Deep Tomato Disease Classification Using EfficientNetB3 with Multi-Source Dataset and Hybrid Augmentation

Abstract

Tomato is a globally valuable crop, but its growth and quality are affected by numerous plant diseases. Traditional disease diagnosis techniques are primarily time-consuming, expert-thirsty, and unsuitable for large-scale adoption. Here, we propose a deep learning method to multi-class classification of tomato leaf diseases using the EfficientNetB3 model. In difference to the aforementioned research in which controlled datasets were utilized, in our approach a real-world multi-source image dataset with 25,851 images across 11 classes, among which are present common diseases including bacterial spot, early blight, septoria leaf spot, and healthy class, was used. As an approach to reduce class imbalance and increase the stability of our model, we utilized a hybrid data augmentation strategy consisting of geometric, noise injection-based, and GAN-based synthetically generated data. The model was trained using transfer learning from weights of ImageNet and evaluated with accuracy, precision, recall, and F1-score metrics. It achieved test accuracy of 98.74% at an F1-score of 98.2%, confirming impeccable generalization. These results guarantee the scalability and reliability of EfficientNetB3 as a use case for precision agriculture in broad, particularly for real-world and variable conditions, and could be transferred to other environments and crops with similar issues.