Determining Model Reliability for Leaf Disease Classification and Detection Approach by Evaluation of Performance Measures

Abstract

This research paper presents a comparative analysis of three deep learning architectures—VGG16, DenseNet121, and the proposed EfficientNetV2B2—for the identification of plant diseases through image classification. The study reveals distinct differences in training efficiency, convergence speed, and computational demands among the models. VGG16, characterized by its substantial parameter count of 15 million, exhibited slower convergence and signs of over fitting despite achieving high training accuracy. In contrast, DenseNet121, with only 7.3 million parameters, demonstrated remarkable efficiency and quick convergence, achieving a classification accuracy of 96.25%. The proposed EfficientNetV2B2 excelled with a classification accuracy of 99.29%, effectively handling complex disease patterns and leaf textures. All models successfully classified healthy leaf classes, attributed to their distinct visual characteristics. The findings highlight the importance of model depth, transfer learning, and data preprocessing in enhancing the reliability of automated plant disease classification. EfficientNetV2B2 stands out as the most robust and scalable model suitable for real-world precision farming applications, while the limitations and potential trade-offs of each architecture warrant further investigation across diverse plant species, diseases, and imaging conditions.