Document Type : Research Paper

• pardazesh tasvir

Introduction

Tomato (Solanum lycopersicum) is one of the most important agricultural crops worldwide and plays a critical role in food security and agricultural sustainability. However, tomato production is continuously threatened by a wide range of fungal, bacterial, and viral diseases that can significantly reduce yield and quality if not detected at early stages. Timely and accurate disease diagnosis is therefore essential for effective crop management, reduction of economic losses, and minimization of excessive pesticide usage. Conventional disease identification based on visual inspection relies heavily on expert knowledge, is time-consuming, subjective, and often unreliable, particularly when different diseases exhibit visually similar symptoms.

Recent advances in precision agriculture have accelerated the adoption of artificial intelligence and computer vision techniques for automated plant disease detection. Deep learning, and especially convolutional neural networks (CNNs), has demonstrated strong capability in learning discriminative features directly from leaf images. Nevertheless, many existing studies primarily report results from a single training run and focus mainly on accuracy, while overlooking the statistical stability and reliability of model performance. This limitation raises concerns regarding the robustness and practical applicability of proposed models. To address this gap, the present study provides a systematic comparative evaluation of a baseline CNN trained from scratch and several transfer learning architectures, with a particular emphasis on performance consistency across multiple independent runs.

Materials and Methods

The experiments were conducted using the tomato subset of the publicly available PlantVillage dataset, which includes RGB images of healthy leaves and nine common tomato diseases, namely Bacterial Spot, Early Blight, Late Blight, Leaf Mold, Septoria Leaf Spot, Two‑Spotted Spider Mites, Target Spot, Tomato Mosaic Virus, and Tomato Yellow Leaf Curl Virus. After data refinement, the dataset was divided into training, validation, and test sets following a standard split protocol.

A total of six deep learning models were implemented using the PyTorch framework. These models included a custom Baseline CNN trained from scratch and five transfer learning architectures—ResNet-18, ResNet-34, ResNet50, EfficientNet‑B0, and DenseNet-121—initialized with ImageNet pre-trained weights. All input images were resized to 224×224 pixels and normalized using ImageNet statistics. Data augmentation techniques, including random rotations and horizontal flips, were applied to enhance generalization.

To ensure statistical reliability and reduce the influence of random weight initialization, each model was trained and evaluated under 25 fully independent runs. Model performance was assessed using accuracy, precision, recall, F1-score, Matthews correlation coefficient (MCC), area under the ROC curve (AUC), and confusion matrices. Mean values and standard deviations across the repeated runs were reported to capture both performance level and stability.

Results and Discussion

The experimental results revealed a pronounced performance gap between the baseline model and transfer learning architectures. The Baseline CNN exhibited clear signs of overfitting and high variability across runs, leading to inferior generalization and unstable classification behavior. In contrast, all transfer learning models achieved substantially higher accuracy and more consistent performance.

Among the evaluated architectures, DenseNet-121 emerged as the most reliable model, achieving the highest average accuracy and F1-score while exhibiting the lowest standard deviation across the 25 independent runs. This indicates a high degree of statistical stability and robustness to random initialization. Qualitative analysis of confusion matrices further showed that DenseNet-121 significantly reduced misclassifications among visually similar disease classes, demonstrating superior class-level discrimination. While other transfer learning models such as ResNet and EfficientNet also delivered strong performance, their variability across runs was slightly higher compared to DenseNet-121.

Conclusions

This study presents a comprehensive and statistically reliable evaluation of deep learning architectures for tomato leaf disease classification. The findings confirm that transfer learning is crucial for achieving both high accuracy and robust generalization in agricultural image analysis. The baseline CNN trained from scratch failed to deliver consistent and dependable results, whereas pre-trained deep architectures showed clear advantages. DenseNet-121, in particular, provided the best balance between classification accuracy and statistical stability across repeated executions. The main contribution of this work lies in highlighting the importance of multi-run stability analysis and qualitative confusion matrix evaluation as essential criteria for selecting reliable models in real-world smart agriculture applications. Future research will explore extending this framework to field-acquired datasets and investigating transformer-based models with an emphasis on improving generalization and stability.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Declaration of Generative AI and AI-assisted technologies in the writing process

there is nothing to disclose.

Data availability statement

Data available on request from the authors.

Acknowledgements

The authors would like to express their sincere appreciation to the Vice-Chancellor for Research of the University of Tehran for the moral support provided during the implementation of this study. The authors also thank the respected reviewers for their constructive structural and scientific comments.

Ethical considerations

The authors avoided data fabrication, falsification, and plagiarism, and any form of misconduct.

Conflict of interest

The authors declare no conflict of interest.