The hyperspectral data used in the experiment were collected using the GTM-900Pro tobacco comprehensive test bench (Chuangheyi Electronic Technology Development Co., Ltd. Shanghai, China). The GTM-900Pro consists of three modules: a sample visual inspection module, a near-infrared spectrometer (950–1650 nm, 351 bands), and the integrated industrial electronic control module (Figure 1).

1. Camera 2. Near infrared spectrograph 3. Scanning platform 4. Light source 5. Computer.

During data acquisition, the ambient lighting was kept constant to avoid interference from external light sources. To minimize the adverse effects of temperature and humidity fluctuations on spectral data acquisition, the room temperature was strictly maintained at 25 ± 1 °C throughout the entire data acquisition process. Tobacco samples were placed horizontally on the inspection stage, and black and white reference plate corrections were applied to the raw hyperspectral images to minimize noise caused by factors such as illumination and camera current variations. A standard reflective whiteboard was placed vertically in front of the imaging lens, and a single frame of whiteboard data corresponding to the current slit was captured for calibration. By attaching a lens cover to the lens, the black and white reference correction can be performed by acquiring the corresponding dark frame data. The calculations are detailed in Equation 1:

where R indicates the corrected sample image, Iraw denotes the original image of the sample, Idark signifies a blackboard correction image, and Iwhite represents a whiteboard correction image.

During spectral data acquisition, the instrument measures the transmittance at 81 points of each sample, through 81 holes arranged in a 9×9 matrix and calculates the average of these measurements to represent the spectral data of the tobacco piece. In traditional machine learning, a 70:30 or 80:20 split between training and test sets is commonly used. A 70:30 division has the advantage of not requiring an independent validation set, providing more stable evaluation results that can reduce the risk of overfitting. Therefore, in this experiment, 70% of the samples were randomly selected from each level as the training set, and 30% as the test set (Toleva, 2021). The sample set in this study comprised flue-cured tobacco leaves collected from eight counties/districts in Kunming over two consecutive years (2023 and 2024), thus capturing geographical and interannual variability to a certain extent. In addition, an independent test set with 535 samples was established, which was completely independent of the training data and had a consistent grade distribution with the latter. Therefore, no additional external validation is required to confirm the credibility of the developed model (Godin et al., 2015).

The test set accuracy reflects the ability of a model to generalize to unseen data and serves as a key performance indicator (Liu and He, 2025). The test set accuracy is the probability of correctly predicting an outcome across all test set samples, which is calculated as shown in Equation 6:

where TP denotes the number of correctly categorized positive samples; TN denotes the number of correctly categorized negative samples; FP denotes the number of negative samples incorrectly categorized as positive samples; and FN denotes the number of positive samples incorrectly categorized as negative samples.

A confusion matrix, also known as a likelihood or error matrix, is an important tool for comparing the classification results with actual values. It effectively represents the accuracy of the classification results and is widely used to evaluate classifier performance (Theissler et al., 2022).

