Rice Leaf Lesions Detection Using Computer Vision
Agriculture is very important for a country like India where it contributes up to 16% to the national GDP, 82% of the farmers are small and marginal, and their problems rise as irrigating the fields is becoming tougher day by day because of the varied rainfall patterns. Due to improper irrigation, rice crops that require an extensive waterlogged land area for long periods of the day have begun to get infected by different types of diseases that are having a severe effect on the country's agrarian sector output. As paddy is widely consumed which makes it an important crop but it doesn’t remove the likelihood of it being very fragile and susceptible to diseases. Rice is one of the most harmed and wasted crop according to UN crop health reports, among the many we have tried to concentrate over three majorly found diseases and then aimed to provide solution for these and then develop over the solution for other thousands of different trivial rice plant diseases. So we have come forward to solve this issue by informing the farmers much early about their crops being at risk through a web application that can diagnose the rice plant remotely and instantly. At first, proper disease detection is necessary to properly diagnose it. Due to limited data available online we decided to create our own smaller dataset from eclectic sources. In this paper, we present a self-designed algorithm that will detect the disease of rice crops based on four major disease classes. Our model is a hybrid of the EffientNetB5 model designed for smooth implementation on web applications. We will also compare it with other existing methodologies such as the VGG16 architecture model. We proposed a deep learning hybrid model that was trained over selected images of the diseases gathered from standard datasets of Kaggle and few from Google Images. After which training was done for few hours. Training required the use of frameworks such as Keras and TensorFlow for the machine learning part. This trained model was deployed on a web application interface developed by the team from scratch. . The application used our hybrid model that will really reduce the load of servers and resulted into very quick analysis of the images using the machine learning model in the backend.