Abstract
Rice is grown almost everywhere in the world, especially in Asian countries, because it is part of the diets of about half of the world's population. However, farmers and planting experts have faced several persistent agricultural obstacles for many years, including many rice diseases. Severe rice diseases might result in no grain harvest; hence, in the field of agriculture, a fast, automatic, less expensive, and reliable approach to identifying rice diseases is widely needed. This paper focuses on how to build a lightweight deep learning model to detect rice plant diseases more precisely. To achieve the above objective, we created our own CNN model "LW17" to detect rice plant disease more precisely in comparison to some of the pre-trained models, such as VGG19, InceptionV3, MobileNet, Xception, DenseNet201, etc. Using the proposed methodology, we took UCI datasets for disease detection and tested our model with different layers, different training-testing ratios, different pooling layers, different optimizers, different learning rates, and different epochs. The Light Weight 17 (LW17) model reduced the complexity and computation cost compared to other heavy deep learning models. We obtained the best accuracy of 93.75% with the LW17 model using max pooling with the "Adam" optimizer at a learning rate of 0.001. The model outperformed the other state-of-the-art models with a limited number of layers in the architecture.