Non-destructive prediction and visualization of major chemical components in tobacco leaves using hyperspectral imaging

The chemical composition of tobacco leaves plays a crucial role in determining their aroma, flavor, and smoking quality. Hyperspectral imaging (HSI) offers a rapid, non-destructive means of detecting and visualizing key chemical constituents in tobacco leaves. In this study, 240 flue-cured tobacco samples from various grades and production areas in Yunnan province, China, were analyzed. Hyperspectral images were collected across the 967.05-2561.33 nm spectral range, and an improved region of interest (ROI) extraction method—combining spectral difference analysis with superpixel clustering—was proposed to effectively eliminate interference from background, leaf veins, and irregular structures. Spectral data were preprocessed using a combination of standard normal variate (SNV) transformation and first derivative (FD), followed by principal component analysis (PCA) for dimensionality reduction. Partial least squares regression (PLSR) models were developed under a unified modeling framework to simultaneously predict six chemical components: nicotine, total sugars, reducing sugars, total nitrogen, potassium, and chlorine. The results showed that the SNV+FD preprocessing strategy could effectively enhance model performance. The coefficients of determination for cross-validation (<italic>Q</italic>²) for nicotine, total sugars, reducing sugars, and total nitrogen all exceeded 0.89, with the lowest root mean square error of cross-validation (RMSECV) reaching 0.09. In an external test with 20 independent samples, the coefficients of determination for prediction (<italic>R</italic>²) were 0.930, 0.908, 0.854, and 0.915, respectively. The average relative deviation (RD) was less than 5%, and the residual prediction deviation (RPD) values were all above 2.5, which verified the stability and predictive capability of the models. The established models enabled pixel-level visualization of chemical constituent distribution, revealing distinct heterogeneity across different leaf regions. The proposed ROI extraction and modeling method provides an efficient, non-destructive, and visual approach for evaluating tobacco quality and supporting precision processing.