<![CDATA[Manual detection of pests on mustard greens (caisim) is a major constraint in reducing harvest productivity, as manual methods are inefficient, time-consuming, and require specialized expertise. Furthermore, deep learning models often suffer from overfitting when applied to limited agricultural datasets. This study aimed to develop and compare the effectiveness of a Convolutional Neural Network (CNN) from scratch model versus the VGG16 transfer learning architecture for automatic classification of healthy and pest-affected mustard leaf images. A dataset of 1,000 images was used for training and testing across four experimental scenarios (A to D), with Percobaan C being the optimized CNN from scratch model (using data augmentation) and Percobaan D using VGG16. The results showed that the VGG16 transfer learning model achieved the highest test accuracy of 95.0% (F1-score: 0.95), while the optimized CNN from scratch model achieved 92.0% (F1-score: 0.92). Therefore, transfer learning with VGG16 is the most effective and optimal approach, demonstrating superior performance and efficiency by achieving high accuracy without complex data augmentation.]]>
Copyrights © 2026
