Machine learning algorithms realized soil stoichiometry prediction and its driver identification in intensive agroecosystems across a north-south transect of eastern China

Soil carbon (C), nitrogen (N), and phosphorus (P) are required components to maintain ecosystem structure, function, and services. Accurate soil nutrient stoichiometry assessments are crucial for precisely managing agricultural and natural ecosystems. However, direct measurement and evaluation of soil characteristics can be costly and time-consuming. The development of statistical and machine learning-based methods for predicting soil C:N:P stoichiometry and microbial dynamics is of great significance. The objective of this study is to compare the performance of four machine learning models, i.e., support vector machine, random forest, extreme gradient boosting, and gradient boosting decision tree, in predicting soil C:N:P stoichiometry and net N mineralization rate and to evaluate their applicability to different agricultural land use types and climate zones. Our results showed that extreme gradient boosting (average R-2 > 0.81, RMSE <16.39, RPD > 2.67) and gradient boosting decision tree (average R-2 > 0.77, RMSE <16.40, RPD > 2.32) models performed the best in predicting C:N:P stoichiometry, demonstrating high accuracy and stability. Machine learning models produced higher accuracy in the vegetable field (except for C:N) than in the rice paddy field with average accuracy improvement of 42.9 %. The prediction performance in warm temperate and subtropical regions was inferior to cold regions. Feature importance assessment suggests that electrical conductivity, total N, and water-filled pore space may have significant predictive roles in the rice paddy field, while mean annual precipitation, total P, and silt content could be important factors in the vegetable field. When predicting the net N mineralization rate, soil texture may emerge as a crucial factor in the rice paddy field, whereas moisture content may play a key role in the vegetable field. Thus, machine learning models can be recommended to predict soil C:N:P stoichiometry and net N mineralization rate for precise agricultural practices.