Deploying YOLOv10 for High-Precision Date Branch Detection in Smart Date Palm Harvesting

Abstract

Automated detection of date branches is a key enabler technology for smart agricultural robotics, facilitating efficient and precise harvesting in oasis environments. This paper presents a real-time vision-based system for the detection of date branches using the YOLOv10 object detection algorithm. A diverse dataset of over 8,000 annotated images was collected from the Kebili region in Tunisia, reflecting real-world challenges such as variable lighting, occlusion, and the frequent use of protective coverings. Images were annotated and augmented, resulting in a robust training set of 5,530 images. The YOLOv10 model was trained and deployed on a Jetson Orin NX, achieving a mean Average Precision (mAP@0.5) of 93.5%, a recall of 98%, and a precision of 100% at a confidence threshold of 0.959, with an inference speed of 85.3 ms per image. Comparative analysis with YOLOv8-m demonstrated that YOLOv10-m not only matches detection accuracy but also achieves higher computational efficiency and a smaller model size, making it highly suitable for real-time deployment in resource-constrained agricultural settings. The results validate the effectiveness and practicality of the proposed approach for automated date branch detection, paving the way for fully automated harvesting solutions in smart agriculture.