Comparison of ResNet50, ResNet101, and ResNet152 Architectures in Image-Based Rice Leaf Disease Classification

Abstract

Rice leaf diseases are one of the main threat that can reduce rice crop productivity especially if they are not detected at an early stage. Conventional disease identification still has limitations because it relies on visual observation and the experience of farmers. Therefore, this study proposes a rice leaf disease classification approach based on digital images using deep learning methods. This study aims to compare the performance of three Residual Network architectures, namely ResNet50, ResNet101, and ResNet152. The dataset used was collected from three public Kaggle datasets, consisting 7.322 images divided into four classes (healthy, hispa, sheath blight, and brown spot). The dataset was split into training, validation, and testing sets with a ratio of 70:20:10 and processed through image preprocessing and data augmentation. All models were trained using a transfer learning approach with the same training configuration to ensure a fair comparison. Model performance was evaluated with the test sets using loss, accuracy, and confusion matrix analysis. The experimental results show that ResNet101 achieved the best performance with a loss value of 0,0146 and an accuracy of 0,9973. Followed by ResNet50 with an accuracy of 0,9918, and ResNet152 with an accuracy of 0,9837. These results indicate that ResNet101 provides the best balance between network depth and classification performance.