EVALUASI KINERJA MODEL YOLOv11 UNTUK KLASIFIKASI TINGKAT KEMANISAN BUAH NANAS

Abstrak

Sweetness is a primary determinant of pineapple quality, yet its conventional measurement requires cutting the fruit open and is therefore unsuitable for large-scale, non-destructive sorting. Existing image-based studies on pineapple sweetness generally treat the task as whole-image classification on pre-isolated fruit and report aggregate accuracy, leaving unclear how a model that must localize and classify simultaneously behaves at category boundaries. This study aims to measure and diagnose the performance limits of a single-view YOLOv11 detector in classifying pineapples into three sweetness categories, namely Asam (sour), Manis Ideal (ideal) and Sangat Manis (very sweet). A dataset of 4,475 annotated instances was trained for 50 epochs and evaluated using precision, recall, mean average precision, a confusion matrix, and confidence-threshold sensitivity curves. The model reached an overall mean average precision of 0.555 at an intersection-over-union threshold of 0.5, with a peak F1-score of 0.58; per class, Asam was the most reliable (0.695) while Manis Ideal (0.505) and Sangat Manis (0.465) lagged behind. At the default threshold only 32 to 64 percent of instances per class were classified correctly, far below the recall range observed on the training curves. The novelty of this work lies in the diagnostic characterisation of that gap: errors concentrate almost symmetrically between visually adjacent categories, not between the extremes, and reported recall substantially overstates behaviour at a real operating threshold. A single viewpoint without class-imbalance handling therefore cannot reliably separate boundary categories, which gives the empirical basis for the multi-view refinement pursued next.