<![CDATA[Tomato leaf diseases pose a significant challenge to agricultural productivity, as delayed or inaccurate disease identification can substantially reduce crop yield and quality. Recent advances in deep learning have demonstrated considerable potential for automated plant disease classification; however, many existing approaches rely heavily on transfer learning and computationally intensive convolutional neural network (CNN) architectures. This study proposes an optimized MobileNetV2 framework trained entirely from scratch for tomato leaf disease classification by integrating AutoAugment for automated data augmentation, the AdamW optimizer for effective parameter optimization, and a Cosine Annealing Learning Rate scheduler to achieve stable model convergence. Experiments were conducted using the PlantVillage tomato leaf dataset comprising ten disease categories, including healthy leaves. The proposed model was trained for 50 epochs without employing pre-trained ImageNet weights. Experimental results demonstrated stable convergence, with the training loss decreasing from 1.5569 to 0.5543 and the validation loss decreasing from 2.5431 to 0.5754, indicating good generalization performance without significant overfitting. The proposed framework achieved a best validation accuracy of 92.89%, with 92.95% precision, 92.89% recall, and an F1-score of 92.87%. These findings indicate that the integration of AutoAugment, AdamW, and Cosine Annealing effectively enhances the learning capability of MobileNetV2 even without transfer learning. The proposed approach provides a lightweight and computationally efficient solution suitable for deployment on resource-constrained devices, making it a promising alternative for real-time tomato leaf disease diagnosis in smart agriculture applications.]]>
