Statistical Reliability Assessment of Deep Learning Architectures for Tomato Leaf Disease Classification

Accurate and timely detection of tomato leaf diseases is essential for reducing crop losses, improving productivity, and supporting sustainable agriculture. Despite the growing success of deep learning approaches in this field, most existing studies report results based on a single training run and primarily focus on peak accuracy, while neglecting statistical stability. This limitation weakens the reliability and real-world applicability of many proposed models. To address this gap, this study introduces a stability-oriented evaluation framework for a comprehensive comparison of deep learning architectures for tomato leaf disease classification. A baseline convolutional neural network trained from scratch was systematically compared with state-of-the-art transfer learning models from the ResNet, GoogLeNet, EfficientNet, and DenseNet families using a standard benchmark dataset. To ensure a fair and realistic assessment, all models were trained under 25 fully independent runs with random initialization. Performance was evaluated using multiple metrics, including accuracy, precision, recall, F1-score, Matthews correlation coefficient, area under the ROC curve, and the standard deviation of results across repeated runs. The results demonstrate that transfer learning models consistently outperform the baseline CNN not only in terms of average accuracy but also in statistical stability. Among all evaluated architectures, DenseNet121 achieved the most reliable performance, with a mean accuracy of 0.996 and the lowest standard deviation (0.0011). Qualitative analysis of confusion matrices further confirmed reduced inter-class misclassifications. These findings highlight the importance of multi-run stability analysis for selecting dependable deep learning models in practical smart agriculture systems.