An Automatic Diagnosis System of Rice Leaf Diseases Based on ResNet-50 Architecture

Rice diseases exert a substantial detrimental influence on both the yield and quality of rice crops. An accurate and expeditious identification of rice diseases is important to minimize the adverse consequences associated with them. At present, identifying rice diseases mainly relies on manual methods, including visual inspection of disease spots on rice leaves. In this study, we proposed an automated diagnostic method for detecting rice leaf diseases using ResNet-50 convolutional neural network (CNN), aiming to enhance the efficiency and convenience of disease detection. The image dataset of diseases in rice leaves was obtained from Kaggle and consists of three distinct forms of diseases: leaf blast, hispa, and brown spot. In addition, several image augmentation methods were implemented to enhance the size of the dataset and avoid the overfitting problem. The ResNet-50 CNN was employed to create the required prediction model using a substantial dataset of 3737 images representing three distinct categories of rice leaf diseases, including leaf blast (1200 images), hispa (1437 images), and brown spot (1100 images). The ResNet-50 model was optimized by tuning different hyper-parameters, resulting in very high average accuracy for both the training dataset and the test dataset of 98.64% and 98.33%, respectively, by running 85 epochs. The trained model was deployed using Gradio web application for the purpose of diagnosing diseases in rice leaves.