Tomato (Solanum lycopersicum L.) plants of the cultivar Cherry were grown in 200 mL pots containing a commercial potting substrate, in a walk-in plant growth chamber under controlled conditions (25-27 °C, humidity of approximately 60%, photoperiod of 12/12 h and light intensity 30W). Plants were divided into three groups of three plants each (nine plants in total), being a) one group of plants inoculated with Pseudomonas syringae pv. tomato DC 3000 (Pst) bacteria, b) a second group of plants inoculated with Xanthomonas euvesicatoria LMG 905 (Xeu) bacteria, and c) a third group of plants was treated with sterile distilled water only (Control group) ( Figure 1 ). Plants were physically separated to avoid cross-contamination.

Plants were inoculated in the laboratory, at the growth stage of 5-6 fully expanded leaves, by spraying until they became fully wet, and run-off occurred. The bacterial suspensions used for these inoculation assays consisted of 1 x 10⁸ cells/mL. They were prepared from 48-h-old cultures of Pst grown in KB medium (peptone, 20.0g; K₂HPO⁴, 1.5g; MgSO⁴, 1.5g; glycerol, 10 mL; agar, 15g; distilled water up to 1.0 liter), and of Xeu cultures grown in YDC medium (yeast extract, 10.0g; dextrose, 20.0g; CaCO₃, 20.0g; agar, 15.0g; distilled water up to 1.0 liter). The inoculated plants were then covered with transparent polythene bags for 48 h to increase the relative humidity that fosters bacterial entry into plant tissues through natural openings such as stomata (Lamichhane, 2015). Plants were monitored daily for symptom development for 18 days ( Figure 1 ).

During the inoculation period, to verify if the bacteria cultures used in these inoculation tests were viable, 20 μL of Pst solution and 20 μL of Xeu solution were cultured in Petri dishes containing KB and YDC media, respectively. After 48 h was possible to observe the bacteria growth in both nutrient media, proving that bacteria were viable at the moment of inoculation.

After the last spectral measurement, sample preparation for bacterial isolation was performed for all the leaflets. Leaflets were excised from plants using a sterile scalpel (Fernandes et al., 2017). Bacterial isolation was carried out as defined by Fernandes et al. (2017; 2021). Briefly, each sample of excised leaflet tissue was disinfected by immersion in 70% ethanol followed by washing with sterile distilled water (SDW), and then macerated with SDW in extraction bags. The suspensions obtained, and corresponding dilutions, were streaked on KB (samples inoculated with Pst bacteria), and on YDC medium (samples infected with Xeu pathogen). Characteristic colonies from these two bacteria species (milky white colonies in the case of Pst, and mucoid yellow colonies in the case of Xeu) were selected for growth on fresh nutrient agar medium to ensure purity.

Pst characteristic symptoms resemble small greasy dark stains (circular or slightly angular), that become brown to black, and appear randomly on the leaflets (often on the youngest or the ones located at the edge of the canopy plant). These lesions may typically show a yellow halo of various sizes. They are about 2–3 mm and can develop and coalesce (especially in the presence of moisture), affecting large areas of the leaf, that may later become necrotic and desiccate (Blancard, 2012). In turn, Xeu characteristic symptoms comprise small, circle, or slightly angular, translucent, and water-soaked lesions, which turn brown with time. They appear randomly in leaflets, and eventually become necrotic spots, with light gray centers and dark margins, which also can become surrounded by a yellow hallow with time. Smaller lesions can coalesce into each other forming larger injuries, whose diameter can range from 2 to 3 mm. In severe cases, tissues in the center of a lesion become dry and fall out, leading to “shot-hole” symptoms (Ritchie, 2000; Blancard, 2012).

