Electronic Theses and Dissertation

YOLO berukuran medium, yaitu YOLOv8m, YOLOv9c, YOLO11m, YOLO13m, dan YOLO26m, menggunakan tiga strategi pelatihan: baseline, weighted sampling, dan optimasi hiperparameter berbasis genetic algorithm. Evaluasi dilakukan menggunakan precision, recall, F1-score, mAP50, dan GFLOPs. Hasil menunjukkan bahwa optimasi genetic algorithm memberikan peningkatan performa paling konsisten pada seluruh model. YOLOv9c memperoleh mAP50 tertinggi sebesar 0,8110, sedangkan YOLOv8m menawarkan keseimbangan terbaik antara akurasi dan efisiensi komputasi dengan mAP50 0,8043 dan kompleksitas 105 GFLOPs. Temuan ini diharapkan dapat menjadi referensi empiris dalam pemilihan arsitektur segmentasi berbasis YOLO untuk aplikasi pemantauan perkebunan sawit secara presisi.

Accurate identification of individual oil palm trees is an important component in supporting precision agriculture and sustainable plantation management. The integration of UAV imagery and deep learning has become an effective approach for large-scale plantation monitoring; however, most studies still employ bounding box-based object detection methods. Although effective for counting trees, such an approach provides only limited spatial information and is unable to support agronomic analyses that require pixel-level canopy characterization. To address these limitations, this study proposes the application of deep learning-based instance segmentation on UAV imagery of oil palm plantations to obtain pixel-level segmentation of each individual tree. This study compares five medium-sized YOLO models, namely YOLOv8m, YOLOv9c, YOLO11m, YOLO13m, and YOLO26m, using three training strategies: baseline, weighted sampling, and genetic algorithmbased hyperparameter optimization. The evaluation was conducted using precision, recall, F1-score, mAP50, and GFLOPs. The results show that genetic algorithm optimization yields the most consistent performance improvement across all models. YOLOv9c achieved the highest mAP50 of 0.8110, while YOLOv8m offered the best balance between accuracy and computational efficiency with an mAP50 of 0.8043 and a complexity of 105 GFLOPs. These findings are expected to serve as an empirical reference for selecting YOLO-based segmentation architectures for precision oil palm plantation monitoring applications.