Pengembangan Sistem Penghitungan Semangka Berbasis Kecerdasan Buatan Menggunakan Algoritma YOLO

Abstract

Accurate watermelon counting plays an important role in precision agriculture, particularly for yield estimation, harvest planning, and post-harvest distribution management. In practical field conditions, watermelon counting is still largely performed through manual visual estimation, which are time-consuming, prone to human error, and inefficient for large cultivation areas. To address these limitations, this study develops an artificial intelligence–based automated watermelon counting system using object detection algorithms from the You Only Look Once (YOLO). Aerial image was collected using an unmanned aerial vehicle (UAV) over a watermelon cultivation area at PT. East West Seed Indonesia. The captured images were processed into orthomosaic imagery to ensure geometric consistency and uniform spatial scale. The orthomosaic data were then annotated using QGIS and cropped into 640×640 pixel image patches to meet the input requirements of YOLO-based models. In total, 900 images containing 10.893 annotated watermelon objects were used for model training and evaluation. This research compares the performance of YOLOv8, YOLOv9, YOLOv10, and YOLOv11 based on precision, recall, mAP@50, mAP@50–95, processing time,and performance stability across different UAV flight latitude. The experimental results indicated that YOLOv8 consistently achieved the most stable performance, particularly in detecting small watermelon objects at higher capture altitudes, as indicated by higher precision and recall values with lower standard deviation. In contrast, YOLOv9 exhibited greater performance variability under changes in object scale. Overall, YOLOv8 is the most suitable model for implementing an automated UAV-based watermelon counting system. The proposed system supports data-driven agricultural management and contribute to the advancement of precision agriculture.