A colony PCR was performed to validate the presence of both bacteria species on tomato leaflets isolates. PST2 (Vieira et al., 2007) and XV14 (Albuquerque et al., 2012) were the chosen markers, for Pst and Xeu, respectively, with amplicon lengths of 200, and 713 bp, correspondingly. A 20 µL PCR reaction mix consisted of 1 × DreamTaq Buffer (ThermoFisher Scientific, Waltham, MA, USA), 0.2 mM of each deoxynucleotide triphosphate (dNTP) (Grisp, Porto, Portugal), 0.2 mM of each forward and reverse primers, 1 U of DreamTaq DNA Polymerase (ThermoFisher Scientific, Waltham, MA, USA) and 10 µL of DNA isolate solution. Sterile distilled water was used as the negative control. PCR cycling parameters were defined as stated by Vieira et al. (2007) for Pst, and Albuquerque et al. (2012) for Xeu. PCR products were then separated by electrophoresis on a 0.8% agarose gel (1 × TAE buffer) and visualized using Xpert Green DNA stain (Grisp, Porto, Portugal) with a Molecular Imager Gel Doc XR+ System (Bio-Rad, Hercules, CA, USA).

Figure 1 presents the main procedures for spectral measurements in the experimental setup. Hyperspectral point-of-measurements (HS-POM) were collected in vivo from the adaxial side of healthy and diseased leaflets of the nine tomato plants in the study, in a dark room. For each plant, spectral assessments were performed randomly on nine points of different leaflets, belonging to the 4^th, 5^th, and 6^th expanded leaflets.

Hyperspectral data were acquired using a Hamamatsu Photonics K.K. TM Series C11697MB spectrometer, which covers a wavelength range of 200-1100 nm with a spectral resolution of 0.6 nm. A transmission optical fiber bundle (FCR-7UVIR200-2-45-BX, Avantes, Eerbeek, The Netherlands) with a range of 200-2500 nm was used along with a stainless-steel slitted reflection probe that was positioned 0.5 cm above the sample surface to capture the leaflet’s spectral signal and direct it to the spectrometer’s entrance lens. A white LED light was placed underneath the leaflet to provide uniform illumination to its entire abaxial surface. The spectral range of the LED emits light from 390 to 800 nm. Therefore, the LED spectra were used as a reference to the spectral range measured by the spectrophotometer and to check measurement and light emission stability ( Figure 1B ). The hyperspectral data were collected using specialized evaluation software (SpecEvaluationUSB2.exe, Hamamatsu Photonics K.K., Japan).

The performance of the modeling approach in detecting bacterial diseases in tomato leaflets was assessed using only the spectral region of 400 to 800 nm, approximately. This decision was based on the spectral wavelength range of the light LED source used (where possible useful information could be retrieved) and due to the observation of spectral noise near the limits of the equipment’s spectral range, which could negatively affect the performance of the classification process. Therefore, a total of 944 features (wavelength) were used in the development of the prediction modeling ( Figure 2 ).

Preprocessing data was performed following spectra normalization ( Figures 2 , 3 ). This approach aimed to standardize the data to a common scale, enabling meaningful comparison and analysis across different scenes or datasets. Additionally, it aims to decrease spectral signal oscillations, related to measurement equipment specifications (including devices’ internal noise), variations in data assessment conditions (comprising differences in global spectral trend, total energy, high-frequency noise, and/or local background) (Randolph, 2006), associated to changes in environmental conditions or induced by the operator in the moment of assessment (e.g. variations in sample-sensor distance, uneven illumination conditions, choice of leaflets sample point location, appropriate scan parameters, spectral calibration, among others). This results in model abilities improvement by aiding in class separation (Randolph, 2006; Guezenoc et al., 2019). Furthermore, this process enables the elimination of the spectral response of both the sensor and light source, making possible the transfer of the acquired classifier to a different sensing device. Spectral data retrieved from measurements in tomato leaflets S ( λ n ) m r a w were normalized through their division by the white LED source spectral signature S ( λ n ) r e f e r e n c e (considering the time of exposure of the spectral measurements), through the computation of the following forming (Equation 1):

Seeking bacterial tomato disease classification, spectral signatures from leaflets were then grouped to perform an applied predictive modeling approach related to the plants’ experimental condition. Leaflets were classified according to the plant treatment group and their health status, including the classes: i) healthy, including all the measurements which were performed before bacteria inoculation, and the remaining assessments that were made in non-inoculated plants considered as control plants; ii) non-symptomatic Pst; iii) non-symptomatic Xeu; iv) symptomatic Pst; and v) symptomatic Xeu. All the spectral data collected from tomato leaflets on different dates were unified in a single classification model ( Figures 1 , 2 ).

