Application of the deep learning transformer architecture (RF-DETR) for monitoring pollinator trees in orchard rows

Relevance . Accurate monitoring of pollinizer trees is essential for optimizing pollination and increasing the yield of apple orchards. Existing remote sensing methods lack the necessary detail for recognizing individual trees by cultivar within densely planted linear rows. This paper proposes an automated method for monitoring pollinizer trees in orchard rows, based on the advanced RF-DETR deep learning transformer architecture. Methods .The method relies on analyzing imagery from aerial and ground-based platforms and enables the recognition of pollinizers based on the presence of persistent fruits. A labeled dataset (2500 images), verified by expert agronomists, was used for training, ensuring high data quality. Training and comparative analysis of model configurations (Nano, Small, Medium, Large) on the labeled dataset demonstrated that the RF-DETR Medium model achieves high recognition accuracy (mAP@50:95 = 0.658) and enables real-time data processing (5.9 FPS). This combination of features allows the model to be used both on onboard computers of robotic systems and for data processing on server hardware. Field testing of the method in a control row successfully identified a deficit of pollinizer trees and precisely localized problematic sections of the orchard. Results . The obtained results confirm that the use of transformer models allows for moving away from subjective and labor-intensive monitoring methods while ensuring scalable analysis.