Automated Pruning Decisions in Dormant Canopies using Instance Segmentation

Abstract. Pruning is an operation vital to orchard health and yield. However, pruning is also a laborious process requiring substantial human resources. As such, interest in automated pruning is growing. Automated pruning systems must possess robust machine vision capable of making proper pruning decisions. Deep neural networks are powerful tools for machine vision, and we demonstrate how deep neural networks can be used in an automated pruning system. A pruning rule in the UFO cherry architecture is to remove vigorous (or large diameter) leaders. Stereo images of UFO cherry trees were collected using active and natural lighting. Images were annotated for two classes of objects—trunks and leaders. Two instance segmentation networks (Mask R-CNN) were trained to detect leaders—one using active lighting images and one using natural lighting images. Deep stereo matching enabled generation of synthetic images to increase the size of our training dataset, and large learning rates were employed to accelerate learning (called super-convergence training). Predictions from the active and natural lighting Mask R-CNNs were compared to ground truth annotations for mask IoU, precision, recall, and probability of correctly identifying the largest leader. The active lighting Mask R-CNN demonstrated higher mask IoU, precision, recall, and probability of selecting the largest leader than the natural lighting Mask R-CNN. Overall, the active lighting Mask R-CNN correctly identified the largest leader in 94% of test images. Our results indicate instance segmentation is a robust approach to making automated pruning decisions in the UFO cherry architecture.