Fig 3 presents the training and validation accuracy and loss curves obtained over 50 epochs. The training accuracy shows a rapid increase during the early epochs, indicating that the model quickly learned meaningful visual features from the haricot bean leaf images. As training progressed, the accuracy continued to improve gradually and stabilized at 96.75% by the final epoch. Similarly, the validation accuracy followed an upward trend and reached 88.41%, demonstrating that the model maintained good generalization performance on unseen validation data. The loss curves further support this observation. Training loss decreased steadily throughout the training process, reflecting effective optimization of the categorical cross-entropy objective function. Validation loss also showed a general downward trend, although minor fluctuations were observed across epochs. These variations are expected in stochastic gradient-based learning due to mini-batch updates and real-time data augmentation. Importantly, there was no significant divergence between training and validation curves, suggesting that overfitting was effectively controlled. The incorporation of dropout layers, batch normalization and augmentation techniques contributed significantly to maintaining stable convergence.
Although a moderate gap exists between training and validation accuracy, the difference remains within an acceptable range for deep learning models trained on real-world agricultural datasets. This indicates that the model learned robust and discriminative features such as lesion patterns, rust formations and discoloration characteristics associated with different disease classes. Evaluation on the independent test set further confirms the effectiveness of the proposed model. The network achieved a test accuracy of 99.13% with a test loss of 1.7738, demonstrating strong predictive capability on unseen haricot bean leaf images. The high test performance indicates excellent generalization ability and suggests that the trained AlexNet-based architecture successfully captured disease-specific visual patterns. Overall, the results validate the suitability of the proposed approach for accurate and reliable automated bean leaf disease classification in practical agricultural applications.
Fig 4 presents the confusion matrix obtained from the test dataset for the four-class haricot bean leaf disease classification task. The matrix provides a detailed breakdown of correct and incorrect predictions made by the model across the classes. The results indicate that the model achieved near-perfect classification performance for most categories. Specifically, 173 samples of angular leaf spot, 172 samples of bacterial blight and 173 samples of bean rust were correctly classified, with no misclassifications observed for these three classes. This demonstrates that the model effectively captured distinctive disease-related visual characteristics such as lesion boundaries, discoloration patterns and rust pustules. For the healthy leaf category, 167 samples were correctly classified, while 6 samples were misclassified as bean rust. This minor misclassification may be attributed to the presence of very early or mild rust symptoms, which are visually indistinguishable from healthy leaf surfaces, as well as variations in illumination, background noise and image acquisition conditions. Nevertheless, the overall classification performance remains exceptionally high. The dominance of diagonal elements and the minimal off-diagonal values confirm excellent class separability and strong generalization capability.
Table 2 presents the detailed classification report of the proposed AlexNet-based model evaluated on the test dataset. The results show outstanding performance across all four categories. Both angular leaf spot and bacterial blight achieved perfect precision, recall and F1-score values of 1.0000, indicating that all samples in these classes were correctly identified without any false positives or false negatives. For the bean rust class, the model achieved a precision of 0.9665, recall of 1.0000 and an F1-score of 0.9830. The perfect recall indicates that all actual bean rust samples were correctly detected, although a small number of samples from other classes were predicted as bean rust, slightly lowering the precision. Similarly, the healthy leaf class achieved a precision of 1.0000, recall of 0.9653 and an F1-score of 0.9824. The slightly lower recall value corresponds to the small number of healthy samples misclassified as bean rust, as observed in the confusion matrix. The overall test accuracy reached 99.13%, with both macro and weighted averages of precision, recall and F1-score equal to 0.9913. The macro average indicates balanced performance across all classes, while the weighted average confirms that class distribution did not bias the results.
Fig 5 illustrates the receiver operating characteristic (ROC) curves and precision-recall (PR) curves for all four classes. All four classes achieved an area under the curve (AUC) value of 1.0000, indicating perfect separability between positive and negative samples for each category. The ROC curves closely follow the top-left boundary of the graph, which reflects an optimal classifier with maximum sensitivity and specificity. Similarly, the precision-recall curves demonstrate exceptional classification performance. The average precision (AP) for each class is reported as 1.0000, indicating near-perfect precision across all recall levels. The PR curves remain close to the upper-right corner of the graph, further validating the model’s strong predictive capability, especially in correctly identifying positive disease cases.
Fig 6 presents representative test samples along with their predicted labels and confidence scores generated by the proposed AlexNet-based model. The first sample corresponds to a healthy leaf, correctly classified as healthy leaf with a confidence score of 0.9991, demonstrating the model’s ability to accurately recognize normal leaf texture and color distribution. The second image represents bean rust, which was correctly predicted with a perfect confidence of 1.0000, highlighting the model’s strong capability in detecting rust-specific visual characteristics such as yellowing and lesion patterns. The third example shows a bacterial blight leaf correctly classified with a confidence of 0.9991, indicating that the model effectively captures irregular necrotic lesions and edge discoloration associated with this disease. Similarly, the fourth image of angular leaf spot was accurately predicted with a confidence score of 0.9997, reflecting precise detection of angular-shaped dark lesions and scattered spotting patterns. The consistently high confidence values across all examples demonstrate the robustness and reliability of the trained model. These qualitative results further support the quantitative findings from the confusion matrix, classification report and ROC analysis, confirming the effectiveness of the proposed approach for automated haricot bean leaf disease classification.
Previous studies have applied AlexNet and other deep learning architectures for plant disease classification across different crops, demonstrating varying levels of performance depending on dataset size, crop type and model design.
Chen et al., (2022) utilized a modified AlexNet on a large-scale tomato dataset consisting of 22,930 images across 10 classes and achieved an accuracy of 98%, highlighting the effectiveness of CNN-based models for multi-class classification tasks. Similarly,
Kewat (2023) applied AlexNet to an ImageNet-based plant disease and pest dataset and reported 96.50% accuracy, confirming its suitability for agricultural image analysis.
In contrast, studies using smaller datasets reported relatively lower performance.
Sholihin et al., (2024) implemented a modified AlexNet for cassava classification using 1,400 images and achieved 87% accuracy, indicating the influence of dataset size and variability on model performance.
Fang et al., (2025) applied AlexNet for bean leaf disease classification (angular leaf spot, rust and healthy) and obtained 95.4% training accuracy but only 82.81% overall accuracy, suggesting challenges in distinguishing visually similar classes under real conditions.Furthermore, comparative architecture studies highlight the advantage of advanced models.
Mishra et al., (2025) evaluated CNN, AlexNet, ResNet and EfficientNet on potato (2,152 images) and mango (4,000 images) datasets, where EfficientNet achieved the highest accuracy of 97.8%, outperforming other architectures in terms of generalization.
Compared to these existing studies, the present work demonstrates superior performance. The proposed modified AlexNet model, enhanced with batch normalization and dropout, achieved 99.13% test accuracy on a dataset of 3,447 haricot bean images. Despite using a moderate-sized dataset, the model shows improved robustness and generalization capability, outperforming several previously reported approaches. This confirms the effectiveness and competitiveness of the proposed architecture for plant disease classification tasks.