Machine learning insights into long-term soil fertility legacy and seasonal weather effects in corn–soybean rotation

Abstract

Multiple factors, such as accumulated fertility legacies, crop rotation, soil properties, and seasonal weather variables, complicate long-term studies due to their nonlinearity and interdependencies. Therefore, we hypothesize that a machine learning (ML)-based data analytics approach can help address this gap and identify the most influential variables associated with the long-term yields of corn (Zea mays) and soybeans (Glycine max). The objectives of the study were to identify the key soil and weather variables associated with corn and soybean yield variability, to assess crop-specific variables, and to evaluate the robustness of a single nonlinear ML approach for long-term yield prediction. For that purpose, a 43-year soil fertility experiment was conducted in Topeka, KS, with 20 combinations of nitrogen (N), phosphorus (P), and potassium (K) fertilizer rates. Variables such as temperature, precipitation, irrigation, soil pH, soil P and K levels, soil texture, and N–P–K fertilization rates (both annual and cumulative) were combined into a Random Forest (RF) ML model. For corn, N fertilizer, soil P, and soil pH were important variables associated with yield variability, whereas soybean yields were more associated with weather variables, particularly rainfall, evapotranspiration, and soil P. The RF model performed well for corn (R² = 0.81, RMSE = 1449 kg ha⁻¹) and soybean (R² = 0.74, RMSE = 527 kg ha⁻¹) yield estimation. These results demonstrate that ML models can help identify key predictors of yield variability and provide robust yield estimates, thereby supporting improved long-term crop management decisions.