Abstract:
This project responds to an urgent need for early diagnosis and identification of diseases affecting the rice leaves, which are the most common and dangerous to food security worldwide. Rice is one of the main foods for a large part of the population of the world, and several diseases affect it and lead to a decrease in yield and quality. To tackle this challenge, a deep learning-based approach was taken, leveraging convolutional neural networks (CNNs) to automatically classify four prevalent rice diseases: Brown Spot, Leaf Blast, Neck Blast, and Healthy leaf are four diseases of the plant. About 4,078 rice images were taken from the Kaggle of rice leaves and the authors generated a new and enhanced dataset of 11,062 rice Images using image data augmentation including rotation, shifting, and flipping. Hence, it tested various architectures of convoluted neural networks such as Efficient Net (B0, B1, B2, B3), ResNet50, and InceptionV3, among others on this boosted dataset. Altogether, after the experiments on the given models have been carefully compared and analysed, EfficientNet-B1 showed the highest results with the accuracy reaching the level of 96.84% on the test set because it is one of the most efficient and powerful model. This model showed a fair level of accuracy in all the disease classes and proved to have a good capacity to generalize when tested on an entirely different set of data known as the cross-dataset. We have applied the knowledge derived from this successful development and evaluation of this system powered by AI to enhance early and accurate rice disease detection to move closer to better crop management, higher yield, and improved food security.
