S18 - Session O3 - Recent work on robotic pruning of upright fruiting offshoot cherry systems

Authors: Joseph Davidson *, Alex You, Woo Hyun Maeng, Achyut Paudel, Ramya Jayaraman, Ashley Thompson, Manoj Karkee, Cindy Grimm

The current standard practice of broad-acre orchard management does not result in targeted actions that are optimal for individual fruit trees n this reduces the impact of management decisions and wastes resources while falling short on achieving the yield and quality potential of individual blocks. Our team's overall goal is to improve fresh-market apple quality and yields by matching nitrogen fertilizer to nitrogen demand for individual trees through a combination of i) automated sensing, ii) learning algorithms, iii) decision support tools, and iv) precision application with variable rate technology. In this paper, we present a computer vision and localization system that uses a YOLOv3 neural network architecture for detecting tree trunks, allowing us to construct semantic maps of orchard environments as well as capture tree features in a nondestructive manner. These features, such as canopy vigor and trunk cross-sectional area, are indicative of overall plant nutrition. Using camera, IMU, and GPS data collected from a Jazz apple orchard block, we demonstrate that the network achieves a respectable mean average precision for trunk detection of over 95% and that the semantic maps generated using the trunk detection algorithm allow for robust robot localization and identification of individual trees.