<![CDATA[The ripeness level of bananas is a crucial factor that affects the quality, taste, and selling value of the commodity, but the manual sorting process that is commonly carried out is still subjective, inconsistent, and time-consuming. This study aims to implement and evaluate the performance of a VGG16-based Convolutional Neural Network (CNN) architecture in automatically classifying the ripeness level of bananas. The research dataset consists of 5,616 digital images obtained from the Roboflow Universe platform and grouped into six specific classes: freshripe, freshunripe, overripe, ripe, rotten, and unripe. The system development methodology includes data division using stratified splitting techniques, image pre-processing with data augmentation strategies to prevent overfitting, and the application of transfer learning. The model was trained using the Stochastic Gradient Descent (SGD) optimization algorithm with a learning rate of 0.001 for 25 epochs on GPU-based hardware. Performance evaluation was conducted in depth using a confusion matrix, F1-Score metrics, and Precision-Recall curve analysis. The experimental results showed that the VGG16 model achieved an overall accuracy of 97.13%. Class-by-class analysis shows perfect performance in the freshunripe category, although there is a slight decrease in precision in the ripe class due to the similarity of visual characteristics with the overripe class. The stability of the training and validation accuracy curves also indicates that the model has good generalization capabilities. This study concludes that the VGG16 architecture is a reliable and accurate solution to support the efficiency of smart farming systems.]]>
Copyrights © 2025