Data classification was, thus, performed seeking the unraveling of spectral phenotyping differences between i) healthy and non-symptomatic diseased tissues (early diagnosis), ii) healthy and diseased tissues (showing visual modifications due to changes in chemical and structural composition), ii) healthy and diseased tissues affected by different bacterial etiological agents (which present distinct host-pathogen specific interactions), iii) and diseased tissues infected by different bacteria species (responsible for causing similar visual symptoms but showing different pathogenic dynamics).

Multi-scale interference in plants’ tissue promotes superimposition on hyperspectral data, resulting in autocorrelations in their spectral signal at several scales (Martins et al., 2022). To mitigate the effects of high dimensional, redundant information, several methodologies have been cited in the state-of-the-art, including dimensionality reduction (DR) approaches (Lapajne et al., 2022; Reis-Pereira et al., 2022). DR techniques are a class of predictor transformations. They can reduce data by creating a minor set of predictors that aim to retain most of the information contained in the original variables. Usually, these approaches generate new predictors which are functions of the original ones (signal extraction or feature extraction techniques) (Kuhn and Johnson, 2013).

This study examined a DR approach called Linear Discriminant Analysis (LDA), generally computed as a pre-processing. It is a supervised learning algorithm used for classification tasks. LDA is usually applied as a feature extraction technique, performed to reduce the dimensionality of the data while maximizing the class separability. It projects the high-dimensional data onto a lower-dimensional space while preserving the discriminative information between classes. In brief, data is projected onto a linear subspace that maximizes the ratio of between-class variance to within-class variance. Thus, the projected data points are as far apart as possible in the new space, while the points belonging to the same class are as close as possible. Therefore, LDA contributes to reducing the problem’s computational complexity and avoiding overfitting. It can also be useful for visualizing the data in a lower-dimensional space, helping interpret patterns in data (Tharwat et al., 2017). Furthermore, this technique was applied since our dataset is not linearly separable, and LDA can organize it in another space with the maximum possible linear separability (Sachin, 2015).

LDA feature space loadings (also called coefficients or weights) were additionally used to infer the most relevant wavelength variables, through the computation of the interquartile range (IQR) for the weights. A threshold at 1.5 times the IQR beyond the upper quartile was established. This process aimed to increase sensitivity to the weight distribution, enabling the capture of outliers and extreme values. An applied predictive classification model was later computed to help deal with the non-linearity of the data.

A Support Vector Machines (SVMs) algorithm was chosen to integrate this modeling strategy. This supervised machine learning algorithm performs classification based on the concept of optimal separating hyperplane (Vapnik, 1999; Mosavi et al., 2021). SVMs are nonlinear approaches that discover the most extensive margin between two classes in feature space. These approaches aim to decrease the error test and model complexity (Ballabio and Sterlacchini, 2012). SVMs can present distinct hyperparameters and kernel forms, which convert raw data inputs from the original user space into kernel space through a user-defined feature map (Patle and Chouhan, 2013; Ding et al., 2021). This study used a radial basis function (RBF) kernel was used since it allows SVMs to capture non-linear relationships between input features and target variables. It may also allocate distinct weights to different points since they learn the decision surface according to the relative importance of the data points in the training set (being well-suited for handling outliers and noisy data) (Xulei et al., 2005). Moredetailed information about the SVM algorithm, including relevant principles and calculation formulas, can be found in Ballabio and Sterlacchini (2012) and in Chang and Lin (2011). The parameters of the SVM applied corresponded to the default values of the algorithm implemented in the ‘Scikit-learn’ machine learning library (Pedregosa et al., 2011), which also can be found in Table 1 .

0The datasets were divided into training data (70% of random observations) and validation data (30% of the remaining observations) (Kuhn and Johnson, 2013), following a holdout method (Lantz, 2019). The training and validation sets combined the pairs of concurrent measurements of the group and health status and the corresponding values of the predicting variables. A resampling strategy was performed as stated in Reis-Pereira et al. (2022) to reduce the possibility of overfitting (Berrar, 2019; Valier, 2020).

