# Prediction of Quality Substance Content of Hakka Stir-Fried Green Tea Based on Multiple Features of Near-Infrared Spectroscopy

## TL;DR

This study uses near-infrared spectroscopy and machine learning to predict the quality of Hakka stir-fried green tea based on its chemical content.

## Contribution

A novel method combining multiple NIRS feature extraction techniques and regression models for predicting tea quality indicators.

## Key findings

- The CARS + AFD + BC feature combination achieved the best overall prediction performance.
- Ridge regression outperformed PLSR for predicting theanine, tea polyphenols, and soluble sugar.
- PLSR provided better predictions for water extract content.

## Abstract

The contents of biochemical components, such as theanine, tea polyphenols, water extract, and soluble sugar in Hakka stir-fried green tea (HSGT), serve as important indicators reflecting the intrinsic quality of tea leaves. In this study, 171 HSGT samples are collected, and their near-infrared spectroscopy (NIRS), together with the contents of the four indicators, are determined. The aim is to establish prediction models for these four indicators by extracting multiple features from the NIRS data. First, the NIRS data is preprocessed. Then, multiple features are extracted using competitive adaptive reweighted sampling (CARS), adaptive Fourier decomposition (AFD), fast Fourier transform (FFT), continuous wavelet transform (CWT), and band combination (BC). Finally, ridge regression (RR) and partial least squares regression (PLSR) models are constructed based on the NIRS features to predict the four indicators. Experimental results show that the model combining multiple features, namely CARS + AFD + BC, delivers the best overall performance. Specifically, the RR model based on multiple features provides the most accurate predictions for theanine, tea polyphenols, and soluble sugar, while the PLSR model performs better for water extract. This study provides a rapid and accurate method for detecting the substance content in HSGT.