An adaptive multi-stage hybrid ensemble with iterative feature optimization for oil palm seed system analytics

Abstract

Oil palm seed systems involve sequential dependencies from upstream fertilization and productivity to germination quality and downstream distribution market outcomes, yet these stages are commonly analysed as isolated prediction tasks. This study proposes an adaptive multi-stage hybrid ensemble framework with iterative feature optimization for integrated oil palm seed system analytics. The framework is built from heterogeneous real-world data covering fertilization, production, germination, certified seed distribution, producer sales, and market share. These data are structured into three connected analytical stages: seed productivity, germination quality, and downstream distribution market analytics. At each stage, engineered features are refined through iterative feature optimization and modelled using a hybrid ensemble of Random Forest, XGBoost, LightGBM, and CatBoost, followed by adaptive weighting and stacking. The resulting predictions are then propagated across stages, allowing upstream productivity signals to inform germination modelling and both productivity and germination signals to inform downstream analytics. The results show that the oil palm seed system is more effectively represented as a connected multi-stage predictive structure than as separate stage-specific models. Ablation analysis further identifies iterative feature optimization and cross-stage propagation as the most influential components, particularly for downstream prediction.