Tomato leaf disease classification using DenseNet-121 with data augmentation and fine-tuning

Abstract

In recent years, accurate classification of agricultural images has become increasingly important to support precision farming and crop disease monitoring. However, achieving reliable performance remains challenging due to visual similarity between disease classes and dataset variability. This study presents an applied evaluation of DenseNet-121 combined with data augmentation and fine-tuning for multi-class tomato leaf disease classification. Experiments were conducted using a publicly available tomato leaf image dataset consisting of 5,000 images across 10 classes. All images were resized to 64×64 pixels and split into 80% training and 20% testing sets using a stratified strategy. Data augmentation was applied exclusively to the training data to improve generalization. The experimental results show a progressive performance improvement across training stages, achieving a final classification accuracy of 98.44% with a loss of 4.72% after fine-tuning. Per-class evaluation indicates strong performance across most disease categories, with minor confusion observed among visually similar classes. While the results demonstrate the effectiveness of the proposed training strategy under controlled experimental conditions, further validation using real-field images is required. Overall, this study shows the potential of DenseNet-121 with transfer learning to support tomato leaf disease classification in precision agriculture applications.