MACHINE LEARNING-BASED APPROACHES TO IMPROVE CROP YIELD PREDICTION IN AGRICULTURE IN ADAMAWA STATE, NIGERIA

Abstract

Accurate yield forecasting is critical for sustainable agriculture in developing countries like Nigeria. This study evaluates three machine learning models: Artificial Neural Network (ANN), Decision Tree (DT), and Support Vector Machine (SVM) for predicting crop yields across five agriculturally diverse Local Government Areas in Adamawa State, Nigeria. Using environmental data from 2019 to 2023, including rainfall, temperature, and soil moisture, we trained models to forecast yields (kg/ha) for staple crops like maize, sorghum, and groundnuts. Results showed that ANN and DT outperformed SVM, achieving high accuracy (R²: 0.93–0.98) and low errors (RMSE: 5.5–10.9). While the ANN excelled in Madagali (R²=0.97, RMSE=6.7), the DT performed best in Mubi North (R²=0.98, RMSE=5.5), whereas the SVM consistently underperformed (R²: 0.42–0.46; RMSE: 29.8–36.6) in all study areas. The result also showed rainfall and soil moisture were stronger yield predictors than temperature. These results demonstrate ANN and DT’s suitability for yield prediction in diverse environments. We recommend that models which prioritize rainfall and soil moisture over temperature be employed in crop yield prediction in the study areas and ANN or DT tools are to be adopted to aid smallholder farmers in farm resource planning. These will reduce prediction errors by 76 to 82% compared to traditional methods, enabling proactive resource allocation and enhancing food security in Northern Nigeria.