Plant Disease Identification Using Efficient Fine-tuning of Deep Learning Models

The development of diseases in plants can significantly affect production in agriculture. Deep Learning in recent years has proved to be a promising technique to develop an automatic system for disease diagnosis in plants. The current manuscript focuses on how fine-tuning the state-of-the-art convolution neural network models can be performed for image-based plant disease diagnosis. The models fine-tuned involve VGG-16, Inception V3, ResNet 50, ResNet 101, ResNet 152, MobileNet, NasNet Mobile, and DenseNet 121. In our experiments, both fine-tuned VGG-16 and ResNet-50 were able to classify the diseases more unambiguously than the other models in the case of the PlantVillage dataset. The VGG-16 and ResNet 50 models obtained a test accuracy score of 97.18% and 97.90% respectively. The Inception v3 and the MobileNet models performed best on the maize leaf dataset with a test score of 92.28% and 92.62% respectively. This manuscript can help the researchers get a brief idea about how a model pre-trained can be fine-tuned for a particular task, how it can help reduce training time, show improved performance, avoid overfitting, provide transfer of knowledge, and how explainable artificial intelligence can be used to increase the explainability and the interpretability of the models.