PVKK-Stacking: A novel stacking method based on spectral data diversity and meta-feature optimization for adulteration analysis of Tieguanyin tea

Abstract

Tea adulteration detection poses a critical challenge in food safety regulation, particularly for oolong tea, due to the similar physicochemical properties among its varieties and the limitations of conventional identification methods. Spectroscopic techniques, with their non-destructive, accurate, and efficient characteristics, offer an effective solution to this challenge. However, given the complexity and diversity of adulterant compositions, traditional individual models based on spectral analysis generally suffer from limited prediction accuracy and poor cross-variety generalization, making them inadequate for detection across different adulteration scenarios. To address this technical bottleneck, a novel PVKK-Stacking method was proposed. This method combined multiple spectral preprocessing and variable selection methods to generate diverse spectral data for the base models, thereby provided comprehensive complementary information for the stacking ensemble learning. Furthermore, base models were pruned through K-means clustering and subsequently integrated using a KPLSR meta-model. This study focused on detecting adulteration in Tieguanyin, a classic oolong tea, using four other highly similar oolong tea varieties to prepare adulterated samples. Experimental results demonstrated that the PVKK-Stacking method achieved certain improvements over the compared individual model and ensemble methods, and exhibited exceptional stability, demonstrating reliable cross-variety generalization capability. Moreover, as an ensemble learning method, PVKK-Stacking has been optimized to reduce prediction time. The proposed method provides a highly reliable solution for detecting oolong tea adulteration, holding significant practical value in ensuring tea quality and safety.