The Driving Forces and Spatial Predictions of Soil Total Nitrogen and Soil Total Phosphorus Using Machine Learning and Explainable AI: A Case Study of Grasslands in Qinghai Province, China

Xinze Guo, Yiming Xu, Zhenqiang Liu, Youquan Tan, Tengfei Fan

Soil total nitrogen (TN) and soil total phosphorus (TP) are key soil quality indicators and provide critical ecological functions in the grasslands. This study analyzed the driving factors of TN/TP in the grasslands of Qinghai Province based on Shapley additive interpretation (SHAP) analysis. Four machine learning methods, namely random forest (RF), XGBoost 3.2.0, support vector machine, and Cubist, were used to establish spatial prediction models for TN/TP. Vegetation factors (Net Primary Production and Normalized Difference Vegetation Index) and precipitation-related factors (Aridity Index and Mean Annual Precipitation) were the most important variables for TN, indicating plant productivity and precipitation are strongly associated with TN accumulation. Elevation and temperature-related factors (Mean Annual Temperature and evapotranspiration) were the most important variables for TP, demonstrating that elevation-mediated temperature was the major factor affecting the TP accumulation. XGBoost and RF were the optimal models for TN and TP, respectively. TN exhibited a decreasing spatial trend from east to west, while the northwestern and southwestern areas showed relatively higher and lower TP, respectively. Total TN and TP stocks were estimated to be 3.57 × 108 t and 0.88 × 108 t, respectively. This study provides data support and suggestions for sustainable soil nutrient management in the grasslands on the Qinghai-Tibet Plateau.