Enhanced quality monitoring during black tea processing by the fusion of NIRS and computer vision

Abstract

Polyphenol and catechin are key components in black tea processing, contributing to both taste and color quality. However, the rapid detection methods that are applicable throughout the processing stages are lacking. Here, we explored the potential of miniature near-infrared spectroscopy and self-built computer vision. Fresh tea leaves, and the samples from withering, rolling, fermentation, and drying steps were collected for in-situ data acquisition in a tea factory. Data from two sensors were fused, competitive adaptive reweighted sampling and Pearson correlation analysis were employed to select effective variables from spectral and color variables, respectively. And the linear partial least squares (PLS) were used for modeling. The results showed that PLS models based on low-level data fusion could not effectively improve the prediction accuracies compared to single data. By contrast, middle-level data fusion achieved the best prediction accuracies for both polyphenol and catechin, with average root mean square error of prediction of 0.66 ± 0.12 and 1.06 ± 0.11 g/100 g, and residual prediction deviations of 5.41 ± 0.99 and 4.03 ± 0.38, respectively. Overall, this study demonstrated the enhanced predictive capability of fused spectral and imaging systems for polyphenols, overcoming the low predictive accuracy of single sensors.