Purpose: Rapid prediction of tobacco nicotine content in tobacco industries has become essential to maintain a stable and reliable cigarette quality. This research deals with combining hyperspectral images (HSI) and chemometric models to predict nicotine content in powdered tobacco samples. Methods: Fifty-seven dried powdered tobacco leaf samples were scanned using a hyperspectral camera followed by image processing. The region of interest (ROI) was selected for calculating average spectra. The average spectra and the destructive measurements of nicotine concentration in the samples were used to develop four regression models based on partial least squares regression (PLSR), support vector regression (SVR), random forest (RF), and PLSR–variable importance in projection (PLSR–VIP). The models were evaluated using leave-one-out cross-validation (LOOCV) and on 15% test dataset. Results: The PLSR outperformed (R2=0.93, RMSE= 0.21%) SVR- and RF-based nicotine prediction models using the entire 970–1700-nm range. Five bands centred at 976.15 nm, 1452 nm, 1575.5 nm, 1592.3 nm, and 1698.9 nm were identified as effective wavelengths for nicotine content prediction and used by the PLSR–variable importance in projection (PLSR–VIP) model to produce satisfactory validation performance (R2=0.91, RMSE= 0.30%). The LOOCV yielded R2 values ranging between 0.89 and 0.93 for the evaluated models. Conclusion: The PLSR-VIP model with 96% fewer wavelengths than the full range PLSR highlighted its potential for a more simplistic nicotine prediction mechanism. The HSI plus chemometric model approach has shown the potential to predict tobacco nicotine content rapidly.
- Partial least squares regression
- Random forest
- Support vector regression
- Variable importance in projection