The contents of nicotine (Nic), Cl⁻, K₂O, reducing sugar (RS), and total sugar (TS) in different tobacco grades were determined, and the corresponding sugar-nicotine ratio (S/N) and potassium-chloride ratio (K/Cl) were calculated (Table 1). As the primary contributor to the strength and satisfaction of tobacco products, the Nic content directly determines the core quality attributes of the product. Consequently, variations in Nic content across leaf positions and grades are essential considerations in the formulation of tobacco products (Darkis and Hackne, 1952). The Nic content of the upper leaves was significantly higher than that of the middle and lower leaves; however, the differences in the Nic content of B1F, B2F, and B3F were minimal. These differences can be precisely used in tobacco product formulations to meet different quality requirements. For high-end flue-cured cigarettes with a rich and strong impact, the formula prioritizes the use of upper leaves, leveraging their high Nic content to enhance smoking satisfaction and form the core flavor profile. B3F leaves are specifically incorporated into products such as low-tar cigarettes, providing a lower tar content while maintaining smoking satisfaction through their high nicotine content (McKinney et al., 2014). RS and TS are positively correlated with the smoothness of smoke; in particular, fructose and glucose can balance irritation by regulating the production of acidic substances. This property makes the differences in sugar content among leaf components and grades the core means for regulating smoke palatability in tobacco product formulation (Stepanov et al., 2025). The RS and TS contents of the lower leaves were significantly higher than those of the middle and upper leaves, with C3L and X2F exhibiting significantly higher RS and TS levels than the other grades. These differences are precisely applied in tobacco product formulations to achieve specific smoke quality objectives. For mid-to-low-end flue-cured cigarettes characterized by mildness and low irritation, the formula incorporates a larger proportion of lower leaves with high sugar content. It leverages the high RS and TS contents of lower leaves to enhance smoke smoothness, reinforcing the effect of balancing acidic substances and reducing irritation. For mid-range products aiming for moderate sweetness and smoothness without blandness, middle leaves can serve as the main component, blended with an appropriate amount of X2F-grade tobacco leaves. This not only ensures the sweet flavor of the smoke but also balances the taste layers and reduces irritation (Talhout et al., 2006). The S/N ratio directly affects the sensory acceptability and smoking behavior of tobacco by modulating the synergistic effects of Nic stimulation and sugar sweetness. This makes the variations in S/N ratios across leaf positions and grades a major consideration in the design of tobacco products (Roemer et al., 2012). The lower leaves had significantly higher S/N ratios than the middle or upper leaves, with C3L and X2F exhibiting significantly higher S/N ratios than the other grades. These differences are systematically used in the formulation of tobacco products. For flue-cured cigarettes with a mild, sweet, and low-irritation taste, the formula prioritizes tobacco leaves with a high S/N ratio, leveraging their strong synergistic effect to neutralize the irritation of smoke and enhance sensory acceptability (Nikolova et al., 2021). K₂O and Cl⁻ are key regulators of tobacco combustibility: K⁺ enhances combustibility by lowering the ignition temperature and promoting cellulose decomposition, while Cl⁻ inhibits combustion speed. Thus, the K/Cl ratio serves as a critical determinant of tobacco combustibility, with ratios exceeding 5 enhancing combustion and values below 1 potentially inhibiting ignition. The variations in K₂O, Cl⁻, and K/Cl ratio among leaf components and grades serve as the core basis for the precise regulation of tobacco combustibility in product formulation (Myhre et al., 1956). X2F had the highest K₂O content, while C1F had the highest Cl⁻ content. The K/Cl ratios of B1F, C2F, C3F, C3L, and C4F were comparable and significantly higher than those of the other grades. However, there was no significant difference in K/Cl ratios among the three leaf positions. These differences are strategically applied in tobacco product formulations to meet various combustion requirements. For high-end flue-cured cigarettes with easy ignition, even combustion, and white ash, the formula prioritizes the use of X2F, B1F, and C3L tobacco leaves while strictly controlling the inclusion of C1F-grade tobacco leaves. This ensures that the overall K/Cl ratio of the leaf blend remains above 5, thereby providing a smoking experience with smooth smoke release and minimal risk of extinguishment (Zhong, 2019). These differences in chemical composition not only distinguish the grades of first-roasted tobacco but also form the core basis for the design of tobacco product formulations in terms of product style, core quality, and consumer preferences.

PCA was performed on the chemical components of the collected tobacco leaves (Figure 3), revealing that the first two principal components cumulatively accounted for 63.8% of the total variance (PC1, 43.81%; PC2, 19.99%). Despite PC1 and PC2 explaining 63.8% of the total variance, leaves of different grades could not be clearly distinguished, likely due to the low dimensionality of the chemical composition data and the fact that tobacco grades are primarily differentiated based on physical characteristics such as maturity, leaf structure, thickness, and oil content. Although chemical composition underlies the physical appearance of tobacco and can influence it to some extent, the relationship between the two is complex. Therefore, chemical composition alone cannot be directly used to differentiate tobacco grades.

The spectra of nine tobacco grades were scanned in the range of 950–1650 nm, and their average values were calculated, with each spectrum consisting of 351 bands (Figure 4). The reflectance spectra of the tobacco leaves are generated through the interaction between light and the leaf tissue. The spectra exhibit distinct peaks at approximately 1180 and 1470 nm. The second overtone or combination frequency vibrations of C–H bonds in fructose, glucose, and portions of cellulose and pectin in flue-cured tobacco leaves give rise to a characteristic reflectance peak at 1180 nm (Wu et al., 2021; Liu et al., 2024). Additionally, surface shrinkage and texture homogenization in flue-cured tobacco leaves enhance the scattering effect at 1180 nm, further contributing to the characteristic reflectance peak. In addition, the low moisture content of flue-cured tobacco leaves reduces O–H bond absorption in water, resulting in a relative increase in reflectance near 1470 nm (Zhang and He, 2011; Liu et al., 2023).

