Real-time hass avocado ripeness detection using close-range UAV imagery and YOLO-based deep learning

Taco V.L.T.; Accostupa W.E.Q.; Espinoza E.S. · Smart Agricultural Technology (2026) · DOI: 10.1016/j.atech.2026.102547

Resumen

Rapid and accurate assessment of Hass avocado ripeness is essential for improving harvest planning, optimizing picking time, increasing productivity, and reducing postharvest losses. This study presents a real-time system for Hass avocado ripeness detection using close-range UAV imagery and YOLO-based deep learning models. A UAV equipped with a camera captured field images, which were transmitted in real time to a processing station through SCRCPY for fruit detection and ripeness classification. The proposed approach consists of two sequential stages. First, a YOLOv8 object detection model was trained using a dataset of 340 manually annotated images generated through real-world adaptation and labeled in Roboflow. The detector achieved a precision of 0.7951, a recall of 0.7849, an mAP@0.5 of 0.8388, and an mAP@0.5:0.95 of 0.6308, demonstrating reliable fruit localization under field conditions. Second, a YOLOv8 classification model was trained on a dataset of 20,231 cropped fruit images derived from an initial collection of 14,710 images categorized into five ripeness stages (unripe, midripe, ripe, fullripe, and overripe). The classification model achieved an accuracy of 0.79, a macro-average F1-score of 0.77, a weighted-average F1-score of 0.78, a Top-1 accuracy of 0.78, and a Top-5 accuracy of 1.00, confirming its effectiveness for ripeness assessment. The complete system was evaluated under real agricultural conditions with variable illumination and partial fruit occlusion, demonstrating that the integration of close-range UAV imagery and deep learning provides an effective tool for automated avocado ripeness assessment in precision agriculture. The proposed approach was evaluated using close-range UAV imagery acquired in a commercial orchard; therefore, additional validation under different orchards, environmental conditions, and acquisition scenarios is recommended to further assess its generalization capability.

Instituciones: Universidad Nacional de San Agustín de Arequipa