<![CDATA[The productivity of chili (Capsicum annuum L.) in Indonesia faces significant challenges due to leaf diseases, which are estimated to reduce harvest yields by up to 35%. Conventional detection methods relying on visual observation often lack accuracy due to the high visual similarity between disease symptoms. This study focuses on a comparative evaluation of three leading Deep Learning architectures VGG16, ResNet50, and InceptionV3 in classifying six types of chili leaf diseases using a public dataset. The research implements a high-resolution image strategy (512 x 512 pixels) to maximize the extraction of disease texture features. The methodology employs a Transfer Learning approach with a standardized hyperparameter tuning scheme. Experimental results indicate that the use of high-resolution images significantly impacts model accuracy. The VGG16 architecture achieved the best performance with a testing accuracy of 99.83% and an F1-Score of 1.00, outperforming ResNet50 (99.75%) and InceptionV3 (84.00%). Confusion Matrix analysis demonstrates that VGG16 possesses superior stability in distinguishing disease classes with high visual similarity, such as Bacterial Spot and Cercospora. The study concludes that architectures preserving deep spatial information, such as VGG16, are more effective for high-resolution image-based plant disease diagnosis compared to more complex architectures that perform aggressive feature compression.]]>
Copyrights © 2026