A comparison of the average reflectance spectra (Figure 4) of nine grades of first-roasted tobacco revealed that the spectra of B1F and B2F were similar; those of B3F, C1F, and C2F were similar; and those of C3F, C3L, C4F, and X2F were similar. Aside from slight differences in color intensity and oil content, the appearance characteristics of B1F and B2F first-roasted tobacco leaves are largely consistent, resulting in minimal spectral variation and similar spectral profiles. The spectral curves of tobacco leaves of different grades show that the spectral reflectance of B1F, B2F, C1F, and C2F is lower than that of B3F, C3F, and C4F within the range of 1000–1400 nm. This is because B1F, B2F, C1F, and C2F have higher maturity and a looser leaf structure. In addition, B3F contains lower levels of RS and TS than B1F and B2F, while C1F and C2F have significantly higher nicotine content than C3F and C4F. These differences result in variations in the spectral reflectance of hydrogen-containing functional groups such as C–H and O–H, ultimately causing differences in the reflectance spectra between B1F, B2F, C1F, C2F and B3F, C3F, C4F within the range of 1000–1400 nm (Lu et al., 2023b).

From a molecular spectroscopy perspective, these observed spectral differences stem from the overtone and combination vibrations of hydrogen-containing functional groups, directly tied to leaf chemistry (Yang et al., 2015). Within this band, the spectral features of key chemical components in tobacco leaves stem from the vibrations of specific chemical bonds. For carbohydrates (e.g., RS and TS), characteristic reflectance originates from O–H stretching-bending combination bands and C–H vibrations. Higher sugar content generally results in weak reflectance at 1200–1300 nm and 1400–1500 nm, lowering overall spectral reflectance levels (Soares et al., 2019). For Nic, its molecular functional groups (pyridine ring, N–CH₃) exhibit significant absorption at 1100–1200 nm and 1300–1400 nm due to C–H and N–H vibrations (Yang et al., 2015). Consequently, the observed differences in average reflectance spectra result from the combined effects of the following factors: higher-grade leaves such as B1F and B2F generally contain higher sugar content and appropriate nicotine levels, leading to greater concentrations of O–H and C–H groups, stronger overall absorption in the 1000–1400 nm region, and thus lower reflectance. In contrast, the relatively lower sugar content in B3F weakens its O–H related absorption, while the higher nicotine in C1F and C2F enhances absorption from C–H and N–H groups. Variations in the concentrations of these specific groups directly determine differences in reflectance intensity at characteristic wavelengths. In summary, the patterns and variations in average reflectance spectra essentially reflect the combined influence of the concentrations of internal chemical groups (O–H, C–H, N–H) and the physical structure of tobacco leaves across different grades.

The lower leaves of first-roasted tobacco had the highest reflectance, followed by the middle leaves, with the upper leaves showing the lowest reflectance. Variations in the structures of tobacco leaves lead to differences in optical behavior. The upper leaves are usually thicker, and the cells are more tightly arranged, which results in reduced light penetration, shorter scattering paths, and lower reflectance. The middle and lower leaves are thinner with a higher proportion of spongy tissues, resulting in increased light scattering and higher reflectance. Significant differences were observed in the chemical compositions of first-roasted tobacco leaves. The Nic content was highest in the upper leaves and lowest in the lower leaves, while the RS and TS contents were lowest in the upper leaves and highest in the middle and lower leaves. Due to the strong absorption characteristics of hydrogen-containing groups (e.g., C–H, O–H, and N–H) in the 1200–1400 nm range, the differences in chemical compositions among leaf positions further amplified variations in their spectral reflectance (Soares et al., 2019). Therefore, different parts and grades of tobacco exhibit distinct spectral responses and characteristic spectral features.

