<![CDATA[Paddy (Oryza sativa L.) is a strategic staple food commodity in Indonesia, yet its production is frequently disrupted by various plant diseases that cause significant yield losses each year. Conventional visual disease identification is inefficient and prone to error, necessitating the adoption of more reliable, automated diagnostic technology. This study compares the performance of two pre-trained convolutional neural network architectures, VGG16 and ResNet50, in classifying rice leaf images into four categories: Brown Spot, Leaf Blast, Healthy Rice Leaf, and Rice Hispa. Both models were fine-tuned using transfer learning under an identical experimental configuration, including the same data split, optimizer, learning-rate schedule, and augmentation pipeline, to ensure a controlled architectural comparison. Model evaluation was conducted using accuracy, precision, recall, F1-score, the Matthews Correlation Coefficient (MCC), confusion matrix analysis, training convergence behaviour, inference-time computational efficiency, and Grad-CAM interpretability visualization. Experimental results show that ResNet50 achieved a test accuracy of 99% and an MCC of 0.9874, outperforming VGG16, which achieved a test accuracy of 95% and an MCC of 0.9306. Confusion matrix analysis revealed that ResNet50's errors were concentrated almost exclusively within the visually similar brown spot–leaf blast class pair, whereas VGG16 exhibited additional confusion between the healthy rice leaf and rice hispa classes. ResNet50 also demonstrated faster and more stable training convergence, more spatially coherent Grad-CAM activation patterns, and substantially higher throughput under batched inference (356.27 vs. 207.63 images/second at batch size 32), while VGG16 retained a marginal latency advantage under single-image inference. These findings indicate that, under the configuration examined in this study, ResNet50 is the more suitable architecture for rice leaf disease classification, offering an advantageous combination of accuracy, interpretability, and computational efficiency for potential deployment in agricultural monitoring systems.]]>
