Machine learning for fertilizer recommendation in Ghana

Maize is a major staple in Ghana and Sub-Saharan Africa (SSA), yet yields remain low due to soil fertility constraints, inefficient fertilizer use, and blanket recommendations that ignore local variability. While mechanistic models such as QUEFTS improve understanding of crop–nutrient relations, their application is limited by high data demands and weak integration of farmer risk and socioeconomic realities. This thesis evaluates machine learning (ML) as a complementary approach for site-specific maize fertilizer recommendations in Ghana. Using multi-season, multi-location field trial data, Random Forest (RF) was trained to predict yield, agronomic efficiency, and uncertainty. RF explained up to 81% of yield variation, with nitrogen rate, rainfall, temperature, soil organic carbon, and bulk density as key drivers. Field validation showed ML-based recommendations were more site-specific and cost-effective than conventional practices, particularly in the Guinea Savanna and Forest–Savanna Transition zones. Overall, ML offers flexible, data-driven, and risk-aware decision support that can enhance productivity, profitability, and resilience in smallholder maize systems.