Determination of nicotine in tobacco samples by near-infrared spectroscopy and boosting partial least squares

The fast and non-destructive quantization of nicotine is an important task in tobacco industry and near infrared (NIR) spectroscopy appears to provide a chance. However, the successful use of NIR spectroscopy has relied on chemometrics to construct a calibration model with satisfactory accuracy. This work investigated the feasibility of the combination of NIR spectroscopy and boosting partial least squares (boosting-PLS) for the determination of nicotine content in tobacco samples. Full-spectrum PLS and its two most representative modifications, i.e., uninformative variable elimination PLS (UVEPLS) and moving-window PLS (MWPLS) as well as the corresponding local versions of boosting-PLS, i.e., using boosting-PLS in the informative regions found by UVPLS and MWPLS were used for comparison. Prior to modeling, a total of 81 samples were first sorted and then split into 27 subsets, each containing 3 samples. By randomly picking one sample for testing out of each subset, 10 training/test sets were built. Based on a statistical comparison, it is revealed that both boosting-PLS and its local versions can improve the calibration but boosting-PLS behaves the best, meaning that boosting-PLS is not only feasible but can avoid doing a variable/interval selection in such a task.