In this study, we used healthy, asymptomatic, and symptomatic infected cocoa leaves as biological samples. The Theobroma cacao variety we studied is named Amelonado. These leaves were collected with the support of the Central Biotechnology Laboratory of the Centre National de Recherche Agronomique (CNRA) in two fields located in Côte d’Ivoire. The healthy samples were collected on a CNRA experimental field located at Adiopodoumé Km 17, while the collection of leaves infected with the swollen shoot virus was carried out in a plantation located at Garango in Bouaflé department where swollen shoot disease is widespread [27]. These experimental cocoa trees plantations were established at the same period and are subject to the same cultivation practices. However, rainfall in Adiopodoume is slightly higher than in Bouaflé. Each type of leaf was collected from 20 trees with 3 leaves per tree, for a total of 60 leaves per type. So the biological samples consisted of 60 healthy leaves, 60 asymptomatic infected leaves, and 60 symptomatic infected leaves. Figure 1 shows the three leaf types we used.

The hyperspectral reflectance signatures of cocoa leaves were acquired using an Ocean Optics USB 4000 spectrometer, which has a wavelength range from 350 to 1100 nm and a spectral resolution of 0.22 nm. The setup includes a bifurcated optical fiber (QR400-7-VIS-BX), a 20 W halogen source (HL-2000-HP-FHSA), a RPH probe tip, and a laptop computer for data storage and processing. Specialized data acquisition software (SpectraSuite) was used, and the spectrometer was calibrated with a diffuse reflectance reference surface (WS-1). Further details of the experimental setup can be found in previous work [21][28]. Figure 2 illustrates the experimental setup used.

Figure 1. Types of leaves submitted for study. (a) healthy leaves; (b) asymptomatic leaves; (c) symptomatic leaves.

Figure 2. Experimental device for acquiring hyperspectral data.

Measurements were performed on 180 cocoa leaves. For each leaf sample, three mean reflectance spectra were measured on the upper, middle, and lower parts of the leaf’s adaxial surface. A total of 540 spectra were then collected, with 180 spectra per leaf type.To reduce the influence of noise inherent to the hyperspectral reflectance device at the limits of its spectral range, only spectral reflectance data between 400 and 1000 nm were used for processing and analysis. These data were then preprocessed using three methods: the Multiplicative Scattering Correction (MSC), the Standard Normal Variable (SNV), and the first order Savitzky-Golay derivative (SG-D1).

The MSC method effectively eliminates scattering effects while preserving useful spectral information, so that the spectra are as close as possible to a reference spectrum. Generally, the average spectrum of the entire dataset is considered as the reference spectrum [29][30]. SNV performs standard normalization on each spectral data point to reduce light scattering effects [31]. SG-D1 is highly sensitive to noise. It is an effective technique for suppressing background effects and enhancing subtle, weak spectral features useful for evaluating target parameters [32][33].

Principal component analysis (PCA) is a widely used technique for reducing the dimensionality of high-dimensional data. It transforms a set of potentially correlated variables into a new set of uncorrelated variables, called principal components, through orthogonal transformation [34]. PCA of raw and preprocessed spectra can provide crucial information on data separation [35]. We generated the principal components of the spectral data and visualized the distribution of samples of healthy, asymptomatic, and symptomatic leaves in three dimensions. This visualization will identify the different clusters, thus providing a better understanding of the underlying structure of the raw and preprocessed spectral data.

Vegetation indices (VIs), obtained by mathematical transformation from spectral bands, generally highlight the spectral characteristics of leaves [36]. Numerous studies have demonstrated the effectiveness of VIs for monitoring crop diseases [21][37]-[39]. Taking into account the spectral range of our spectrometer, we selected eight vegetation indices from the literature related to leaf pigments or structure and crop diseases. These indices were used as variables in the development of classification models for the three types of cocoa leaves. Using spectral data, a program developed in MATLAB allowed us to calculate these eight vegetation indices. Table 1 presents the definitions, calculation formulas, and bibliographic references for each VI.

