<![CDATA[Tomato leaf disease classification based on digital imagery has become an important approach in supporting smart agriculture, particularly for early detection of plant disease attacks. This study aims to compare the performance of several lightweight Convolutional Neural Network (CNN) architectures, namely MobileNetV3-Small, MobileNetV2, and EfficientNet-B0, in classifying tomato leaf diseases using the PlantVillage dataset. The dataset consists of 3,628 images distributed across 10 classes (9 disease classes and 1 healthy class), with a data split scheme of 80% for training and 20% for validation. Performance evaluation was conducted using classification reports, confusion matrices, and interpretability analysis through Grad-CAM and feature map visualization. The experimental results show that all models achieved very high accuracy, exceeding 99%. EfficientNet-B0 obtained the best performance with a validation accuracy of 99.59%, followed by MobileNetV2 at 99.45% and MobileNetV3-Small at 99.04%. However, model complexity increased along with accuracy, where EfficientNet-B0 had the largest number of parameters and FLOPs. Grad-CAM analysis revealed that higher-accuracy models demonstrated more precise activation focus on leaf lesion regions. This study confirms that lightweight CNN architectures are capable of delivering excellent classification performance while offering strong potential for deployment in plant disease detection systems on resource-limited devices]]>
Copyrights © 2026
