| Summary: | Addressing the persistent challenge of low maize productivity in Malawi requires spatially explicit and nutrient-specific fertilizer recommendations that align with soil heterogeneity and economic constraints. Existing blanket recommendations often overlook localized variability in soil properties, climate, and nutrient responses, limiting their effectiveness. This study aimed to (1) evaluate and compare the predictive performance of multiple machine learning algorithms for maize yield estimation in Malawi; (2) identify the most important yield-determining features through recursive feature elimination (RFECV) and SHAP-based interpretation; (3) examine interaction effects between key nutrient inputs and environmental variables using two-dimensional partial dependence plots; and (4) translate model outputs into site-specific nutrient management insights for precision agronomy in smallholder maize systems. A Light Gradient Boosting Machine (LightGBM) model was selected based on performance and interpretability and trained on 20 selected features including nutrients management, soil, climate, topography, and hydrological covariates. Hyperparameter optimization and SHAP analysis enhanced both accuracy and model transparency. We generated fine-scale (250 m resolution) maize fertilizer recommendations for N, P, K, S, and Zn using the optimized LightGBM model. Results revealed substantial spatial variation in nutrient requirements, with nitrogen and phosphorus showing the widest ranges in both agronomic and economic rates. The economic optimum rates were slightly lower than agronomic ones, highlighting the importance of cost-benefit considerations in fertilizer planning. Spatial patterns also indicated that highly weathered soils in southern and central Malawi often required higher inputs, while nutrient-rich areas in the north exhibited lower optimal rates. These findings support a shift from national blanket recommendations to precision-guided input planning using high-resolution geospatial data and data-driven models.
Keywords: Artificial intelligence; machine learning; nutrient; maize; LightGBM; SHAP
|
|---|