Autonomous Pollination System for Tomato Plants in Greenhouses: Integrating Deep Learning and Robotic Hardware Manipulation on Edge Device

Abstract

Air pollination is important for plants like tomatoes as it plays a vital role in the reproduction and productivity of crops. Tomatoes, being predominantly selfpollinated, depend on airflow to transfer pollen from the anther to the stigma within the flower. In isolated environments like greenhouses, ensuring effective air pollination presents a significant challenge. Due to limited natural airflow, greenhouses often rely on mechanical ventilation systems which cannot adequately replicate the natural pollination process. In this work, a low-cost but efficient autonomous pollination system is developed for tomato plants in greenhouses that utilize both deep learning capability as well as robotic hardware manipulation. The deep learning model is based on the YOLOv8 Nano architecture that is customized with a transformer module which achieved an mAP@50 score of $\mathbf{9 4. 6}$ percent for mature flower detection. Explainable AI (XAI) like EigenCAM and XGrad-CAM approach showed correct visualization over the flower area with minimum noise. The developed model was implemented on the Nvidia Jetson Nano where the Bot-SORT tracking algorithm was used to avoid over-pollination of the same flower. For real-time application, a Graphical User Interface (GUI) app was developed which functioned together with the hardware system using closed-loop feedback control on a pan-tilt mechanism for targeting each mature tomato flower. This technique ensured consistent pollination throughout the day and offers efficiency, precision, and cost-effectiveness by increasing yield.