Machine Learning for Fertilizer Recommendation in Ghana

Information on site-specific fertilizer recommendations for efficient nutrient use and yield prediction prior to harvesting is critical for adequate planning along the agricultural value chain in Ghana. Conventional methods to generate fertilizer recommendations are time-consuming, expensive and practically not possible to cover large areas, hence unable to arrive at accurate site- and crop-specificity. In this study, a novel approach that addresses prevailing challenges to existing methods is proposed. A machine learning model will be developed and tested for maize to derive fertilizer recommendations to improve the yields and ensure efficient nutrient use in Ghana. We develop and apply a Random Forest model that predicts nutrient use efficiencies and its associated yields from maize yield datasets. We will derive fertiliser recommendations from the calibrated Random Forest model and compare them with recommendations derived from existing empirical method such as QUEFTS. Furthermore, we will develop other machine learning models such as the support vector machine, extreme gradient boosting and the K nearest neighbors algorithms and evaluate their predicted nutrient use efficiencies and yields. We will then develop a fertilizer recommendation strategy that accounts for uncertainties in the model predictions that will be relevant for actors in the fertiliser value chain. The results from this research will be highly relevant to actors in the fertilizer value chain for decision making to boost maize production in Ghana.