Tobacco Mosaic Virus (TMV) and Potato Virus Y (PVY) are widespread virus diseases in fields and cause massive economic losses to crops (Scholthof et al., 2011; Khateri et al., 2014; Korbecka-Glinka et al., 2021). Both PVY and TMV infect a wide range of plants, especially tobacco and other members of the family Solanaceae, causing symptoms such as leaf mosaic, vein clearing, and deformation (McDonald and Singh, 1996; Yang and Klessig, 1996; Quenouille et al., 2013). The key to effective viral disease control is to directly monitor the occurrence and prevalence of diseases. However, the traditional methods of visual or molecular identification are time-consuming, inefficient, and destructive. At the same time, TMV and PVY infections are difficult to be separated and could develop into severe symptoms rapidly under an ideal environment, which leads to missing the best control period ( Figure 1B ). Therefore, automatic identification of the disease occurrence and severity degree of plants in the field in time will be of great benefit for precise prevention by guiding the chemical application where and when needed at an appropriate dose, further controlling the spread of TMV and PVY in time and avoiding great production loss. Furthermore, the automated disease identification can be integrated into innovative disease-resistant plant breeding process, expediting the phenotyping process and yielding time savings compared to the visual assessment by human raters.

Modern agriculture has benefited greatly from the high-tech vision solutions such as artificial intelligence, machine learning, etc. For example, spectrum technology has often been used in precision agriculture to fill gaps in continued human monitoring. Spectral reflectance captures crop biomass, disease information, and crop quality (Gnyp et al., 2014; Martínez-Martínez et al., 2018). The principle of spectroscopic technology is to identify the content and composition of substances using the different characteristics of different substances such as light absorption and reflection (Fang et al., 2015). In particular, at the onset of PVY and TMV infection, leaf structural characteristics and chlorophyll levels begin to change. These changes will further trigger the reflection spectrum. Hence, taking advantage of these fluctuations in the reflection spectrum, we could carry out disease detection and monitoring thanks to modern techniques avoiding irretrievable yield loss caused by missing the best control time. The advantages of applying spectral technology to monitor crop diseases are fast, non-destructive, and wide-area detectable. It has been verified in previous studies that the usage of leaf spectral information could effectively monitor and distinguish leaf disease (Huang et al., 2012; Hu et al., 2016; Kong et al., 2018; Martínez-Martínez et al., 2018; Long et al., 2021; Cai et al., 2022; Fernández et al., 2022). However, there are scarce studies to explore the application value in the detection of virus diseases.

Furthermore, searching for sensitive wavelength bands is the focus of spectroscopic technology applications. The optimal wavelength of the leaf spectrum can be quickly and accurately located by machine learning (Steddom et al., 2005; Li et al., 2022; Lv et al., 2022). Machine learning algorithms can process a large number of data sets with irregular surfaces, find the potential probability distribution of the data, and make predictions, which can be used for the diagnosis and prediction of crop diseases. Machine learning is mainly classified into supervised classification and unsupervised classification. Supervised classification algorithms include support vector machine (SVM), random forest (RF), decision Tree, KNN algorithm, linear regression, and other methods often used in the binary classification regression analysis. Unsupervised classification algorithms are mainly used for cluster analysis. Most of the applications have been implemented using supervised variants of machine learning algorithms rather than unsupervised ones. Besides, among the above algorithms, RF training speed is fast, having strong model generalization ability. The SVM algorithm is suitable for analyzing finite samples, overcoming the shortcomings of some other binary support vector machines, and improving multi-classification accuracy (Uddin et al., 2019).

Machine learning holds immense potential for enhancing accuracy. Hence, the fusion of spectral data and machine learning techniques has been harnessed for leaf disease diagnosis (Lamba et al., 2021). Nevertheless, spectral data often carries inherent noise, making the careful selection of appropriate algorithms and models paramount in the accurate identification of tobacco diseases. By analyzing the full-length spectra (350-2500nm) of healthy and diseased tobacco leaves, this study aims to establish robust and effective classification models for TMV and PVY diseases by two classifiers: support vector machine (SVM) and random forest (RF) ( Figure 1 ). The outcomes of this research will be helpful to facilitate the early-stage disease type and the severity level assessments, scientifically informed strategies for preventing and controlling leaf diseases.

Spectroscopy has been used in precision agriculture for fast and non-destructive determining crop disease epidemic situations (Martínez-Martínez et al., 2018; San-Blas et al., 2020; Cai et al., 2022; Fernández et al., 2022; Li et al., 2022; Lv et al., 2022). Machine learning in leaf disease recognition is superior to traditional methods such as regression and clustering analyses (Ingale and Baru, 2019). Combined with automatic extracting spectral characteristics to improve the precision of the model, machine learning could be better applied to precision agriculture to guide the diagnosis of agricultural diseases. Here, these two technologies (hyperspectral and machine learning) were utilized to diagnose the type and severity degree of virus diseases PVY and TMV on tobacco. The optimal model for plant virus disease diagnosis was explored by setting up six combinations with three spectral data preprocessing methods (MSC, SNC, and SavGol) and two machine learning classifiers (SVM and RF).

