Stacked Ensemble Machine Learning for Crop Yield Prediction Using Agro-Meteorological and Remote Sensing Features

Accurate pre-harvest crop yield prediction is a critical enabler of food security planning, agricultural insurance underwriting, market price stabilisation, and precision farming resource allocation. Traditional agronomic models rely on process-based crop simulation frameworks that require extensive parameterisation and exhibit limited transferability across geographies and cultivar types. This study proposes a two-level stacked ensemble machine learning model that integrates agro-meteorological variables, satellite-derived vegetation indices, and soil physicochemical properties to predict rice (Oryza sativa L.) grain yield at the district level across Tamil Nadu, India, for the Kharif and Rabi seasons of 2018–2022. A dataset of 1,240 district-season observations was assembled from the Tamil Nadu Department of Agriculture yield records, India Meteorological Department gridded climate data, MODIS NDVI time series, and the National Bureau of Soil Survey soil database. Ten predictive features were selected through mutual information-based feature ranking. The stacked ensemble architecture employs support vector regression (SVR), random forest, and XGBoost as base-level learners, with ridge regression as the meta-learner trained on out-of-fold base predictions. The proposed stacked model achieves root mean square error (RMSE) of 0.29 t/ha, mean absolute error (MAE) of 0.22 t/ha, and R² of 0.95 on the held-out test set, outperforming all individual base learners and a benchmark linear regression model. Seasonal residual analysis confirms consistent prediction accuracy across both Kharif and Rabi seasons, with no systematic bias attributable to seasonal climate differences. Feature importance analysis identifies seasonal rainfall, soil moisture, and temperature as the three dominant predictors, collectively accounting for 55.6% of model variance explained, with NDVI contributing 14.2% as the primary remote sensing-derived feature. The results demonstrate the practical viability of the stacked ensemble approach for district-level rice yield forecasting in data-scarce agricultural environments.