMICS-PFT especially in non-forest soils. Digital maps of terrestrial SOC stocks generated using all of the models showed a similar spatial distribution, with higher values in south-eastern and south-western Australia, but the magnitude of the estimated SOC stocks varied. The mean ensemble estimate of SOC stocks was 30.3 t ha−1, with k-means coupled with multiple linear regression generating the highest estimate (mean SOC stocks of 38.15 t ha−1) and MIMICS-PFT generating the lowest estimate (mean SOC stocks of 24.29 t ha−1). We suggest that enhancing process-based models to incorporate newly identified drivers that significantly influence SOC variation in different environments could be the key to reducing the discrepancies in these estimates. Our findings underscore the considerable uncertainty in SOC estimates derived from different modelling approaches and emphasise the importance of rigorous out-of-sample validation before applying any one approach in Australia.

中文翻译：

---

使用机器学习和基于过程的建模对澳大利亚土壤有机碳进行整体估计


摘要。土壤有机碳（SOC）的空间明确预测是旨在减轻土壤退化和评估碳封存潜力的有效土地管理策略的重要基础。在这里，我们使用 1000 多个现场观测数据训练了两种机器学习模型（一种随机森林模型和一种与多元线性回归模型相结合的 k 均值模型）和一种基于过程的模型（垂直解析的 MIcrobial-MIneral Carbon Stabilization，MIMICS ，模型）来预测澳大利亚表层 30 厘米土壤的 SOC 储量。使用两种不同的方法针对不同位点分组优化 MIMICS 参数：植物功能类型 (MIMICS-PFT) 和最有影响力的环境因素 (MIMICS-ENV)。所有模型在 SOC 预测方面均表现出良好的性能，在样本外验证期间 R2 值大于 0.8，其中随机森林最为准确；此外，研究发现森林中的有机碳比农田以外的非森林土壤中的有机碳更容易预测。 MIMICS-ENV 对大陆尺度 SOC 的预测性能优于 MIMICS-PFT，特别是在非森林土壤中。使用所有模型生成的陆地 SOC 储量数字地图显示出相似的空间分布，澳大利亚东南部和西南部的值较高，但估计的 SOC 储量的大小有所不同。 SOC 库的平均集合估计为 30.3 t ha−1，k 均值与多元线性回归相结合生成最高估计（平均 SOC 库为 38.15 t ha−1），MIMICS-PFT 生成最低估计（平均 SOC 库） 24.29 t ha−1)。 我们建议，增强基于过程的模型以纳入新发现的显着影响不同环境中 SOC 变化的驱动因素可能是减少这些估计差异的关键。我们的研究结果强调了不同建模方法得出的 SOC 估计值存在相当大的不确定性，并强调了在澳大利亚应用任何一种方法之前严格的样本外验证的重要性。