Optimization of Support Vector Machines and Window Sampling for E-Nose-Based Classification of Black and Green Tea Aroma Bambang Heru Iswanto, Huffaz Muhammad Abdurrofi Baith, Abdullah Mu^adz Muflih, Haris Suhendar
Universitas Negeri Jakarta
Abstract
Aroma classification is critical for quality assurance and authentication in the global tea industry, where traditional sensory evaluation remains limited by subjectivity and inefficiency. This study aims to optimize both sampling window duration and Support Vector Machine (SVM) classifier performance for the objective, rapid classification of black and green tea aromas using a low-cost electronic nose (E-Nose). The custom-built E-Nose, integrating eight MQ-series gas sensors, measured volatile profiles from Indonesian black and green tea samples across three sampling windows-30, 60, and 90 seconds. For each sample, statistical features were extracted from sensor responses, and classification was performed using SVM models with linear and radial basis function (RBF) kernels. Model selection and validation employed leave-one-out cross-validation and grid-based hyperparameter tuning. Results show that a 60-second sampling window is sufficient for near-perfect classification (accuracy ≥-97.5%), with the linear SVM achieving perfect separation at 90 seconds. Principal Component Analysis (PCA) confirmed clear feature-based separation of tea classes. These findings demonstrate that rapid, objective, and reliable tea aroma authentication can be achieved using simple machine learning models and short sampling durations with a low-cost E-Nose
