Application of machine learning approaches to predict ammonium nitrogen transport in different soil types and evaluate the contribution of control factors.

The loss of nitrogen in soil damages the environment. Clarifying the mechanism of ammonium nitrogen (NH-N) transport in soil and increasing the fixation of NH-N after N application are effective methods for improving N use efficiency. However, the main factors are not easily identified because of the complicated transport and retardation factors in different soils. This study employed machine learning (ML) to identify the main influencing factors that contribute to the retardation factor (Rf) of NH-N in soil. First, NH-N transport in the soil was investigated using column experiments and a transport model. The Rf (1.29 - 17.42) was calculated and used as a proxy for the efficacy of NH-N transport. Second, the physicochemical parameters of the soil were determined and screened using lasso and ridge regressions as inputs for the ML model. Third, six machine learning models were evaluated: Adaptive Boosting, Extreme Gradient Boosting (XGB), Random Forest, Gradient Boosting Regression, Multilayer Perceptron, and Support Vector Regression. The optimal ML model of the XGB model with a low mean absolute error (0.81), mean squared error (0.50), and high test r (0.97) was obtained by random sampling and five-fold cross-validation. Finally, SHapely Additive exPlanations, entropy-based feature importance, and permutation characteristic importance were used for global interpretation. The cation exchange capacity (CEC), total organic carbon (TOC), and Kaolin had the greatest effects on NH-N transport in the soil. The accumulated local effect offered a fundamental insight: When CEC > 6 cmol kg, and TOC > 40 g kg, the maximum resistance to NH-N transport within the soil was observed. This study provides a novel approach for predicting the impact of the soil environment on NH-N transport and guiding the establishment of an early-warning system of nutrient loss.