1) Dataset Description

The dataset used in this study consists of cocoa leaf images from plants belonging to the Forastero group, specifically from the Amelonado sub-variety and Trinitario group. The images were collected from three plantations in Côte d’Ivoire. The experimental plantation of the National Centre for Agronomic Research (CNRA) located in Bouaflé, in the central-western region of Côte d’Ivoire (Forastero) and a local farmer’s plantation in the same area (Forastero group). The last is a local farmer’s plantation in Divo (Forastero and Trinitario group), located in the southern part of Côte d’Ivoire, a region known for high cocoa production.

The dataset comprises 6000 images (2456 px × 3275 px), organized as follows: A first folder containing 2700 images of healthy plant leaves and the second folder containing 3300 images of infected plant leaves showing visible symptoms.

The images were captured in uncontrolled, real-world conditions during normal daylight hours, under clear skies (between 10 a.m. and 4 p.m.), with ambient temperatures ranging from 28˚C to 32˚C, solar irradiance between 200 W/m² and 500 W/m², and humidity levels between 70% and 90%. The distance between the camera and the subject ranged from one to five meters.

Figure 1 and Figure 2 display images of healthy cocoa leaves and leaves showing symptoms respectively.

2) Data Pre-processing

Data pre-processing is a critical phase in any image classification process, as it significantly influences both the performance and efficiency of the models.

For our data pre-processing, we utilized the CLAHE (Contrast Limited Adaptive Histogram Equalization) algorithm. CLAHE is an image processing technique designed to enhance local contrast while mitigating the noise typically amplified by global histogram equalization [30] [31] . The CLAHE pre-processing method consists of the following steps:

Upon processing all blocks, the resulting image displays enhanced local contrast without the artifacts commonly associated with global histogram equalization.

Figure 3 below are examples of images following the application of the pre-processing algorithm.

3) Data Augmentation

To improve the model’s generalization by exposing it to various perspectives of the original images, we applied data augmentation techniques. This was achieved by rotating the images at three angles: 90˚, 180˚, and 270˚, followed by flipping each of these rotated images ( Figure 4 ). This approach effectively increases the size of our dataset by a factor of eight. The advantage of this technique lies in its ability to diversify the dataset while enhancing the model’s resilience to noise. As a result, it improves the model’s accuracy and sensitivity by optimizing feature detection and reducing biases.

Following the pre-processing steps, we present in Table 1 the proportions of the different categories of leaves in the dataset.

4) Hardware

We selected Python as the programming language for implementing our project due to its simplicity, versatility, and the extensive range of libraries it offers. Python is also supported by most online platforms. For hardware, we used an HP All-in-One desktop PC equipped with an Intel(R) CORE I7-12700T 2.4 GHz processor, 16 GB RAM, a 1 TB SSD, and an NVIDIA Quadro P400 GPU. Additionally, we utilized the free Kaggle online platform for testing, which provides 16 GB of disk space, 16 GB RAM for the CPU, and 16 GB RAM for the GPU.

We implemented three other algorithms in parallel using the pre-trained models VGG-19, Inception V3, and DenseNet-201.

1) CNN Algorithm used

a) VGG-19

The VGG-19 model is a deep convolutional neural network belonging to the VGG (Visual Geometry Group) family. Developed by the Visual Geometry Group at the University of Oxford, it consists of 19 convolutional layers using a 3 × 3 kernel. The model employs a nonlinear activation function and max-pooling for subsampling. Additionally, it includes two fully connected layers, each with 4096 nodes, followed by a Softmax layer for classification [32] [33] .

b) Inception V3

Inception V3 is a widely adopted convolutional neural network architecture designed for classification tasks. The Inception V3 module forms the backbone of the GoogLeNet network. Typically, the Inception V3 module contains three different convolution sizes and a max-pooling operation. After performing convolutions on the output from the previous layer, channels are aggregated, and pooling of non-network elements occurs. Inception V3 introduces kernel factorization to break down larger convolutions into smaller ones, optimizing computation [34] .

c) DenseNet-201

The Dense Convolutional Network (DenseNet) architecture includes four variants: DenseNet121, DenseNet169, DenseNet201, and DenseNet264. These networks use layers with 12 filters. In our study, we used DenseNet-201, where each layer has direct access to both the original input image and the gradients of the loss function. This architecture significantly reduces computational cost, making DenseNet-201 an excellent choice for image classification tasks [35] .

