Comparative Analysis of CNN Models for Plant Disease Classification Using Augmented PlantVillage Dataset

Abstract

Plant diseases pose significant threats to agricultural production, and need early and precise identification to avert them. Traditional approaches rely on subjective, time-consuming, and error-prone manual assessment. This study investigates the performance of CNN models with different capacities for the classification of plant diseases using the extended PlantVillage dataset. Within the scope of the study, transfer learning models such as DenseNet-121, EfficientNet-B3, MobileNet-V3, and ResNet-50 were compared. The results reveal that, despite MobileNet-V3 having lower capacity than other models, it exhibits superior performance, indicating no direct relationship between model capacity and performance. MobileNet-V3 provides the highest accuracy among the models, achieving 99.85% with a parameter size of 2.4 million. This study investigates the effectiveness of various CNN designs in classifying plant diseases, with a particular focus on identifying a lower-capacity model suitable for use, especially on mobile devices.