Performance Comparison of VGG-19 and DenseNet-121 Architectures for Rice Plant Disease

Abstract

Rice (Oryza sativa L.) is a major food source that often faces the challenge of crop failure due to various plant diseases. These diseases not only reduce productivity, but are also exacerbated by farmers' limited knowledge in recognizing symptoms and reliance on manual diagnosis that takes a long time. This study aims to compare the performance of two Convolutional Neural Network (CNN) architectures, namely VGG-19 and DenseNet-121, in classifying rice plant diseases based on image processing. Low accuracy and overfitting are problems that are often observed when small datasets are used to train deep learning models, such as Convolutional Neural Networks (CNN). In this study, modifications were made to the VGG-19 and DenseNet-121 architectures so that the model can achieve good accuracy and reduce the risk of overfitting despite using small datasets. The dataset consists of 11,790 images in 9 classes, which are divided into 7545 training data, 1887 validation data, and 2358 testing data. After the training data is segmented, the total number of images in the dataset is 23,580. Before modification, the DenseNet-121 model achieved the highest accuracy of 50.45% and F1-score of 44.83%, while VGG-19 achieved the highest accuracy of 13.84% and F1-score of 7.39%. After making modifications to both models, the test results show that DenseNet-121 achieved an accuracy of 97.76% and F1-score of 96.31%, while VGG-19 achieved an accuracy of 84.82% and F1-score of 87.52%. The advantage of DenseNet-121 lies in its ability to process features more efficiently, resulting in more accurate predictions than VGG-19. This research contributes to the selection of the best model architecture to support automatic diagnosis of rice plant diseases, which is relevant to the agricultural sector in Indonesia.