Hybrid regression–classification framework with probability-guided decision refinement for green tea quality assessment using electronic nose

Accurate prediction in tea quality assessment is essential for ensuring consistency between production standards and market expectations. However, conventional classification models often struggle to distinguish subtle aroma differences captured by electronic nose (e-nose) sensors, particularly in borderline quality categories. To address this challenge, this study proposes a hybrid regression–classification framework that integrates ensemble tree-based classifiers with regression-guided decision refinement using probability-based thresholding. In the proposed framework, regression models produce continuous aroma quality scores that are interpreted through domain-defined thresholds, while classifier probabilities are analyzed to determine a reference probability, denoted as p Ref , identifying low-confidence predictions. Predictions below this threshold are subsequently refined using regression-based classification, enabling a confidence-guided decision mechanism. Experiments conducted on two independent datasets (Crop_1 and Crop_2) demonstrate that the optimized classifiers achieve strong predictive performance, with baseline balanced accuracies of 0.9827 and 0.9272, respectively. When integrated with the proposed hybrid framework, the balanced accuracy increases to 0.9846 on Crop_1 and 0.9311 on Crop_2 through improved specificity in low-confidence prediction regions. These results indicate that combining regression outputs with probabilistic classification enables more reliable decision behavior without degrading overall predictive performance. The proposed framework provides a practical and computationally efficient approach for intelligent sensor-based quality assessment systems in agricultural monitoring applications.