2) Classifiers

a) Sigmoid Classifier

The sigmoid classifier employs the sigmoid function as an activation function within a model, commonly used in logistic regression or neural networks. The sigmoid function is mathematically defined by the following formula:

$\xi =\frac{1}{1+{\text{e}}^{-x}}$ (1)

This function takes an input value xxx and transforms it into an output ranging between 0 and 1, making it particularly suitable for binary classification tasks where the objective is to predict the probability of belonging to a specific class. In neural networks, the sigmoid function is often applied for several reasons:

However, the sigmoid function has certain limitations. It can lead to slow learning due to the very small gradients produced for large or small input values. This inefficiency, particularly as the function approaches 1 or 0, can hinder the learning process.

b) XGBOOST Classifier

XGBoost (Extreme Gradient Boosting) is an optimized machine learning library designed for implementing gradient boosting algorithms. Introduced by Tianqi Chen in 2014, it has become one of the most popular and efficient classifiers for both classification and regression tasks in machine learning [36] . XGBoost is recognized for the following:

c) SVM Classifier

The Support Vector Machine (SVM) is a supervised learning algorithm used for both classification and regression tasks. It seeks to maximize the margin between classes by finding an optimal hyperplane that separates them. The support vectors are the points closest to this hyperplane.

SVM uses kernel functions (linear, polynomial, RBF) to handle non-linear data by projecting it into a higher-dimensional space where separation becomes possible. The C parameter controls the trade-off between maximizing the margin and minimizing classification errors.

SVM is valued for its robustness and ability to generalize well, especially in high-dimensional spaces. However, it can be computationally expensive and often requires fine-tuning of parameters.

3) Evaluation Metrics

To evaluate the performance of the models in our study, we used several evaluation metrics, including accuracy, mean squared error (MSE), recall, F1 score, and the Matthews correlation coefficient (MCC), each calculated using the following formulas:

a) Accuracy

Accuracy is the simplest and most commonly used metric to evaluate the performance of deep learning algorithms. It represents the percentage of correct predictions out of the total predictions made on the selected sample.

$\text{Accuracy}=\frac{\text{PC}}{\text{PT}}$ (2)

b) Mean Squared Error (MSE)

Mean Squared Error measures the average of the squared differences between the predicted values and the actual values of the training data.

$\text{MSE}=\frac{1}{n}{\displaystyle {\sum }_{i=1}^n{\left(y_i-{\hat{y}}_i\right)}^2}$ (3)

c) Recall

Recall assesses the true positive rate, which is the proportion of leaves showing symptoms that were correctly identified. It measures the model’s ability to detect all infected leaves.

$\text{Recall}=\frac{\text{VP}}{\text{VP}+\text{FN}}$ (4)

d) F1 Score

The F1 score is the weighted average of recall and precision, balancing both metrics to provide a single performance measure.

$\text{F}1\text{score}=2\ast \frac{\text{Precision}\ast \text{Recall}}{\text{Precision + Recalll}}$ (5)

e) Matthews Correlation Coefficient (MCC)

MCC is used to evaluate the quality of binary classifications, offering a more balanced assessment than accuracy, especially in cases with imbalanced datasets.

$\text{MCC}=\frac{\text{VP}\ast \text{VN}-\text{FP}\ast \text{FN}}{\sqrt{\left(\text{VP}+\text{FP}\right)\left(\text{VP}+\text{FN}\right)\left(\text{VN}+\text{FP}\right)\left(\text{VN}+\text{FN}\right)}}$ (6)

where:

These metrics were used to assess the overall performance and robustness of the implemented algorithms.

