Machine learning for fertilizer recommendation in Ghana

Maize is a major staple crop in Ghana and across Sub-Saharan Africa (SSA), yet yields remain far below their potential due to persistent soil fertility constraints, inefficient fertilizer use, and reliance on blanket recommendations that fail to capture site-specific variability. Mechanistic models such as QUEFTS have advanced understanding of crop–nutrient relationships, but their application in SSA is limited by high data requirements and weak consideration of farmer risk and socioeconomic realities. Recent advances in machine learning (ML) provide new opportunities to generate data-driven, site-specific, and adaptive fertilizer recommendations, but questions remain regarding model reliability under data-scarce conditions, uncertainty quantification, and alignment with smallholder decision-making. This thesis evaluates the potential of ML to support maize fertilizer recommendations in Ghana by addressing yield prediction, agronomic efficiency, uncertainty, and farmer risk preferences. Using large, multi-season, multi-location datasets of maize field trials, Random Forest (RF) and other ML models were trained and rigorously validated. Results showed that RF explained up to 81% of yield variation and moderately predicted agronomic efficiency, with nitrogen fertilizer, rainfall, temperature, soil organic carbon, and bulk density emerging as key drivers. Comparative analysis of ML algorithms demonstrated that tree-based models (RF and XGBoost) outperformed other approaches under heterogeneous and data-scarce conditions. Field validation experiments across major agroecological zones revealed that ML-derived fertilizer recommendations were generally more site-specific and cost-effective than conventional approaches, particularly in the Guinea Savanna and Forest–Savanna Transition zones. While mechanistic models performed well in some environments, ML approaches often achieved comparable or higher yields with improved net profit margins. Importantly, integrating predictive uncertainty and farmer risk preferences through a utility-based framework showed that risk-averse farmers prefer lower fertilizer rates than technically optimal recommendations, especially in marginal environments. Overall, the thesis demonstrates that ML can complement mechanistic models by providing flexible, data-driven, and risk-aware decision support. By explicitly accounting for uncertainty, data limitations, and farmer behaviour, this research lays the foundation for farmer-centered agronomic decision-support systems that enhance productivity, profitability, and resilience in smallholder maize systems in SSA.