Explainable Deep Transfer Learning for Robust Tomato Leaf Disease Classification

Abstract

Automated identification of plant diseases is crucial for advancing precision agriculture and enabling farmers to make informed, timely decisions. This study presents a deep learning-based framework for multi-class classification of tomato leaf diseases using transfer learning with the VGG-19 architecture. A dataset comprising 10,000 images across ten classes, including nine disease categories and one healthy class, was preprocessed and augmented to improve model robustness and generalization. The training strategy employed a two-stage approach: initial feature extraction with frozen, pre-trained layers, followed by selective fine-tuning to adapt the convolutional features to the target domain. Comprehensive evaluation using accuracy, precision, recall, F1-score, and confusion matrices demonstrated the model’s high discriminative capability, achieving an overall accuracy of 93% on the validation set. The results further revealed strong performance in identifying most disease categories, while highlighting classification challenges between visually similar classes, such as Tomato Mosaic Virus and Tomato Target Spot. The contributions of this research include the development of an optimized training pipeline, a reproducible evaluation framework, and insights into the role of transfer learning for agricultural image classification. The findings highlight the potential of deep learning to support automated tomato disease monitoring, with implications for improving crop health management and enhancing agricultural productivity