Quantitative analysis of tobacco blending proportions based on hyperspectral imaging and data fusion

The rapid and accurate detection of tobacco blending proportions is essential for quality control in the tobacco industry. This study proposes a method for the quantitative analysis of tobacco components based on multispectral fusion, integrating visible-near-infrared (Vis-NIR) and near-infrared (NIR) spectral data. The method employs the minimum covariance determinant (MCD) for anomaly detection and constructs a quantitative model using partial least squares regression (PLSR). The experimental data comprise two matrices of dimensions 400 × 90 and 220 × 90, each containing 90 samples. Experimental results demonstrate that multispectral fusion significantly improves the model's quantitative analysis performance compared to using a single spectrum. The adopted preprocessing strategy effectively reduces noise interference and enhances feature extraction capability. When predicting tobacco silk content, the fused spectral model achieved the highest prediction accuracy with R² of 0.8873. The innovation of this study lies in the proposed multispectral data optimization fusion and preprocessing strategy, which facilitates rapid detection of tobacco constituents and offers an optimal and efficient method. This approach provides a reliable technical solution and advances spectral detection technology in the tobacco and related industries.