YOLOv11 for Classification of Strawberry Quality and Ripeness

Strawberries are economically important in Indonesia, driven by strong consumer demand and their rich nutritional profile. Traditional harvesting methods, which rely on manual visual inspection, are often inefficient and prone to errors. Real-time multi-object detection presents a promising solution to enhance automation in harvesting, ripeness classification, and post-harvest processing. This study assesses the performance of four YOLOv11 variants—YOLOv11N, YOLOv11S, YOLOv11M, and YOLOv11L—in detecting strawberries across five quality and ripeness categories: Unripe, Grade B Half Ripe, Grade A Half Ripe, Grade B Fully Ripe, and Grade A Fully Ripe. A dataset originally consisting of 3,055 high-resolution strawberry images was expanded through data augmentation to 7,940 images. These were subsequently split into training (7,330 images), validation (305 images), and testing (305 images) sets. Using the AdamW optimizer, cosine annealing learning rate scheduling, a batch size of 16, and an input resolution of 640x640 pixels, all models were trained in the same way. Performance was evaluated based on Precision, Recall, F1-Score, [email protected], [email protected], and inference time. With a Precision of 0.869, Recall of 0.878, F1-Score of 0.87, [email protected] of 0.830, and the fastest inference time of 3.6 ms, the findings show that YOLOv11N performed the best overall, making it appropriate for real-time deployment. YOLOv11M provided a balanced trade-off between accuracy and speed, while YOLOv11S offered competitive accuracy with lower inference latency. YOLOv11L demonstrated strong detection capabilities but with the slowest inference time. These findings affirm the efficacy of YOLOv11-based models in facilitating scalable and intelligent systems for precision agriculture.