面向亚热带丘陵区小流域土壤有机碳空间预测的四种模型构建及性能比较

土壤是陆地生态系统最大的碳库,在提升生态系统服务功能和调节气候变化等方面发挥关键作用。对复杂多变环境下土壤有机碳(SOC)含量的精确预测将有助于正确评估区域土壤质量和碳汇功能。本研究以亚热带丘陵区一个典型小流域为研究对象,以地形、气候和植被三类环境变量为驱动因子,分析支持向量机回归(SVR)、随机森林(RF)、极端梯度提升算法(XGBoost)和轻量级梯度提升机(LightGBM)四种不同的机器学习算法在土壤(0~20 cm)SOC含量预测中的精度差异,并筛选影响SOC分布的主要环境影响因素。结果表明,RF模型、XGBoost模型和LightGBM模型均能较好预测SOC含量,以RF模型的表现相对最佳(R^(2)=0.540),其预测精度优于XGBoost(R^(2)=0.528)和LightGBM模型(R^(2)=0.504)。而SVR模型的预测精度(R^(2)=0.427)低于模型预测精度的最低可接受值0.50,并不适用于亚热带丘陵地貌SOC含量的预测。相关分析表明,在亚热带丘陵地貌区,地形(主要为海拔)对几种模型预测的贡献最大,是预测SOC的重要环境变量。基于四种模型预测的SOC数字制图显示,SOC空间分布趋势总体相似,均表现为北部区域、西南和东南边缘区域SOC含量较高,而中部区域SOC含量普遍偏低。

Comparison of four machine learning in predicting soil organic carbon content in a small watershed in the subtropical hilly area

As the largest C pool in terrestrial ecosystems, soil plays a major role in the enhancement of ecosystem services and the regulation of climate change. An accurate prediction of soil organic carbon (SOC) content in areas with complex and variable environments will assist in assessing soil quality and carbon sink functions at regional scales. In this study, a typical small watershed in a subtropical hilly region was selected as the research object, and four machine learning algorithms namely, support vector machine regression (SVR), random forest (RF), extreme gradient boosting algorithm (XGBoost) and light gradient boosting machine (LightGBM), were used to predict the SOC content in the soil surface layer (0~20 cm). Three types of environmental variables, including topography, climate and vegetation, were utilized as environmental factors. The purpose was to determine the effectiveness of different algorithms in predicting SOC content, and to screen the primary environmental influences affecting SOC distribution. Among the four models, the RF model performed the best in predicting SOC with RF (R2=0.540), and its prediction accuracy was superior to that of XGBoost (R2=0.528) and LightGBM (R2=0.504). Contrary to this, the SVR model had relatively low prediction accuracy (R2=0.427), therefore it was not suitable for predicting SOC content in subtropical hilly landscapes. As a result of the correlation analysis, it was found that topography (primarily elevation) played the most significant role in the model prediction in the subtropical hilly landscape area. In the digital mapping made by four model predictions, it was generally found that the trends of SOC spatial distribution were similar. Each showed a higher SOC content in the northern, southwestern and southeastern marginal regions, while the central region exhibited a low SOC content.