4) Implementation of the Algorithms

a) Architecture of the Algorithms

Our algorithm follows the general architecture of convolutional neural networks. It consists of a convolutional layer, a subsampling layer, a normalization layer, and a fully connected layer (e.g. Figure 5 ).

b) Hyper parameters used

We implemented this algorithm using the Python programming language with TensorFlow libraries, notably Keras. In the implementation, we used the following hyper parameters:

1) Scenario 1

Scenario 1 involves detecting cocoa leaf diseases using feature extractors from VGG19, DenseNet, Inception V3, and Custom CNN with fully connected RNN layers as classifiers.

2) Scenario 2

Scenario 2 involves detecting cocoa leaf diseases using feature extractors from VGG19, DenseNet, Inception V3, and Custom CNN with XGBOOST classifiers.

3) Scenario 3

Scenario 3 involves detecting cocoa leaf diseases using feature extractors from VGG19, DenseNet, Inception V3, and Custom CNN with SVM classifiers.

This section presents all the experimental results obtained from our work, as well as a comparative study of the results from the VGG19, Dense Net, Inception V3, and Custom CNN algorithms with a sigmoid, XGBOOST and SVM classifier. For the three scenarios, we used the same dataset described earlier.

To highlight the differences and similarities between the three scenarios, we will use the standard metrics described above.

1) Results of scenario 1

Table 2 and Figure 6 below summarie the results of the tests conducted with the previously mentioned custom CNN, using the evaluation metrics discussed earlier for Scenario 1.

Table 2 presents the metric values collected during the validation and testing phases of the deep learning models used. We obtained accuracy values ranging from 92.61% to 96.30%, and recall values from 98.99% to 100%, demonstrating the overall predictive performance of the models and their ability to correctly classify both positive and negative cases. Precision values ranged from 84.67% to 91.44%, and F1-scores varied from 91.38% to 92.66%, confirming the balanced capability of the models to identify positive instances while maintaining a low false-positive rate.

In this scenario, the performance of the models, particularly VGG19, DenseNet, Inception V3, and a custom CNN with a sigmoid classifier, yielded very encouraging results. The best performance was achieved with our custom CNN.

2) Results of Scenario 2

Table 3 and Figure 7 below summarize the results of the tests conducted with the previously described Custom CNN, using the specified metrics for Scenario 2.

The four models were evaluated using the XGBOOST classifier, which significantly improved their performance as observed in Table 3 . Accuracy values ranged from 95.94% to 99.10%, recall from 99.83% to 98.98%, precision from 90.78% to 97.77%, and F1-scores varied from 95.09% to 98.87%.

In this scenario, the performance of the four models significantly improved with the XGBOOST classifier. The best results were achieved with our custom CNN.

The graphs in Figure 7 , below provide an overview of the performance (accuracy and loss) of each mode.

3) Results of Scenario 3

Table 4 and Figure 8 below summarize the results of the tests conducted with the previously described Custom CNN, using the specified metrics for Scenario 2.

In Scenario 3, the previous models were evaluated using the SVM classifier. We obtained accuracy values ranging from 94.31% to 98.31%, recall from 99.12% to 99.64%, while precision values ranged from 90.92% to 97.85%, and F1-scores varied from 95.08% to 98.58%. We observed a significant improvement in the results for all four models with the SVM classifier, although the performance remained below that of the XGBOOST classifier in Scenario 2. All four models showed better performance with both the SVM and XGBOOST classifiers.

The graphs in Figure 8 provide an overview of the performance (accuracy and loss) of each mode.

4) Results Analysis

The increase in accuracy based on the test data is presented in Table 5 below.

Table 5 above presents the variations in efficiency observed across the three scenarios during the validation and testing phases of the different models. We observed an Accuracy increase ranging from 2.54% to 3.47%. These values underscore the overall improvement in results in the second scenario.

All four models showed enhanced performance with both the SVM and XGBOOST classifiers.

However, the custom CNN stood out with an accuracy of 97.77% and a recall of 99.98%.

A comparative analysis of the three scenarios highlights the efficiency and precision gains brought by the use of XGBOOST in Scenario 2. These results illustrate the positive impact of integrating XGBOOST on model performance. The improved performance indicates a better ability of the models to generalize on validation and test datasets, reducing classification errors and increasing overall robustness.

Our study shows that the XGBOOST classifier is highly effective in classifying Swollen Shoot symptoms from cocoa leaf images.