Advancing sustainable agriculture in Atlantic Canada through HYDRA-SE: A hybrid decision and regression stacked ensemble for yield prediction

Crop yield prediction is a fundamental component of global food security and sustainable agricultural planning. Potato, a key staple and cash crop in Atlantic Canada, presents challenges for yield prediction due to the complex nature of the relationships between soil parameters and crop production. In this research, we developed a machine learning (ML) framework that integrates multi-method feature selection and stacked ensemble learning to predict potato yield. The datasets consisted of soil analytical, geospatial, and productivity parameters from potato fields across Prince Edward Island (PEI) and New Brunswick (NB) during the 2017–2018 growing seasons. Stepwise regression-based feature selection identified eight significant features from 33 initial features that were included for training 16 baseline ML algorithms. The five top-performing models were used to further develop the proposed ensemble and stacked ensemble learning methods. Overall, the results indicated that the CatBoost, XGBoost, LightGBM, random forest (RF), and bagged trees outperformed other traditional ML methods. The proposed hybrid yield decision and regression approach- stacked ensemble (HYDRA-SE) framework achieved the highest accuracy (R 2 = 0.88 and RMSE = 3.75) . These results illustrate the significant potential of advanced ensemble learning methods to produce reliable and robust predictive capacity for potato yield prediction at a field level, regional level, and perhaps at the national level and can provide a decision-support system for potato precision agriculture (PA) in Atlantic Canada.