Machine Learning Approaches for Predicting Maize Yield from Remote Sensing Indices in the Nigerian Savanna

This study develops and validates machine learning models for predicting maize yield from satellite-derived spectral indices in the Nigerian savanna agro-ecological zone, contributing to improved regional crop monitoring capabilities. Timely and accurate crop yield forecasting is essential for food security monitoring and agricultural planning, but ground-based yield surveys are expensive, slow, and spatially limited. Remote sensing-based yield prediction models offer a more scalable approach, but model performance is highly sensitive to the choice of spectral indices, prediction algorithm, and model calibration data quality. This study combines Sentinel-2 and Landsat 8 spectral data including NDVI, EVI, LAI, and LSWI indices with 280 ground-truth yield observations from maize fields in Kaduna, Niger, and Kogi States collected over three growing seasons. Four machine learning algorithms are compared: support vector regression, random forest, gradient boosting, and artificial neural network. Model training uses 80 percent of observations with 20 percent holdout validation. Cross-validation across seasons tests model temporal transferability. Findings reveal that the gradient boosting model achieves the highest prediction accuracy with R-squared of 0.81 and RMSE of 0.42 tonnes per hectare on the validation set, outperforming multiple linear regression at R-squared of 0.64. NDVI at tasselling stage and EVI at grain fill are the most important predictor variables. Model performance transfers adequately across years with minor recalibration. The study recommends the gradient boosting model for integration into Nigeria's agricultural early warning system.

Keywords: machine learning, crop yield prediction, remote sensing, maize, Nigerian savanna.