Deep-Learning-Driven Insights into Nitrogen Leaching for Sustainable Land Use and Agricultural Practices

Nitrate leaching from soil presents a significant threat to soil health, as it can result in nutrient loss, soil acidification, and structural damage. It is crucial to quantify the spatial heterogeneity of nitrate leaching and its drivers. A total of 509 observational data points regarding nitrate leaching in northern China were collected, capturing the spatial and temporal variations across crops such as winter wheat, maize, and greenhouse vegetables. A machine learning (ML) model for predicting nitrate leaching was then developed, with the random forest (RF) model outperforming the support vector machine (SVM), extreme gradient boosting (XGBoost), and convolutional neural network (CNN) models, achieving an R2 of 0.75. However, the performance improved significantly after integrating the four models with Bayesian optimization (all models had R2 > 0.56), which realized quantitative prediction capabilities for nitrate leaching loss concentrations. Moreover, the XGBoost model exhibited the highest fitting accuracy and the smallest error in estimating nitrate leaching losses, with an R2 value of 0.79 and an average absolute error (MAE) of 3.87 kg/ha. Analyses of the feature importance and SHAP values in the optimal XGBoost model identified soil organic matter, chemical nitrogen fertilizer input, and water input (including rainfall and irrigation) as the main indicators of nitrate leaching loss. The ML-based modeling method developed overcomes the difficulty of the determination of the functional relationship between nitrate loss intensity and its influencing factors, providing a data-driven solution for estimating nitrate–nitrogen loss in farmlands in North China and strengthening sustainable agricultural practices.