Overall, all the models have excellent capabilities in the identification of the type and infection severities of TMV/PVY diseased leaves. In the binary classification model, the SVM classifier performed better compared to RF with over 97% precision ( Table 2 ). Using an SVM classifier, Fernández et al. (2022) also achieved a good separation effect of yellow powder disease four days after cucumber leaves inoculation, and the overall accuracy rate was above 95% (Fernández et al., 2022). For the six-class classification, the performance of the SVM classifier is also obviously good and the best combination is SavGol+SVM ( Table 2 ). However, our results illustrate that different classifiers have distinct diagnostic precisions for different diseases, which is similar to the diagnosis result of rice diseases using SVM (Sethy et al., 2020). The errors of different classifiers were not completely consistent either ( Figures 6 , 8 ), which might be caused by the slight deviations from the original manual definition of severity grades.

Lastly, machine learning was also used to automatically extract spectral features (Barradas et al., 2021). The results illustrated the effective wavelengths for identifying diseases located in visible and near-infrared light bands. In detail, the characteristic wavelengths for identifying PVY diseases were concentrated in the vicinity of 700nm, which was similar to the sensitive wavelength of other leaf diseases (Long et al., 2021). A previous study discovered a strong correlation between the reflectance of the band near 700nm and chlorophyll content, carotene, and other leaf pigments (Chen et al., 2012). Besides, when the leaf is infected by disease pathogens, the visible area will have higher reflectance (Shi et al., 2009; Sankaran and Ehsani, 2013; Wang et al., 2014). The characteristic wavelength of TMV-diseased leaves captured by SVM is near 1800nm. The near-infrared band is related to the internal structure and dry matter of leaves. Therefore, the change in the internal structure of leaves infected by the virus will change the spectral reflectance. But most studies were focused on the spectral wavelength range of 380–1023nm, which only found a similar effective wavelength of around 700nm (Xie et al., 2015; Zhu et al., 2017; Sethy et al., 2020). Here, the full-length spectra study revealed the potential candidate effective bands in the NIR region. Based on the captured effective bands, it is clear that even the spectral reflectance resolution became higher after MSC and SNV treatment in the NIR, but the identification abilities of the models with MSC or SNV were still lower than models with SavGol.

In yet other words, different spectral features could be caused by different diseases, while various feature bands might be caught by diverse classifiers for the same disease further leading to irregular classification accuracies (Kong et al., 2018). Therefore, all the combinations using SVM as a classifier captured the same effective bands of the same disease, while the combinations including the RF classifier had a huge variation in the effective wavelength capture for the same disease leaves. This may be one reason that the RF classifier has underperformed in this scenario.

Given the precise plant disease management, early detection plays a pivotal role in guiding timely interventions and preventing potential losses in production. Identification of the disease type proves invaluable in selecting the proper control strategies and expediting the breeding process. Moreover, the rapid development speed of tobacco viral diseases in the field underscores the challenge of sustaining minimal impact over time. Thus, assessing infection severities becomes crucial in choosing the appropriate intensity of control efforts. This study facilitates a comprehensive investigation into a rapid and non-invasion diagnostic model of the type and severity grades of two important virus disease -TMV and PVY, and further explore the optimal classification models. These classifiers process the categorization of tobacco-diseased leaf severity by capturing the feature wavelengths, paving the way for future large-scale promotion and application. Take, for instance, unmanned aerial vehicle (UAV)-based hyperspectral platforms, which have gained significant prominence due to their lightweight, flexibility, and ease of operation for detecting plant diseases (Vanegas et al., 2018). Nevertheless, there are still certain limitations to consider. Specifically, the acquisition cost of hyperspectral data remains relatively high, regardless of financial or labor expenses. External factors can also influence image quality during data collection, including factors such as measurement timing, light intensity, solar altitude angle, and more.

To sum up, all the classification models examined in this study demonstrated commendable performance in distinguishing PVY and TMV diseases in tobacco. Moreover, the SVM classifier did a better job than RF in the binary and six-class classification of PVY and TMV-diseased tobacco leaves. Additionally, the synergy between the SavGol preprocessing method and the SVM classifier yielded exceptional precision rates exceeding 96% across all classification tasks. In the light of feature wavelengths caught by SVM, specifically the 700nm band of PVY diseased leaf and the 1800nm band of TMV diseased leaf holds significant promise for the development of PVY and TMV disease classification model in the future large-scale monitoring, such as UAV spectral detection. In short, the integration of hyperspectral technology and machine learning offers a promising avenue for early detection of PVY and TMV disease leaves to achieve effective crop management.

The original contributions presented in the study are included in the article/supplementary material. Further inquiries can be directed to the corresponding authors.

XW, HC, YH, JW, and WX designed the experiments. HC, YH, YL, DL, LJ, KH, and HW performed the experiments. HC, YH, XW, JW, and WX analyzed the data. YH and HC wrote the manuscript with input from LG, XW, JW, and WX. All authors contributed to the article and approved the submitted version.