PCA was performed on the collected hyperspectral data of tobacco leaves (Figure 5), which showed that the first two principal components cumulatively accounted for 96.4% of the total variance (PC1, 86.3%; PC2, 10.1%). PC1 is the primary factor for distinguishing tobacco leaf grades, encompassing most of the variance and highlighting the pronounced near-infrared hyperspectral differences between leaves of different grades. Near-infrared hyperspectral data have high dimensionality and contain abundant physical and chemical information of flue-cured tobacco leaves, including macroscopic traits such as moisture content, cell structure, and thickness, as well as the chemical information inferred from these characteristics (Marcelo et al., 2019). Additionally, the grades of flue-cured tobacco leaves are classified based on their external physical characteristics, which explains the high contribution rates of the first two principal components. However, PCA could not effectively differentiate between tobacco grades, likely because fluctuations in the O–H bond reflectance peak at 1470 nm, corresponding to water content, were much stronger than the grade-related chemical signals, such as the C–H bond at 1180 nm (Lu et al., 2021b). Moisture content was consistent across all nine grades and did not differ significantly between grades. Therefore, moisture content is unlikely to be a major driver of variation in the PCA or a key factor distinguishing grades. This overshadowed the spectral differences between grades, including those arising from structural variations, preventing PCA from effectively distinguishing tobacco grades.

Although first-roasted tobacco grades showed differences in spectral data and chemical composition, PCA tended to drown out the spectral differences between different grades. The dimensionality of chemical composition data was relatively low, making PCA ineffective for differentiation.

The PLS-DA model demonstrates consistently higher grade classification accuracy than other classification models across all three preprocessing methods, achieving an accuracy of 98.5% under MSC preprocessing. This is because there is strong collinearity among the bands in the near-infrared spectra. PLS-DA, however, extracts latent variables by maximizing covariance, automatically filtering out noise and emphasizing discriminative bands. Thus, the classification accuracy of PLS-DA is significantly higher than that of the other three models.

The experiment employed the confusion matrix as the performance evaluation index for the model (Figure 10). Values along the diagonal of the confusion matrix represent the proportion of correctly categorized samples, while values on the off-diagonal denote the proportion of incorrectly categorized samples. The classification accuracy for C1F, C4F, and X2F reached 100%, while that for B1F, C2F, and C3F averaged 98.0%, with only a very small portion of samples misclassified. B2F is similar to B1F in both appearance and intrinsic chemical composition, and the near-infrared spectra primarily reflect the overall chemical information of tobacco leaves. If different grades of tobacco differ only slightly in key components, their spectral features may overlap in the principal component space, making it difficult for the model to distinguish them. As a result, 2.1% of B2F samples were miscategorized as B1F. The slight differences in the physical structures of C3L and C3F in the middle leaves were not sufficiently reflected in the near-infrared spectra, leading to 2.1% of C3L samples being classified as C3F. Although a small number of misclassifications occurred, the confusion matrix demonstrates that the MSC-PLS-DA model has strong classification performance, accurately distinguishing between tobacco grades.

The MSC-preprocessed data were input into the SPA and CARS characteristic band selection algorithms, which selected varying numbers of characteristic bands. These numbers were then placed in ascending order according to the ratio of characteristic bands to the full wavelength range and fed into the grade classification model to evaluate the classification accuracy for different band counts. In the SPA, the classification accuracy of the RF model does not increase significantly as the number of feature bands increases (Table 3). The remaining three models have unsatisfactory performance when the number of feature bands is small. However, the classification accuracy of all three models improves as the number of feature bands increases, increasing by 20% for ELM and by 40% for BPNN and PLS-DA. In the CARS algorithm, the classification performance of all four models is unsatisfactory when the number of feature bands is small (Table 4). Similar to the SPA, the classification accuracy of all four models improves as the number of feature bands in the CARS model increases, increasing by 40% for RF, by 30% for BPNN and PLS-DA, and by 10% for ELM. When the number of selected feature bands is limited, the wavelengths chosen by SPA are predominantly concentrated in the spectral end regions. RF owing to its ensemble learning mechanism, exhibits strong robustness to locally concentrated high-information features, enabling effective extraction of discriminative information from such intervals (Bin et al., 2016). In contrast, both BPNN and ELM rely on nonlinear correlations among multiple constituent features, while PLS-DA depends on linear relationships across multiple bands to construct meaningful latent variables (Jiang et al., 2017; Shao et al., 2022). A single spectral interval is insufficient to meet the input requirements of these models, leading to relatively lower classification accuracy. Although the wavelengths selected by CARS cover key spectral regions associated with sugars and nicotine, its selection bias—rooted in the regression coefficients of PLS and compounded by the randomness introduced through Monte Carlo sampling—tends to discard signal bands that are critical for classification (Song et al., 2020). Consequently, the resulting feature subset may lack representativeness, hindering the ability of all classification models to extract effective discriminatory information. Therefore, when the number of feature bands is limited, all classification models based on bands selected by SPA (with the exception of the RF model) or CARS fail to achieve satisfactory classification accuracy.