The statistical analysis of significant differences applied to the independent variables consisted of verifying the differences between healthy and infected (asymptomatic and symptomatic) cocoa leaves based on all the eight VIs. Normality was tested using the Shapiro-Wilk test [48] and equality of variances with the Leneve test [49]. Since non-normality and non-homogeneity of variances were observed in the data for asymptomatic and symptomatic leaves, a statistical difference analysis was performed using the non-parametric Kruskal-Wallis test [50]. This test does not require assumptions about normality and homoscedasticity, making it a robust option for comparing multiple independent samples. Finally, Dunn’s multiple comparisons test [51] was applied with a significance level of 5% using MATLAB software.

Table 1. Vegetation indices used.

To compare the performance of the classifiers based on the all eight vegetation indices evaluated in this study, four machine learning algorithms frequently used for classification were employed, namely: K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Decision Tree (DT), and Random Forest (RF). These methods were chosen because of their high classification accuracy in similar studies [52]-[55].

The KNN algorithm is a supervised machine learning algorithm widely used for classification and predictive regression problems. It is capable to describe nonlinear relationships between collected samples and hyperspectral data. Specifically, for given test samples, a specific distance evaluation method is used to determine the k closest samples [56]. Then, a predictive classification is performed on these k samples. In general, KNN is easy to use and works well with little prior knowledge of the data distribution. Euclidean distance is the most common method for calculating the variations between samples, represented by input vectors [57].

SVM is a supervised algorithm used in machine learning. It performs classification by finding the hyperplane that maximizes the margin between data. The vectors that define the hyperplane are called support vectors. SVM is based on statistical approaches, enabling data classification and assigning each class a specific score that serves as the basis for evaluation. SVM can also be used for regression tasks even if it is more commonly used for classification purposes [58]. Intuitively, good separation is achieved by the hyperplane that has the greatest distance from the nearest training data points of any class. Indeed, in general, the larger the margin is, the smaller the classifier’s generalization error is [59].

A decision tree (DT) is a widely used supervised machine learning technique for classification and regression tasks. The classifier breaks down a dataset into progressively smaller subsets based on their attribute values [60]. The decision tree algorithm has a tree structure where internal nodes represent features, while terminal nodes contain class labels or predicted values. A prediction is made by traversing the path from the root to one of the terminal nodes based on the input feature values. This process divides the data until no further division is possible or when all values of the target variable are identical. Decision trees are easy to understand and interpret. They can handle numerical and categorical variables and are robust to outliers and missing values. Most decision trees consist of a random forest-type classifier that determines the category based on the classes in a particular tree.

Random Forest (RF) is a classification model composed of multiple decision trees, combined to obtain a more accurate and stable prediction. Each tree depends on the values of a randomly sampled vector that follows the same distribution for all trees in the forest. The final classification result is obtained by averaging the classification results of all individual binary decision trees [61]. The RF method has many advantages over other machine learning methods; the most notable one is its high accuracy when a large dataset is used for training [62]. Furthermore, it is simple and quick to implement.

In order to better assess the performance of the classification models used and their stability for the swollen shoot infection detection, 540 observations of the eight vegetation indices were divided into two datasets: a training set (70%) and a test set (30%). The split of samples was performed at the leaf level, so all spectra acquired on the same leaf belong to the same dataset. A confusion matrix was provided for each model, comparing the predicted classes of the test set to the real classes in order to evaluate the performance of classification models based on True Positives (TP), False Positives (FP), True Negatives (TN) and False Negatives (FN) [63]. Accuracy, precision, recall and F1-Score were computed as parameters for evaluating model performance. The formulas for these performance indicators are respectively given by Equation (1), Equation (2), Equation (3) and Equation (4) as follows:

This metric reveals the model’s predictive performance, reflecting its overall ability to correctly classify healthy, asymptomatic, and symptomatic leaves.

