Mobile-Xcep hybrid model for plant disease diagnosis

Abstract: More than 80% of the human diet depends on plants. Diseases in plants can cause food insecurity, colossal crop loss, and low profit for growers. Therefore, autonomous systems are required for the diagnosis of diseases in plants. Transfer learning and fine-tuning were used in the research to develop a hybrid model for disease diagnosis. Two datasets were formed for food grains, rice, and maize. The hybrid model developed was trained and tested on the PlantVillage and developed datasets. The MobileNet model is pre-trained on the ImageNet dataset, and the middle flow of the Xception model is repeated once in our approach. The proposed hybrid model improves performance compared to the pre-trained MobileNet and Xception models fine-tuned for the same task. It achieves a validation accuracy of 0.9890 on the PlantVillage dataset and a testing accuracy of 0.9940 and 0.9968 on the developed maize and rice leaf dataset. A testing accuracy of 0.5569 was observed when trained with maize leaves from the PlantVillage dataset and tested on the maize leaf dataset. A testing accuracy of around 0.6959 was achieved when the proposed model was trained on the maize leaf dataset and tested on the maize leaf images from the plant village dataset. An increase of testing accuracy by 5% and 4% for the maize crops,4%, and 2% for rice crops, and around 0.7% and 2% increase in validation accuracy for the PlantVillage dataset was shown by the proposed hybrid model when compared to the pre-trained models, MobileNet and Xception trained for the same task.