Predicting nitrogen surplus in agricultural lands of China using a hybrid machine learning approach with smaller datasets and fewer features.

Global nitrogen pollution in agricultural lands poses a major environmental challenge, complicating both assessment and mitigation of nitrogen surplus. Nitrogen surplus (NS) is a critical indicator for evaluating nitrogen use efficiency. However, traditional NS estimation methods often require extensive data, which are difficult to obtain in data-scarce regions. In this study, a BN-ML NS prediction model was developed by coupling Bayesian Networks (BN) and Machine Learning (ML) using long-term monitoring data (including climate, soil, and crops) from 2000 to 2019 in China. The key findings are as follows: Nitrogen fertilizer application rate (NR) is the dominant factor influencing NS across the seven regions; however, due to differences in climate, cropping patterns, and soil types, the impact of NR exhibits spatial heterogeneity; The BN-ML model demonstrates strong predictive performance, with R² values exceeding 0.9; compared to traditional NS prediction models, the BN-ML model maintains high accuracy using only half the number of features (e.g., NR, Yield) and fewer than 120 data points; when validated at the small watershed scale, the model achieved over 80 % prediction accuracy. By enabling reliable NS prediction with limited data, the proposed model supports more targeted and sustainable nitrogen management. It can assist national and regional authorities in identifying high-risk areas, optimizing fertilizer use, and formulating region-specific agricultural strategies.