Classification of Rice Diseases Using Leaf Image-Based Convolutional Neural Network (CNN)

Abstract

Rice diseases significantly impact agricultural productivity, making classification models essential for accurately distinguishing rice leaf diseases. Various classification models have been proposed for image-based rice disease classification; however, further performance improvement is still required. This study proposes the use of a convolutional neural network (CNN) to classify rice diseases based on leaf images. The dataset used in this study included leaf images categorized into four conditions: leaf blight, blast, tungro, and healthy. In the initial stage, data preprocessing was conducted, including resizing, augmentation, and normalization. Following preprocessing, a custom CNN architecture was developed, consisting of four convolutional layers, four pooling layers, and three fully connected layers. Each convolutional layer employed a 3 × 3 kernel with a stride of 1 and ReLU activation, while the pooling layers used max pooling with a 3 × 3 kernel and a stride of 2. Using a batch size of 32 and the Adam optimizer, the best test performance was achieved with 100 training epochs and a learning rate of 0.0002, resulting in a training accuracy of 0.9930, a loss of 0.0221, and a test accuracy of 0.9647. Model evaluation demonstrated a balanced performance across precision, recall, and F1 score, each achieving 0.9647, indicating highly effective classification without bias toward any specific class. These findings suggest that the simplified CNN model can deliver competitive classification performance without the need for complex architectures or additional enhancement techniques. The proposed CNN model outperformed existing CNN architectures, such as Inception-ResNet-V2, VGG-16, VGG-19, and Xception.