plant diseases classification using pre-trained and transfer learning models: a study on rice leaves

Ocular detection of pests by phytosanitary specialists, as a very imperative and challenging task, appears to be time-consuming, costly, and associated with human error in today’s farming processes. in modern agriculture, diagnostic softwares by artificial intelligence are advised to be used by farmers themselves with little time and cost and with more accuracy. in this paper, two different datasets of rice leaf disease have been used with two transfer learning methods for diagnosing rice leaf disease. the first method uses a cnn-based output of a pre-trained model with an appropriate classifier. in the second method, freezing bottom layers, fine-tuning weights of the last layers of the pre-trained network, and adding an appropriate classifier to the model are proposed. for this purpose, seven cnn models have been designed and evaluated. next, the weights of the best model among these seven proposed models is used to train a second dataset of rice leaves with more disease classes and fewer images. using these weights, an accuracy of over 96% is reached which is higher than other comparing methods. furthermore, grad-cam, heat map, and roc diagram are used to observe the diagnostic areas of the best model.