Electronic Nose-Driven Classification of Sri Lankan Teas: An Ensemble Learning Approach Based on Aroma Profiles

Geographical authentication of tea is critical for quality assurance, brand protection, and fraud prevention in the global tea market. However, conventional approaches based on sensory evaluation and chemical analysis are often subjective, labor-intensive, time-consuming, and costly, limiting their scalability for routine authentication. This study proposes an integrated electronic nose (E-Nose) and machine learning framework for the geographical origin classification of Sri Lankan tea using aroma profiles. Aroma data were collected from 367 dry tea samples representing seven major tea-growing regions using a custom-developed metal oxide semiconductor (MOS)-based Digi-Nose device equipped with seven gas sensors. The acquired sensor responses were digitally recorded and structured for supervised learning analysis. To ensure reliable and unbiased model training, the dataset was preprocessed and balanced prior to classification. Eight supervised machine learning models, such as ensemble-based (Extra Trees, Random Forest, XGBoost, CatBoost, LightGBM), distance-based (K-Nearest Neighbors), kernel-based (Support Vector Machine), and neural network-based (Multi-Layer Perceptron) approaches, were experimentally evaluated. Model performance was assessed using multiple metrics, including accuracy, precision, recall, F1-score, confusion matrices, learning curve analysis, and $\mathbf{1 0}$-fold cross-validation to verify robustness and generalization. Among all evaluated models, the Extra Trees classifier demonstrated superior performance, achieving an overall classification accuracy of 98.63 %, along with consistently high precision and recall across all tea regions. The findings confirm that the integration of MOS-based E-Nose sensing with ensemble learning techniques provides a reliable, non-destructive, and deployment-ready solution for aroma-based geographical authentication of tea, with strong potential for real-world industrial and regulatory applications.