Different metrics were additionally retrieved to investigate model performance, namely the Confusion Matrix (CM), accuracy score (Equation 2), and F1-Score (Equation 3) whose description is detailed in Reis-Pereira et al. (2022). Furthermore, precision (the fraction of correct positive predictions out of all positive predictions, Equation 4) and recall (fraction of correct positive predictions out of all observed positive samples, Equation 5) were also computed using the following formula, where true positive, false positive, false negative, and true negative values are denoted by TP, FP, FN, and TN, respectively:

All the computational analyses were performed in the Jupyter Notebook software using the libraries ‘Matplotlib’ (Ari and Ustazhanov, 2014), ‘numpy’, ‘pandas’ (Betancourt et al., 2019), and ‘Scikit-learn’ (Pedregosa et al., 2011).

Tomato plants were inoculated with Pst and Xeu bacteria, respectively. After spectral analysis, leaf samples from each treatment were tested for the presence of these bacteria. Proper controls from samples known to be positive and negative for Pst and Xeu bacteria were included to confirm the assay results. After the colony PCR reaction, the amplified products were separated by agarose gel electrophoresis and visualized under UV light. The PCR results showed bacteria-specific bands for each bacteria species, namely a 200-base pair (bp) fragment of Pst, and a 713 bp fragment for Xeu, indicating that Pst and Xeu bacteria were present in each inoculation treatment group. No PCR amplification was observed from samples collected from healthy leaves.

Tomato plants infected with Pst bacteria showed the first visual typical chlorotic symptoms mostly between four and five days after infection (DAI). These spots evolved into necrotic lesions at six to seven DAI. In turn, chlorotic lesions in samples inoculated with Xeu mainly developed among twelve to fifteen DAI, only evolving to the necrotic stage at seventeen to eighteen DAI. Pst-infected plants died 12 DAI ( Figure 4 ).

Table 2 presents the dataset structure used, composed of 3478 spectral point measurements, from which 1377 (39.6%) observations correspond to the healthy class. Of these, 1215 (34.9%) assessments belonged to Control leaflets, 81 to measurements performed on Pst leaflets before bacteria inoculation, and 81 to captures made on Xeu leaflets also before bacterial infection. Spectral records performed before symptom appearance reached the value of 844 (24.3%), where 101 (2.9%) measurements belonged to non-symptomatic leaflets inoculated with Pst, and 743 (21.4%) to leaflets inoculated with Xeu bacteria. Lastly, after symptom development, 1257 (36.1%) spectra were captured (866 – 24.90% – from symptomatic Pst leaflets, and 391 – 11.24% – from Xeu symptomatic tissue). Class imbalance is observed due to the disease infection process’s dynamic, resulting in symptoms appearing throughout the measurements dates at different rates ( Table 1 ). Spectral assessments were performed during 18 DAI for Control and Xeu leaflets. For Pst, the process was only made until 15 DAI because the plants presented high-stress levels, and leaf dehydration after this date, interfering with the spectral signal recording ( Figure 1 ; Table 1 ).

Hyperspectral signatures captured in healthy leaflets showed the typical spectral behavior of healthy green tissues. On the other hand, spectral assessments belonging to disease leaflets (both with Pst and Xeu bacteria) presented deviations in their format ( Figure 5 ). Thus, a more detailed analysis was performed for these measurements to evaluate the spectral modifications caused by the different bacteria, resulting in a higher number of classes in the study. Spectra signatures belonging to Pst inoculated samples had a more distinct spectral curve (for both, non-symptomatic and symptomatic stages) compared to the healthy measurements, showing higher intensity on the wavelength ranges of approximately 430 to 520 nm, and 560 to 680 nm. Nevertheless, the lower intensity was captured from 710 to 800 nm ( Figures 6A, B ). Xeu-inoculated tissues also displayed modification in their spectral signature in these regions. The intensity measured in the first two spectral intervals was marginally higher than the one captured on healthy leaflets. However, a more evident variance was observed in the 710 to 780 nm range ( Figures 6A, C ). When measurements belonging to disease samples were compared, the data showed differences between the samples infected with the different etiological agents. Pst measurements (for both non- and symptomatic stages) demonstrated greater intensity in the ranges of 430 to 520 nm, and 560 to 680 nm, but lower intensity in the 710 to 800 nm interval ( Figures 6A, D ).

