Towards site and context-specific fertilizer recommendations for maize production in Northern Ghana using AgWise framework (Machine learning component)

Agroadvisory services play vital roles in promoting sustainable and productive agricultural practices, particularly in areas that experience fluctuations in climate and limited resources. In Ghana, the Alliance of Bioversity International & CIAT in collaboration with IITA and CSIR aims integrate input supply, insurance, and advisory services as a bundle that provides holistic support to enhance productivity and mitigate the risks associated with maize cultivation. This study aims to improve advisory services concerning fertilizer use by maize farmers from the Guinea Savannah zone of Ghana. The objective is to generate recommendations that are tailored to specific areas as current practice or recommended rates a based on current and legacy soil tests and experiments datasets. Site and context-specific fertilizer recommendations were generated using the AgWise framework. More than 1,740 data points were gathered from scientific publications, field trials and crop cut surveys and used in the model development. To ensure precision, only studies that reported site-specific trial locations were included, with data gathered on nitrogen, phosphorus, and potassium application and their effects on maize yield. To build robust models, a thorough exploration of the extracted covariate data was conducted, encompassing soil characteristics, climate patterns, and topography. After careful pre-processing, a dataset ready for model training was obtained, containing 1532 observations and 47 data columns. The modeling process utilized the H2O package within RStudio. H2O's AutoML functionality was employed to streamline the workflow. A total of ten models were trained using AutoML, and the most promising ones from the output were selected for additional refinement. These included Random Forest (RF), Gradient Boosting Machine (GBM), Deep Learning Artificial Neural Network (ANN), and Generalized Linear Models (GLM). The random forest model emerged as the top performer through rigorous evaluation, based on MAE, RMSE, and R², which was utilized in the prediction. Analysis revealed that nitrogen application rate, soil characteristics, and climatic conditions are the primary determinants of maize yield. Preliminary results indicate significant variability in fertilizer requirements across different districts, reinforcing the need for data-driven, context-specific recommendations. The model successfully generated preliminary fertilizer recommendations for nitrogen (30–140 kg/ha), phosphorus (10–35 kg/ha), and potassium (0–40 kg/ha) across various districts. Expanding the dataset and refining the model will further enhance these recommendations, supporting maize growth, improving yields, boosting farmer productivity, and reducing production risks in Ghana.