<![CDATA[Rice is a primary global food commodity, yet its productivity is frequently threatened by various diseases that significantly reduce both yield quality and quantity. Traditional manual diagnosis by farmers is often subjective, time-consuming, and prone to inaccuracies, necessitating more efficient automated solutions. This research evaluates the ResNet50 architecture for the automated classification of rice leaf diseases through digital image analysis. The study specifically investigates the model's performance on a specialized dataset and analyzes how different data splitting ratios influence accuracy and stability. A public dataset comprising four classes—Hispa, Healthy, Leaf Blast, and Brown Spot—was employed. The data underwent rigorous labeling, pre-processing, and augmentation to enhance sample diversity before being partitioned into training and testing sets using three ratios: 85:15, 80:20, and 90:10. The ResNet50 model was implemented using transfer learning with pre-trained ImageNet weights and fine-tuned on the classification layers. Experimental results reveal that the 85:15 split ratio achieved the highest accuracy of 81.48%, followed by 78.77% for the 80:20 ratio and 76.21% for the 90:10 ratio. These findings suggest that ResNet50 provides competitive performance for rice disease detection. Furthermore, achieving an optimal balance between training and testing data is critical for maximizing model generalization within modern smart farming applications.]]>
