Hierarchical Ensemble Learning with Adaptive Sampling for Rice Phenological Stage Detection under Class Distribution Skewness

Abstract

Accurate identification of paddy growth stages from satellite imagery is a prerequisite for precision agricultural management, yet the inherent temporal imbalance across phenological phases introduces severe classification bias in conventional machine learning pipelines. This study presents the Hierarchical Ensemble Learning with Adaptive Sampling for Rice Phenological Stage Detection (HELASRPD), a two-phase framework that couples cluster-based oversampling (CBO) with SMOTE-ENN hybrid resampling in the first phase, and a stacked hierarchical random forest ensemble with phenology-aware weighting in the second phase. Two publicly accessible benchmark datasets, the IRRI Paddy Dataset (IRRI-PD) and the NDVI Remote Sensing Paddy Dataset (NDVI-RS), are used for comprehensive evaluation. Two novel evaluation metrics, Phase-Weighted F-score (PWF) and Imbalance-Adjusted Kappa (IAK), are introduced to overcome the limitations of standard accuracy measures under skewed class distributions. The proposed model achieves 92.7% accuracy and a macro F1-score of 91.4% on IRRI-PD, outperforming all six competing baselines. Ablation experiments confirm that each architectural component contributes measurably to overall performance. HELASRPD demonstrates that systematic imbalance correction integrated within an ensemble hierarchy yields robust and interpretable phenological stage mapping.