Multi-basis continuous wavelet transform feature wavelengths selection and machine learning with hyperspectral imaging for non-destructive prediction of 1,2-propylene glycol content in cased tobacco leaves

The process of manufacturing tobacco casings constitutes a critical step in the production of tobacco leaves, exerting a substantial influence on the improvement of their physical and chemical characteristics, and consequently, the quality of the final product. Nevertheless, the prevailing approach to casing accuracy detection is predominantly focused on dosage monitoring, with a paucity of attention being paid to real-time effect evaluation. Current hyperspectral-based detection systems encounter difficulties in extracting trace additive features and managing high-dimensional data under limited sample conditions. A multi-basis continuous wavelet transform (CWT) and machine-learning-integrated framework for non-destructive propylene glycol (PG) content prediction were proposed in this paper, addressing precision limitations in tobacco quality monitoring. The preprocessing of hyperspectral imaging data from six PG concentration levels were undertaken via Savitzky-Golay filtering, followed by multiscale decomposition using three CWT basis functions with morlet, Mexican hat, and Gaussian wavelets. A dual optimization mechanism combining correlation threshold filtering and wavelength frequency statistics was developed to enable efficient feature wavelength selection. Furthermore, a stacking regression model was constructed and compared with standalone algorithms. The results demonstrated that the multiscale combined MMG strategy achieved 79.86% dimensionality reduction by selecting 58 feature wavelengths covering adjacent regions in the near-infrared and short-wave infrared (NIR-SWIR) range, significantly enhancing model generalization compared to full-spectrum inputs. Additionally, the stacking regression model attained optimal performance, with a testing set coefficient of determination of 0.9704, under combined MMG input through synergistic complementarity of heterogeneous base learners, with a root mean square error of 0.3188. It is confirmed in this paper that spectral feature interpretability is improved by multi-basis CWT decomposition through complementary wavelet-scale responses, and a novel, to our knowledge, non-destructive PG detection paradigm for industrial tobacco processing is established by the framework. Transferable insights for the hyperspectral analysis of trace components in agricultural products are provided by the methodology, and it could be applied to the non-destructive detection of trace additives for tobacco quality control.