ANALISIS PERFORMA YOLOV8 PADA DETEKSI KEMATANGAN BIJI KOPI DENGAN BERBAGAI KONFIGURASI PELATIHAN

Performance Analysis of YOLOv8 in Coffee Seed Detection with Various Training Configuration

Abstract

Coffee is a strategic plantation commodity in Indonesia, whose quality is strongly influenced by fruit maturity at harvest. However, maturity assessment in the field is still largely conducted manually, leading to subjectivity and low efficiency. This study aims to analyze the performance of YOLOv8 in detecting three coffee fruit maturity levels—unripe, semi-ripe, and ripe—and to evaluate the impact of different training configurations on model performance. The experiments were conducted using the public “Coffee Cherry” dataset from Kaggle, consisting of 432 images with 1000 bounding box annotations. Several configurations were evaluated, including YOLOv8 variants (n, s, m), image sizes, batch sizes, number of epochs, and data augmentation techniques based on HSV color space, object scaling, and mixup. The results show that YOLOv8s provides the best balance between detection accuracy and computational efficiency. The optimal configuration was achieved using an image size of 416, batch size of 8, and 30 epochs, with augmentation improving recall and mAP@50, particularly for challenging classes. Nevertheless, limited dataset size and diversity remain the main constraints affecting performance. Overall, YOLOv8 demonstrates strong potential for coffee fruit maturity detection under real-world conditions.