Quality grading of Zijin cicada tea using a multi-level fusion strategy

Abstract

A Computer Vision System (CVS) and Near-Infrared Spectroscopy (NIRS) were employed for the quality grading of Zijin Cicada tea for the purpose of more scientifically discriminating the bite degree of tea green leafhoppers and substituting the subjective judgment of manual detection. The 163 samples were categorized into different bite degrees, namely, mild biting (grade B), moderate biting (grade A), and severe biting (grade C). Four machine learning models—the Adaptive Boosting algorithm (AdaBoost), Support Vector Machine (SVM), K-nearest neighbors (KNN), and Random Forest (RF)—were established using image, spectral information and fusion information, followed by dimensionality reduction of spectral data via Competitive Adaptive Reweighted Sampling (CARS), the Successive Projections Algorithm (SPA), and Principal Component Analysis (PCA), while image features were optimized using Linear Discriminant Analysis (LDA) and PCA. The results showed that the models based on feature-level, decision-level and hybrid fused information that combine machine vision and spectral technologies outperform single-sensor data in terms of robustness and accuracy. Using SVM, feature-level fusion (LDA-extracted image features + CARS-optimized spectral features) achieved 97.20 % accuracy using SVM. Decision-level fusion (LDA-SVM for images; PCA-RF for spectra) attained 98.15 % accuracy. Hybrid fusion combining LDA image features and SPA spectral features further improved the accuracy to 98.45 %. This study confirms that multi-level fusion of NIRS and machine vision provides an efficient, non-destructive solution for Zijin Cicada tea quality grading.