Electronic Theses and Dissertation

Penelitian ini bertujuan untuk merancang model deteksi hama monyet pada perkebunan durian berbasis algoritma YOLOv8. Dataset yang digunakan berasal dari Roboflow sebanyak 3,380 citra yang dibagi menjadi data training (70%), validation (20%), dan testing (10%), serta ditambah 40 citra lapangan untuk pengujian kondisi nyata. Penelitian membandingkan dua varian model, yaitu YOLOv8n dan YOLOv8s, yang dilatih selama 50 epoch menggunakan optimizer AdamW dengan ukuran citra 640×640 piksel. Evaluasi dilakukan menggunakan metrik precision, recall, F1-score, mean Average Precision (mAP), Intersection over Union (IoU), serta inference time. Hasil validasi menunjukkan bahwa YOLOv8n memperoleh precision sebesar 91.40%, recall sebesar 86.60%, F1-score sebesar 88.90%, mAP@0.5 sebesar 93.70%, dan mAP@0.5:0.95 sebesar 66.61%, sedangkan YOLOv8s memperoleh precision sebesar 90.80%, recall sebesar 88.30%, F1-score sebesar 89.50%, mAP@0.5 sebesar 94.00%, dan mAP@0.5:0.95 sebesar 66.80%. Pada pengujian Roboflow, YOLOv8n memiliki inference time 10.28 ms/foto (97.28 FPS), sedangkan YOLOv8s memiliki inference time 12.76 ms/foto (78.37 FPS). Pengujian video juga menunjukkan YOLOv8n lebih cepat dengan inference time 12.82 ms/frame (77.99 FPS), sementara YOLOv8s memiliki nilai confidence yang sedikit lebih tinggi. Secara keseluruhan, YOLOv8s memberikan akurasi dan kemampuan lokalisasi objek yang lebih baik, sedangkan YOLOv8n lebih unggul dalam kecepatan pemrosesan sehingga lebih sesuai untuk implementasi deteksi hama monyet secara real time.
Kata kunci: YOLOv8, deteksi objek, hama monyet, deep learning, computer vision

This study aims to develop a monkey pest detection model for durian plantations using the YOLOv8 algorithm. The dataset consisted of 3,380 monkey images obtained from Roboflow, divided into 70% training, 20% validation, and 10% testing datasets, with an additional 40 field images for real-world evaluation. Two YOLOv8 variants, namely YOLOv8n and YOLOv8s, were trained for 50 epochs using the AdamW optimizer with an image size of 640×640 pixels. Model performance was evaluated using precision, recall, F1-score, mean Average Precision (mAP), Intersection over Union (IoU), and inference time. Validation results showed that YOLOv8n achieved 91.40% precision, 86.60% recall, 88.90% F1-score, 93.70% mAP@0.5, and 66.61% mAP@0.5:0.95, while YOLOv8s achieved 90.80% precision, 88.30% recall, 89.50% F1-score, 94.00% mAP@0.5, and 66.80% mAP@0.5:0.95. On the Roboflow testing dataset, YOLOv8n achieved an inference time of 10.28 ms/image (97.28 FPS), while YOLOv8s required 12.76 ms/image (78.37 FPS). Video testing also demonstrated that YOLOv8n processed images faster with an inference time of 12.82 ms/frame (77.99 FPS), whereas YOLOv8s produced slightly higher confidence values. Overall, YOLOv8s achieved better detection accuracy and object localization performance, while YOLOv8n provided faster inference, making it more suitable for real-time monkey pest detection applications. Keywords: YOLOv8, object detection, monkey pest, deep learning,computer vision