A Comparative Variability Analysis of Pre-Trained Deep Learning Models for Sugarcane Leaf Disease Classification

Abstract

This reseach provides a variability study of well known pre-trained deep learning models in automated sugarcane leaf disease classification. VGG16, VGG19, ResNet50, InceptionV3, MobileNetV2, DenseNet121 and EfficientNetB0 were used to transfer learning in various training conditions, such as different epochs and learning rates. The experimental findings indicate that EfficientNetB0 provided the highest performance regularly with a highest accuracy, precision, recall, and F1-score of 0.89 at 50 epochs with a learning rate of 0.001, meaning that it extracts features better and also uses fewer parameters. ResNet50 and DenseNet121 were also found to be stable with accuracies that lie between 0.82 and 0.85 and therefore can be used in practical applications. Nevertheless, InceptionV3 demonstrated relatively low accuracy. The effect of using increasing epochs over time past 50 was a small overfitting in certain models, whereas there was little improvement in learning rates. On the whole, EfficientNetB0 can be considered the most appropriate model, although a tailored architecture can be even more effective.