KLASIFIKASI JENIS PISANG (Musa sp.) LOKAL BERBASIS PENGOLAHAN CITRA DENGAN YOLOv8 MENGGUNAKAN ROBOFLOW DAN GOOGLE COLAB

Bananas are one of the most important horticultural commodities in Indonesia and consist of many local varieties with diverse physical characteristics. Differences in shape, size, and peel color often make manual identification difficult and may lead to misclassification, especially among banana varieties with similar appearances. This condition highlights the need for an image-based classification system that can identify banana varieties more objectively and consistently. This study aims to design and implement an image-based local banana classification system using the YOLOv8 algorithm and to evaluate the system’s ability to classify various local banana varieties.
The research method includes collecting banana image datasets through direct field photography and from internet sources, preparing and annotating the data using the Roboflow platform, and training a classification model using the YOLOv8 algorithm via Google Colab. The dataset consists of eight local banana varieties, namely kepok, susu, kirana, raja, tanduk, nangka, kapas, and hurang (red banana). The training process was conducted until the best model weights were obtained and subsequently used for testing and system implementation.
The results show that the developed classification system is capable of recognizing and classifying local banana varieties with a satisfactory level of performance. The system achieved a precision value of 0.58, a recall value of 0.60, and an F1-score of approximately 0.58, indicating that the classification results meet the research objectives. In addition, the mean Average Precision (mAP) value of 0.622 demonstrates reasonably good detection performance, although it decreases to 0.44 under stricter evaluation conditions. These findings indicate that system performance is influenced by the physical characteristics of the bananas and the number of training images available for each variety, where varieties with clearer visual features and larger datasets tend to be recognized more accurately.