Deteksi dan Klasifikasi Kematangan Pisang menggunakan YOLOv8 dan Model Hibrida MobileNetV2–KNN

Abstract

Banana ripeness is a key indicator of fruit quality, shelf life, and commercial value; however, manual assessment remains subjective and inconsistent, highlighting the need for an automated and objective classification system. This study aims to develop a high-accuracy and stable hybrid model for classifying banana ripeness levels and to implement it in a real-time image-based quality inspection system. This research is an experimental study combined with software development using the Waterfall model. The proposed method integrates YOLOv8 for object pre-filtering and banana detection, MobileNetV2 as a lightweight yet representative deep feature extractor, and K-Nearest Neighbor (KNN) as a Euclidean-distance-based classifier. A dataset of 13,478 images representing four ripeness categories was processed through detection, feature extraction, normalization, and classification stages. Performance evaluations using accuracy, precision, recall, F1-score, confusion matrix, and 5-fold cross-validation show that the model achieves an accuracy of 96.57% with balanced precision–recall values and low cross-validation variance, indicating strong stability and minimal overfitting. The main contributions of this study include the design of the YOLOv8–MobileNetV2–KNN hybrid architecture that overcomes the limitations of manually engineered features and pure CNN models, as well as the development of a GUI-based banana ripeness classification system suitable for real-time smart farming applications.