State-of-the-art CNN models for plant disease classification: a comparative study

Plant diseases pose a significant threat to global food security, resulting in substantial agricultural losses. Traditional methods of diagnosing plant diseases rely on manual observation, which is timeconsuming and prone to errors. Recent advancements in deep learning, particularly Convolutional Neural Networks (CNNs), offer a promising solution for automating and improving the accuracy of plant disease classification through image analysis. This study evaluates the performance of five stateof-the-art CNN architectures: VGG16, GoogleNet, EfficientNet B0, ResNet50, and DenseNet201 for plant leaf classification using the PlantVillage dataset, which comprises 54,305 images across 38 classes. Among the models, VGG16 achieved the highest accuracy of 93.75% and recall of 93.75%, with a low loss of 0.48, though it has a larger parameter size of 56.79 MB. GoogleNet followed closely with 88% accuracy, a high F1 score of 0.93, and a balanced size of 85.51 MB, despite a higher loss of 0.96. ResNet50 demonstrated strong performance with 84.36% accuracy and 0.35 loss, but was the most resource-intensive at 100.17 MB. EfficientNet B0, the smallest model at 18.62 MB, achieved 84.37% accuracy, whereas DenseNet201 underperformed, attaining only 69.32% accuracy and 1.06 loss, despite its moderate size of 28.01 MB. These findings highlight trade-offs between accuracy and computational efficiency, with VGG16 and GoogleNet excelling in precision, while EfficientNet B0 presents a compact alternative for resource-limited settings. This study provides valuable insights for selecting optimal CNN models for plant disease detection based on specific agricultural needs.

Text 126259.pdf - Published Version Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (1MB) Official URL or Download Paper: http://pertanika.upm.edu.my/pjst/browse/regular-is...

Actions (login required)

View Item