Exploring and predicting nitrogen fertilizer use efficiency of maize (Zea mays L.)-soybean (Glycine max (L.) Merr.) intercropping systems in China: A combined Meta-analysis and machine learning approach

Abstract

Maize (Zea mays L.)-soybean (Glycine max (L.) Merr.) intercropping systems have been widely promoted in China due to their potential to enhance agricultural land use efficiency through crop complementarity. However, the complex interactions between environmental conditions and management practices create significant challenges for understanding and accurately predicting nitrogen (N) use efficiency under different production conditions. Through a meta-analysis of 330 datasets from 45 experimental sites across China and machine learning approaches, we evaluated the fertilizer N equivalent ratio (FNER) and developed a prediction model for it in maize-soybean intercropping systems. The national average FNER value of maize-soybean intercropping systems in China was 1.41 ± 0.02. The FNER of maize-soybean intercropping systems was significantly correlated with climate conditions, the proportion of maize, N application rate, and temporal niche differentiation (TND). Regions with higher annual precipitation and temperature showed a greater N fertilizer utilization advantage in maize-soybean intercropping. Furthermore, reducing N application rate and the proportion of maize while extending TND can enhance the FNER of maize-soybean intercropping systems. N application rate and TND were identified as the most important input parameters for machine learning-based FNER prediction in maize-soybean intercropping systems. Among the machine learning models, Random Forest and Gradient Boosting models demonstrated superior effectiveness in predicting FNER values using five input variables, including N application rate, TND, soil organic matter content, average annual temperature, and soybean planting density. The current study can provide practical guidance for improving the N use efficiencies of maize-soybean intercropping systems and offer a robust tool for predicting FNER under various environmental and management conditions.