Precision represents the proportion of real positive cases among all samples predicted to be positive. The higher the Precision is, the better the model performs at distinguishing between healthy, asymptomatic, and symptomatic leaves. In other words, high accuracy means fewer incorrect positive predictions.

This indicator shows the proportion of samples correctly predicted to be positive and which actually are. Thus, Recall can be used to assess the completeness of the prediction of healthy, asymptomatic, and symptomatic leaves.

This is the harmonic mean between Precision and Recall, taking into account the overall balance between the two parameters. The F1-Score represents a balanced average between positive predictive value and sensitivity. An optimal value for this metric indicates better algorithm efficiency.

Spectral preprocessing is an essential step in spectral analysis. This approach yields reliable and more accurate spectral data, facilitating subsequent analyses and applications. Preprocessing modifies spectra in various ways. Figure 3 provides an overview of the raw and preprocessed reflectance spectra of healthy, asymptomatic, and symptomatic cocoa leaf samples.

Figure 3. Raw reflectance spectra (a) and preprocessed: MSC (b), SNV (c) and SG-D1 (d).

Observation of the raw spectra in Figure 3(a) reveals that the spectral signatures of our samples are similar, with low reflectance in the visible region, a sharp increase in the far-infrared and higher reflectance in the near-infrared. However, these raw reflectance spectra of healthy, asymptomatic, and symptomatic cocoa leaves exhibit strong dispersion in the 400 - 1000 nm wavelength range. The MSC method corrects the scattering effect in spectral data to bring these data closer to the real spectral information. The spectral signatures in Figure 3(b) are more consistent and the spectral differences due to scattering are corrected. SNV homogenizes the spectra by standardizing each spectrum, eliminating multiplicative noise and reflectance variations. The spectra in Figure 3(c) are normalized to the same scale to facilitate comparison and subsequent analysis. The SG-D1 highlights changes in the position of emission or absorption peaks in the spectral reflectance signature. Figure 3(d) provides more information on the positions of these peaks but this may weaken the relative variation in their intensity.

The spectral curves obtained by pretreatment, while each exhibiting their own characteristics, cannot be intuitively compared in terms of their effects on the reflectance data of our samples. Therefore, in this study, principal component analysis (PCA) was used to analyze the results of the different pretreatment methods in order to identify any clustering trends in the healthy and infected cocoa leaf samples. The PCA results obtained from the raw and pretreated spectra showed that the first three principal components provided useful clustering for healthy, asymptomatic, and symptomatic cocoa leaves. Among the three pretreatment methods, MSC proved to be the best method compared to the others, as shown in Figure 4. Indeed, it yielded clearer clustering trends in the space of the first three principal components.

The best grouping by leaf type (healthy, asymptomatic, and symptomatic) was obtained by combining MSC pretreatment and PCA. This could be attributed to the ability of the MSC technique to correct the scattering effects and eliminate unwanted scattering from the spectral data matrix. Thus, the MSC pretreatment technique is best suited to our raw reflectance data for better classification of the three types of cocoa plants.

The Kruskal-Wallis nonparametric test and Dunn’s multiple comparisons test were performed to evaluate the differences between healthy and infected cocoa leaves using various vegetation indices. The results are summarized in Table 2. They indicate that all vegetation indices of infected plants are significantly influenced by viral infection. Indeed, the calculated vegetation indices show significant differences between healthy and asymptomatic leaves, and also between healthy and symptomatic leaves. Regarding the comparison between the vegetation indices

Figure 4. Representation of the first three principal components of raw reflectance spectra (a), pretreated by: MSC (b), SNV (c), SG-D1 (d) of healthy (H), asymptomatic (As) and symptomatic (Sy) cocoa leaves.

Table 2. Significant differences in vegetation indices of healthy (H), asymptomatic (As) and symptomatic (Sy) cocoa leaves, obtained by the Kruskal-Wallis test for a statistical significance threshold p < 0.05.

