Automated Recognition of Rice Grain Diseases Using Deep Learning

To detect and classify various diseases of plants, the images of leaves are the main wellspring of information. Misidentifying various diseases in agricultural crops can lead to significant economic loss and environmental impacts. Rice is the most predominant food crop in Bangladesh. It possesses around 75% of the grossed harvest area of the country. Early diagnosis of rice grain diseases can save tonnes of agricultural products every year. The traditional manual observation of the crop disease is time-consuming and less accurate. For addressing this issue, researchers have utilized different high-performing Convolutional Neural Networks (CNNs) that perform well in the disease classification tasks. But these networks have very complex architecture and not very much suitable for real-life applications. Moreover, there is a limitation of devices and stable connectivity in agricultural areas. So there requires a lightweight efficiency for applying the automated disease detection system. In this paper, we present a custom memory-efficient Convolutional Neural Network (CNN) namely, RiceNet to automatically detect rice grain diseases. An extensive experiment showed the effectiveness of the proposed custom network (RiceNet) which provides a classification accuracy of 93.75%. We also apply the transfer learning approach through various pre-trained lightweight structures and achieved 97.94% accuracy through EfficientNetB0.