Optimizing rice leaf disease classification through convolutional neural network architectural modification and augmentation techniques

Abstract

This research focuses on advancing the accuracy of rice leaf disease classification through the integration of convolutional neural network (CNN) and deep learning models. With Indonesia ranking third in global rice production, effective crop management is crucial for sustaining agricultural output. This study employs innovative data augmentation techniques, including random zoom and others, to enhance model training robustness. The experimentation involves eight scenarios with varied architectural configurations applied to a residual network-50 (ResNet50) layers model, aiming to optimize disease classification performance. Featuring random zoom without the multilayer perceptron (MLP) component, emerges as the most effective, demonstrating superior accuracy and performance metrics. A grid search is conducted to optimize MLP layers, revealing a three-layer configuration as most effective. We found that the data augmentation and MLP layer can increase the accuracy of the disease classification task. The method proposed in this study is likely to have a much higher proportion of correct disease classification by combining MLP and zoom augmentation. Specifically, the model with three MLP layers and zoom augmentation demonstrated significantly higher accuracy, achieving a test accuracy, precision, recall, and F1-score of 0.92, 0.94, 0.92, and 0.92, respectively.