Two-Stage XGBoost-Based Classification of Chili Fruit Grades Using Weekly Vegetative Features

Accurate prediction of chili fruit grades prior to harvest is essential to support early-stage agricultural decision-making. This paper proposes a machine learning-based classification approach to predict chili grades—A, B, and C— using weekly vegetative features, such as plant height, leaf count, stem diameter, and growth stage. The modeling framework adopts the CRISP-DM methodology, encompassing data understanding, feature engineering, and model evaluation. Three models were developed and compared: a single-stage XGBoost, a multilayer perceptron (MLP), and a proposed two-stage XGBoost pipeline that first distinguishes Grade C from non-C, followed by classification between Grades A and B. Experimental results demonstrate that the two-stage XGBoost model achieved the highest performance, with an accuracy of 0.83 and a macro F1-score of 0.82, significantly outperforming both baselines in addressing class imbalance and feature overlap. Engineered features—particularly growth rate and interaction terms—proved critical in enhancing class separability and prediction accuracy. Despite the limitations and noise present in real-world data, the final model demonstrated strong generalization capability, stability, and robustness across all classes. These results highlight the effectiveness of hierarchical modeling, structured-data learning, and domain-informed features in addressing practical classification challenges. Ultimately, this work contributes to the development of low-cost, interpretable, and field-ready decision-support systems for smallholder farmers through accessible machine learning tools.