A Linear Discriminant Analysis (LDA) was performed to reduce the dimensionality of the normalized dataset, organizing the spectral observations in a new space as the maximum linear separability possible. LDA results were plotted and showed spectral separability between the different classes studied ( Figure 7A ). It was possible to see an evolution pattern through LD 1 for spectral data belonging to healthy, and Pst diseased leaflets regardless of whether they exhibit symptoms or not ( Figures 7A, B ). In turn, healthy and Xeu-diseased leaflets (including, non- and symptomatic data) presented a spectral separation gradient through LD2 ( Figures 7A, C ). When data of diseased leaflets infected with distinct bacteria were compared, it was possible to observe a divergence gradient between the LD1 and LD2, especially at the symptomatic stage ( Figures 7A, D ). Since data presented a non-linear characteristic, overlapping between classes was observed. Thus, these findings demonstrated the efficacy of the LDA technique for reducing the dataset dimensionality to the most important features. LDA’s DR results were, then, applied in the following steps of the modeling process helping in the classification task and avoiding overfitting.

The most relevant wavelength variables for LD1 were assessed based on their coefficients, equaling 44 features. These variables were mostly located in the blue-green and red VIS regions of the electromagnetic spectrum (blue - 434.9, 435.72, 438.17, 438.58, 440.21, 441.44, 442.67, 443.08, 445.53, 445.94, 448.4, 448.81, 494.6 nm; green - 503.74, 508.74, 527.53 nm; red - 556.09, 562.0, 562.84, 590.37, 607.82, 609.1, 611.24, 618.5, 643.36, 650.24, 673.97, 680.02 nm), coinciding with the wavelength absorption range of chlorophylls (430 to 480 nm, and 640 to 700 nm), and carotenoids pigments, namely β-carotenes (whose primary and secondary absorption peaks are respectively located at 450 to 480 nm, and 600 to 650), and xanthophylls (520 to 580 nm) ( Figure 8 ). This coincides with the action of Pst and Xeu bacteria on tomato leaves’ levels of photosynthetic pigments during the infection process.

Other plants whose metabolites are affected by these two bacteria also have their absorption spectrum coinciding with the selected wavelengths of LD1, namely some phenolic compounds (e.g., flavonoids, 400 to 500 nm), and composts derived from chlorophylls decomposition, namely pheophytins (400 to 500 nm, and 600 to 700 nm) ( Figure 8 ).

Applied predictive classification modeling was, then, performed using the LDA-reduced normalized data (including all classes: i) healthy; ii) non-symptomatic diseased Pst leaflets; iii) non-symptomatic Xeu samples; iv) symptomatic inoculated Pst tissues; v) symptomatic Xeu observations) and an SVM algorithm with a Radial Basis Function (RBF) kernel. The model was trained using 70% (2434) of the spectral observations (randomly selected), and then, it was validated using the remaining 30% (1044) of the observations (test set), and the complete dataset. The test set comprised 413 healthy samples, 30 non-symptomatic Pst disease leaflets, 223 non-symptomatic Xeu, 260 symptomatic Pst observations, and 118 symptomatic Xeu.

The developed model performed well for both the test set and the complete dataset. The model achieved an accuracy of 0.85 for the test set and 0.86 for the complete dataset, indicating that it can correctly classify most of the measurements ( Table 3 ; Figure 9 ). Furthermore, it demonstrated high metric values (precision, recall, and F1-score) for all the classes, indicating that it can identify both healthy and infected measurements. In detail, higher precision, recall, and F1-score values were found for the healthy and non-symptomatic Pst leaflets measurements ( Table 3 ). This shows that the model more easily predicted spectral assessments belonging to these classes. Nevertheless, it showed more difficulties in classifying measurements of Xeu inoculated leaflets, especially those captured before symptom appearance (indicated by lower metric scores). It is important to note that the model’s performance was consistent across both the test set and the complete dataset, indicating that the model is robust and can be used to classify new spectral samples accurately.

