Purpose

Riverine floodplain soils are important reservoirs of organic carbon (OC) in terrestrial ecosystems because of their high biomass productivity and OC input via flood events. Substantial knowledge of the riverine floodplain soil OC distribution and its impacting variables are significant for predicting carbon sequestration and emission. This study aimed at (1) predicting SOC in riverine floodplain soils using the multiple linear regression (MLR), M5P, and random forest (RF) models, (2) comparing the model performances, and (3) revealing the significant variables controlling the spatial dynamic of riverine floodplain soils OC.

Materials and methods

In this study, 4227 topsoil samples (0–20 cm) from the Yangtze riverine floodplain were collected and analyzed for soil organic carbon (SOC) and other properties. We identified the key variables impacting SOC and predicted the SOC spatial distribution based on three predictive models (i.e., MLR, M5P, and RF), measured soil attributes, and land-use data.

Results and discussion

The study results indicated that total soil sulfur and nitrogen concentrations were the most important variables affecting the SOC distribution, followed by soil geochemical variables (e.g., Al₂O₃, Na₂O, CaO, and Fe₂O₃), and alkalinity conditions. The M5P and RF models showed higher accuracy in predicting SOC compared with the MLR model. Although RF outperformed M5P in SOC prediction, RF was limited in revealing the relationships between SOC and environmental variables, restricting its interpretability. The land-use analysis highlighted that paddy soils were more conducive to maintaining high SOC concentration than upland soils.

Conclusions

We recommend using the RF and M5P models to efficiently predict the SOC distribution in riverine floodplain soils as RF could produce higher prediction accuracy and M5P can detect the splitting process and identify relevant thresholds for the regression. Paddy management is crucial for SOC sequestration and grain production. Therefore, it is recommended as an efficient approach to enhance soil fertility, mitigate climate change, and ensure food security.

中文翻译：

---

基于机器学习的集约化农业实践对河流漫滩土壤有机碳的预测和关键因素识别

目的

河流泛滥平原土壤是陆地生态系统中重要的有机碳 (OC) 库，因为它们具有高生物量生产力和通过洪水事件输入的有机碳。河流泛滥平原土壤有机碳分布及其影响变量的大量知识对于预测碳固存和排放具有重要意义。本研究旨在 (1) 使用多元线性回归 (MLR)、M5P 和随机森林 (RF) 模型预测河流漫滩土壤中的 SOC，(2) 比较模型性能，以及 (3) 揭示控制河流泛滥平原土壤有机碳的空间动态。

材料和方法

在这项研究中，收集了来自长江泛滥平原的 4227 个表土样品（0-20 厘米），并分析了土壤有机碳 (SOC) 和其他特性。我们确定了影响 SOC 的关键变量，并基于三个预测模型（即 MLR、M5P 和 RF）、测量的土壤属性和土地利用数据预测了 SOC 空间分布。

结果和讨论

研究结果表明，土壤总硫和氮浓度是影响 SOC 分布的最重要变量，其次是土壤地球化学变量（如 Al ₂ O ₃、Na ₂ O、CaO 和 Fe ₂ O ₃）和碱度条件. 与 MLR 模型相比，M5P 和 RF 模型在预测 SOC 方面表现出更高的准确性。尽管 RF 在 SOC 预测方面优于 M5P，但 RF 在揭示 SOC 与环境变量之间的关系方面受到限制，限制了其可解释性。土地利用分析强调，稻田土壤比旱地土壤更有利于保持高 SOC 浓度。

结论

我们建议使用 RF 和 M5P 模型来有效预测河流泛滥平原土壤中的 SOC 分布，因为 RF 可以产生更高的预测精度，而 M5P 可以检测分裂过程并确定回归的相关阈值。稻谷管理对于 SOC 封存和粮食生产至关重要。因此，建议将其作为提高土壤肥力、缓解气候变化和确保粮食安全的有效方法。