In both BPNN and PLS-DA models, the use of characteristic band algorithms results in a significant improvement in classification accuracy as the number of characteristic bands increases. The performance of the BPNN and PLS-DA models is influenced by the information density and signal-to-noise ratio of the input features, while the SPA and CARS algorithms optimize feature subsets, significantly improving data quality and thereby enhancing model performance. Specifically, the SPA, with its low collinearity and small-scale feature subsets, aligns with the sensitivity of BPNN and the projection requirements of PLS-DA, thus significantly improving classification accuracy. In contrast, the CARS algorithm retains a broad range of wavelengths, making it more suitable for tasks requiring comprehensive spectral information. After applying the SPA characteristic wavelength selection method, the accuracy of the SPA-RF model remains essentially unchanged. When the number of selected feature bands reached 70% of the total spectral bands (i.e., 246 bands), the classification accuracy of the SPA-ELM and SPA-PLS-DA models decreased by only 1.7% and 4.5%, respectively, compared to the full-band models. These results demonstrate that a 30% reduction in data dimensionality (i.e., 105 bands)—achieved at the cost of a marginal drop in accuracy—significantly lowers the computational load and storage requirements of the models. This feature selection strategy effectively avoids the redundancy associated with full-band modeling while addressing the performance limitations of low-dimensional feature sets. It thus offers a practical and feasible solution for the lightweight deployment and engineering application of hyperspectral-based tobacco grading technology.

The PLS-DA model achieves a significantly higher classification accuracy than the other three models when classifying grades based on both full bands and selected feature bands. When the number of feature bands reaches 70% of the total, the classification accuracy of the PLS-DA model is comparable to that achieved using the full set of bands, indicating that PLS-DA is well suited for tobacco grade classification. In this study, a model developed by integrating hyperspectral technology with machine learning has classified nine different grades of first-roasted tobacco with an accuracy of 98.5%, achieving accurate discrimination across different parts and grades.

Compared with previous studies, this research mainly focuses on analyzing the near-infrared hyperspectral data of nine common grades of flue-cured tobacco leaves. Wei et al., 2024 used the 1D-CNN model combined with the LAR characteristic band algorithm to classify 10 grades of flue-cured tobacco leaves, achieving an accuracy of 96.3%. Similarly, Lu et al., 2021a used CNN to classify tobacco leaf grades based on global images and local patches, with a classification accuracy of 91.3%. In contrast, the classification accuracy of this study reached 98.5%, representing improvements of 2.2% and 7.2% over the aforementioned studies of Wei et al., 2024 and Lu et al., 2021a respectively. Furthermore, Wei et al., 2024 used the visible–near-infrared (401–1046 nm) spectral range, while Lu et al., 2021a performed a classification based on tobacco leaf images. This study has employed near-infrared hyperspectroscopy (950–1650 nm), which provides greater stability and richer spectral information, making it more suitable for classifying tobacco leaf grades.

In data-driven models for agricultural product quality classification, variations in ecological conditions, cultivation practices, and inter-annual environmental fluctuations across different regions can lead to shifts in spectral features. The core challenge arising from this phenomenon is that when a model is directly applied to crops from other regions or subsequent years, it may encounter novel feature patterns not sufficiently represented in the training data, consequently leading to a degradation in classification performance (Wang et al., 2025). Although the classification model developed in this study demonstrates excellent performance on the experimental samples, it is crucial to acknowledge this inherent limitation. The generality and robustness of the model are not absolute; its long-term and broad applicability relies on a sustained update mechanism. This entails continuously incorporating new samples from diverse regions and harvest years to recalibrate and retrain the model, enabling it to dynamically adapt to the spatiotemporal variability in tobacco leaf spectral characteristics. This limitation is not unique to our model but represents a common challenge that all data-driven agricultural product quality classification models must address to transition successfully into practical application (Wang et al., 2025).