Model predictions for the non-symptomatic Pst class did not present any misclassification in the test set. In the complete dataset, the model accurately predicted 96% of the spectral measurements but missed 1% of the predictions, which it classified as assessments made on non-symptomatic leaflets infected by Xeu ( Figure 9 ). Symptomatic spectral captures performed in Pst diseased leaflets were correctly categorized in 94% of the cases for both the test and complete sets. Nevertheless, the model mistakenly classified these assessments as non-symptomatic Xeu observations in 4% and 3% of the cases, and as healthy samples in 2% when the test set and complete dataset were used, respectively. Predictions of Xeu spectral assessments were more challenging to the model, presenting a higher number of wrong classifications in the non-symptomatic class than in the remaining classes studied. In fact, the model successfully classified 77% of the measurements of this class in the test set, and 78% when all data was used. However, it attributed 11% and 10% of the measurements as healthy, 5% and 8% as symptomatic diseased Xeu leaflets assessments, 3% as non-symptomatic inoculated Pst observations, and 2% as symptomatic Pst captures, when the test set and complete dataset were used, respectively. The model showed more efficacy in identifying symptomatic Xeu leaflets measurements, predicting 83% of these samples in the test and complete datasets. In terms of missed classifications, it predicted 6% and 5% of the assessments as non-symptomatic, 3% and 2% as healthy, 3% and 1% as non-symptomatic spectral captures of Pst infected leaflets, and 1% and 2% as symptomatic Pst, in the test set and complete dataset, respectively ( Figures 9A, B ).

For the complete dataset prediction, we investigated the number of misclassifications per class and date ( Figure 10 ). As expected, the observed tendency for healthy spectral assessments showed a regular number of observations per date (81). Nonetheless, the developed model categorized more samples than the true value per date, except for 7, 13, 17, and 18 DAI. On the other hand, the spectral model consistently underfit the infected Xeu leaflets, regardless of whether they exhibit symptoms or not ( Figure 10A ).

In plants inoculated with Xeu, discrepancies between observed and predicted classes are more evident in the non-symptomatic Xeu class in the observations recorded up to 10 days after infection. During this period, which included seven measurement dates of the non-symptomatic Xeu class, 53 observations were recorded below the predicted value of the developed model. In contrast, the healthy class accumulated 47 observations above the predicted value during the same period. Furthermore, according to the confusion matrix results (All data), 10% (148) of the non-symptomatic Xeu observations were misclassified as healthy. Considering the period up to 10 days after infection (data not shown), out of the 150 observations wrongly classified as healthy, 100 were from the non-symptomatic Xeu class. These results indicate that in the early stages of Xeu-induced disease infection, the symptoms developed in the plant leaflets are not strong enough for the developed model to distinguish them from healthy observations efficiently. Therefore, the non-symptomatic Xeu class, compared to other tested classes, exhibits the lowest model performance indicators (all data: accuracy 0.74, precision 0.78, recall 0.74, and F1-score 0.76). For the non-symptomatic stage, the actual observations presented a stable pattern until 8 DAI, and after a sharp drop was observed until 13 DAI, where the rate of infected leaflets increased up to 65%. A stable number of observations was maintained until 15 DAI, after which a period of exponential increase in observed symptomatic spectral measurements was registered. After this day, all leaflets were symptomatic. The model was rigorous in discriminating non-symptomatic Xeu leaflet measurements after 9/10 DAI, presenting a percentage of error inferior to 10% (correctly classifying 64 of the 71 observations) when about 90% of the sampled leaflets (71 of the initial 81 assessments) still didn’t show any typical symptoms of the disease ( Figure 10B ).

For the prediction of the Pst disease samples, the non-symptomatic phase was very similar for both observed and predicted. Nevertheless, the prediction of the symptomatic phase showed irregularities between the five and seven days (corresponding to the dates were necrosis appeared). Is possible to observe that most of the Pst inoculated leaflets (79%) started to show the first symptoms of the disease 4 DAI. The number of symptomatic sampled leaflets increased until 6 DAI, where all the leaflets assessed were symptomatic ( Figure 10C ).