(*) significant difference, (ns) no significant difference

of asymptomatic and symptomatic leaves, only the GNDVI, Ctr2, PSRI, PRI, and Lic2 indices show significant differences. These five vegetation indices are significantly different for the three cocoa leaves types, so they well discriminated them. This is explained by the fact that these vegetation indices are closely linked to changes in leaf pigments. This interdependence influences their values [64]-[66]. These changes are manifested by the symptoms of chlorosis observed on the symptomatic leaves of cocoa plants infected by the swollen shoot virus.

The distribution of each vegetation index for the three cocoa leaves states is illustrated by the box plots in Figure 5.

Figure 5. Box plot of vegetation indices of healthy (H), asymptomatic (As) and symptomatic (Sy) cocoa leaves.

Compared to healthy leaves, the median values of the vegetation indices NDVI, GNDVI, Datt5, PSRI, and SIPI increase with infection, while those of PRI and Lic2 decrease. There is virtually no overlap between the box plots of healthy and symptomatic leaves vegetation indices. In contrast, there is generally some overlap between the boxes for asymptomatic and symptomatic cocoa leaves. Similarly, for GNDVI, Dattt5 and Lic2, there is slight overlap between the boxes for healthy and asymptomatic cocoa leaves. This demonstrates the difficulty to detect early infection by cocoa swollen shoot virus using threshold values based on individual vegetation indices.

The overall performance of the four evaluated algorithms (KNN, SVM, DT, RF) is excellent, as shown in Table 3. RF stands out as the most efficient algorithm, with an accuracy of 97.59%, a Precision of 97.62%, a Recall of 97.59%, and an F1-Score of 97.59%. These results demonstrate extremely reliable classification and excellent consistency between accuracy and Recall, thus highlighting the robustness of the model. The KNN algorithm also performed well, with an accuracy of 95.93%, a Precision of 96.25%, a Recall of 95.93%, and an F1-Score of 95.92%, making it a competitive alternative. The performance of the SVM algorithm is also remarkable, with an accuracy of 93.89%, a Precision of 94.1%, a Recall of 93.89%, and an F1-Score of 93.91%, but significantly less efficient than the RF and KNN models. In contrast, DT achieved more modest results, with 90.74%, 90.85%, 90.74%, and 90.74% respectively for accuracy, Precision, Recall, and the F1-Score. This indicates a strong classification capacity, but significantly lower stability and reliability than the other three models. These results confirm the relevance of MSC preprocessing for improving the quality of our spectral data, the effectiveness of the computed vegetation indices and reveal the robustness of the RF. Indeed, in this study, RF proved to be the most accurate and robust model for classifying healthy, asymptomatic and symptomatic cocoa leaves. Therefore, we can conclude that cocoa swollen shoot disease can effectively be diagnosed at an early stage using a combination of sensitive vegetation indices as input variables in a classification model.

Table 3. Algorithms performance evaluation.

To further illustrate the predictive performance of the different classification models, for healthy, asymptomatic, and symptomatic cocoa leaves, we present in Figure 6 the confusion matrices of the KNN, SVM, DT and RF classifications.

As illustrated in Figure 6, for a total of 180 samples, the RF model correctly predicted 176 samples, while the KNN, SVM, and DT models predicted 172, 170, and 163 samples, respectively. Despite some differences in predictive accuracy among these models when classifying the three cocoa leaves categories, all four models performed satisfactorily.

The performance obtained with the RF model is similar to that of Kouassi et al. (2024) [17], who used convolutional neural networks combined with different classifiers. The combinations with the XGBOOST classifier to differentiate between healthy and infected cocoa leaves yielded the best performance, with overall accuracies exceeding 95%.

In contrast, Batoo et al. (2025) [18] used a three-layer convolutional neural network to classify infected and healthy cocoa tree samples. This classification reached an accuracy of 88%, lower than that of our models.

Figure 6. Confusion matrices: (a) KNN; (b) SVM; (c) DT and (d) RF.