Assessing the impact of multi-source environmental variables on soil organic carbon in different land use types of China using an interpretable high-precision machine learning method

作者：Feng Wang, Ruilin Liang, Shuyue Li, Meiyan Xiang, Weihao Yang, Miao Lu, Yingqiang Song · 发表于：Ecological Indicators · 年份：2024 · DOI：10.1016/j.ecolind.2024.112865 · 被引用次数：16 · 研究领域：Soil Carbon and Nitrogen Dynamics、Soil Geostatistics and Mapping、Plant Ecology and Soil Science

• TPE-XGBoost is the state-of-the-art technology to fit environmental factors and SOC. • The SHAP clarified positive and negative driving effects of environmental factors. • Temperature and pH have the high response for SOC in different land use types. • Temperature affects the activity of microorganisms and change the content of SOC. To explore the impact of environmental factors on soil organic carbon (SOC) with machine learning (ML) model is of great significance for mitigating climate change and soil carbon sequestration and emission reduction. However, the traditional ML model is limited by the hyperparameter adjustment of artificially trial-and-error experimentation and the inexplicability of fitting process, and the precision and performance of ML model cannot be fully utilized. For the end, this study developed a tree-structured Parzen estimator-extreme gradient boosting (TPE-XGBoost) method based on SHapley additive explanations (SHAP) analysis to analyze the response of climate, human activities, soil properties and terrain for SOC (0-200cm) in different land use types of China. The results of descriptive statistics described the order of SOC content: forest land > grassland > cultivated land > unused land. With the increase of soil depth, the SOC content of all land types decreased continuously, and the values indicate a left-skewed non-normal distribution. The fitting accuracy (R 2 ) of TPE-XGBoost model for SOC content was greater than 0.8. At the depth of 0-5cm,...