Machine learning-based traceability and sensory prediction of cigar tobacco leaves: A dual-dimensional data comparison of volatile and non-volatile components

Geographic traceability of cigar tobacco leaves (CTLs) is crucial for quality control; however, existing methods predominantly rely on single data types. This study integrated volatile organic compounds (VOCs) and non-volatile chemical components (NVCs) to establish a dual-dimensional tracing framework. By applying seven machine learning algorithms, we geographically traced the CTLs from four major Chinese growing regions (Yunnan, Sichuan, Hainan, and Hubei) and explored their predictive capability for sensory attributes. The key findings are as follows: (1) each region exhibits a unique chemical fingerprint; (2) model selection follows a data-environment adaptation principle. For VOCs-based models, linear models outperform nonlinear models in Yunnan, Sichuan, and Hainan, whereas nonlinear models dominate in the environmentally heterogeneous Hubei region. For NVCs-based models, linear models demonstrate superiority across all four regions; (3) in sensory prediction, the predictive capability of the two data dimensions is relatively limited, with only aftertaste and ash color showing slight predictive power, indicating that the relationship between these two data types and sensory attributes requires further exploration. By elucidating the adaptive relationship between data dimensions, environmental characteristics, and algorithmic performance, this study provides a methodological foundation for establishing a dual-dimensional traceability system for CTLs and offers a preliminary direction for future research on sensory prediction.