Ensemble machine learning with limited data: Feature selection and wheat yield prediction in Bangladesh

Predicting wheat yields accurately in smallholder farming systems in Bangladesh is challenging due to the complex interplay of various factors and the lack of high-quality field data. To tackle this problem, we created a stacked ensemble machine learning framework. It includes stepwise feature selection and model-agnostic interpretation techniques using data from 178 farmers in Bangladesh. We ensemble four main models: generalized linear model (GLM), random forest (RF), gradient boosting machine (GBM), and extreme gradient boosting (XGBoost). Our ensemble model outperformed individual models, reducing the root mean square error (RMSE) by 6 % compared to GBM and by 12 % compared to GLM. We found that ground cover, represented by well-covered plots (>60 %) and patchy ground cover plots (<60 %), is the most important factor for predicting wheat yields, followed by the wheat maturity date and the date of the previous monsoon rice harvest. These factors are easily observable and recordable by farmers and extension agents. These findings offer practical solutions for Bangladesh: (1) more reliable yield forecasts can assist with planning inputs and harvests; (2) we can develop digital tools that focus on simple-to-measure indicators like ground cover, which can be easily monitored with satellite images; and (3) this research can support government and NGO programs aimed at making wheat farming more climate-resilient. By demonstrating how ensemble learning can extract valuable insights from limited data, this study contributes to enhancing food security and promoting sustainable farming practices in South Asia’s resource-limited agricultural regions.