Comparative Analysis of Deep Learning Models for Tomato Leaf Disease Classification: Insights and Opportunities

Abstract

This study evaluates the performance of four deep learning models—CNN, ResNet50, VGG16, and MobileNetV2—on the classification of tomato disease images into seven distinct categories: Septoria Leaf Spot, Early Blight, Mosaic Virus, Spider Mites, Target Spot, Leaf Mold, and Healthy Leaf. Using a dataset of 1,100 images, equally distributed across categories, the models were trained on 1,000 images and tested on 100 images to ensure standardized performance evaluation. ResNet50 demonstrated superior performance with an accuracy of 86.8%, precision of 87.72%, recall of 86.8%, and F1 score of 86.74%. VGG16 followed with an accuracy of 78.7% and F1 score of 78.7%, showcasing competitive but slightly lower efficacy compared to ResNet50. The custom CNN model achieved moderate results with an accuracy of 73.9% and an F1 score of 73.57%. Computationally efficient but with lowest performance metrics of 69.4% and 69.52% accuracy and F1 score respectively, MobileNetV2 was an underperformer. Data visualization showed a balanced dataset distribution for unbiased training, and we used data augmentation to improve model generalizability and reduce overfitting. Deep architecture and residual connection demonstrated to play a crucial role in feature extraction and classification. Future work could focus on hyperparameter tuning, more sophisticated architectures (such as EfficientNet) and combining the different architectures in order to maximize performance. It may also help to expand the dataset and to use transfer learning. ResNet50's efficacy for complex image classification tasks is evident from these findings and the potentials for improvement in deep learning based agricultural disease diagnosis are also shown.