Hyperspectral Imaging Technology for Non-destructive Detection of Tobacco Flue-cured Oil Content

In the leaf quality evaluation system, oil content is one of the important indicators to measure the quality of flue-cured tobacco. Traditionally, oil content evaluation mainly relies on the empirical judgment of professionals, which has problems such as strong subjectivity and relatively low efficiency. In recent years, hyperspectral imaging technology, due to its characteristics of combining graphs and spectra, has demonstrated application potential in the field of agricultural product quality detection. Taking a study on flue-cured tobacco oil content detection as an example, this paper introduces the practical application effect of visible-near-infrared hyperspectral technology in this scenario.

Research Background and Experimental Design
The study selected 634 flue-cured tobacco leaf samples from 22 tobacco-growing provinces (autonomous regions) across the country, covering upper, middle, and lower parts. The research team used the FigSpec series hyperspectral imaging system from CHNSpec (including FigSpec-23 and FigSpec-25 cameras) to synchronously collect the spectral information of tobacco leaves in the wavelength ranges of 400-1000nm and 900-1700nm. During the collection process, by fixing the light source angle and camera distance, illumination uniformity was ensured, and the average value after collecting spectral data twice for each sample was used as the original input.

The oil content score was independently evaluated on a 10-point scale by an appearance quality evaluation team consisting of 20 people. The samples were divided into a calibration set (443 samples) and a validation set (191 samples) at a ratio of 7:3. The distribution characteristics of oil content scores in the two sets of samples were consistent with the overall population, providing a reliable foundation for subsequent model construction.

Spectral Preprocessing and Correlation Analysis
The original spectral data contains noise and scattering interference, requiring preprocessing to enhance effective signals. The study compared five single preprocessing methods, including moving average smoothing (MA), multiplicative scatter correction (MSC), standard normal variate (SNV), first derivative (D1), and standardization (SS), as well as their combination strategies.

The analysis results showed that MSC and SNV preprocessing could effectively improve the correlation between spectral reflectance and oil content scores. In the wavelength range of 928.36-1177.03nm, the correlation coefficient increased from 0.076-0.124 of the original spectrum to 0.331-0.640. D1 preprocessing, by strengthening the local variation features of the spectral curves, made the number of strongly correlated bands (|r|≥0.4) exceed 100. These results indicate that reasonable preprocessing strategies help to improve the predictive capacity of subsequent models.

Model Construction and Performance Evaluation
The study used two algorithms, partial least squares regression (PLSR) and support vector regression (SVR), to construct quantitative prediction models for oil content scores. The PLSR model based on the full visible-near-infrared band had validation set RPD values between 1.642 and 1.775 under most preprocessing conditions, among which the validation set R² reached 0.683 and the RMSE was 0.346 after MA preprocessing. The SVR model had a validation set R² of 0.653 and an RMSE of 0.362 under the D1+SS combination preprocessing.

To merge the advantages of both models, the study introduced a weighted average fusion strategy. The fusion model based on the full visible-near-infrared band (PLSR under MA preprocessing and SVR under D1+SS preprocessing) saw its validation set R² increase to 0.721, RMSE drop to 0.324, and RPD reach 1.894, showing a better prediction effect than any single model.

Characteristic Band Selection and Model Optimization
Hyperspectral data contains hundreds of bands, presenting data redundancy problems. The study adopted the successive projections algorithm (SPA) for characteristic band selection. The results showed that after MA preprocessing, the PLSR model constructed with 95 characteristic bands selected by SPA had a validation set R² of 0.685 and an RMSE of 0.345; after D1+SS preprocessing, the SVR model constructed with 56 characteristic bands selected by SPA had a validation set R² of 0.666 and an RMSE of 0.355. The number of characteristic bands was significantly reduced from 428 in the full band, drastically lowering the data dimensionality.

The fusion model of PLSR and SVR based on SPA selection further improved the prediction accuracy, with the validation set R² reaching 0.724, RMSE at 0.323, and RPD at 1.904. This result indicates that characteristic band selection maintains model validity while reducing data redundancy.

Application Outlook
This study demonstrates the feasibility of visible-near-infrared hyperspectral technology in the non-destructive detection of flue-cured tobacco oil content. Compared with traditional manual evaluation methods, hyperspectral technology has potential advantages of objectivity, non-destructiveness, and speed, which can provide reference bases for the development of automatic tobacco leaf grading equipment and the construction of intelligent quality control systems. The FigSpec series hyperspectral imaging system from CHNSpec undertook the core data collection task in this study, verifying its applicability in agricultural material quality detection scenarios.