Mask R-CNN based apple flower detection and king flower identification for precision pollination

Abstract

Identification of individual king flowers within flower clusters is a critical step for developing a robotic apple pollination system. Typically, each cluster has five to six individual flowers, and the king flower can be occluded by the lateral flowers because of their central position in a flower cluster. King flowers share identical features (e.g., color, shape, and size) with other flowers. Apple flower clusters open sequentially from the king flower to the lateral flowers in the time of anthesis, which presents an opportunity for selective pollination. Therefore, it is critical to monitor the flower blooming stage for accurately determining the pollination targets and timing. In this study, a machine vision system was developed to acquire images for two apple varieties in the orchard environment. A Mask R-CNN-based detection model followed by a king flower segmentation algorithm were developed to identify and locate the king flowers from an apple flower dataset throughout the blooming stage from first king bloom to full bloom. The flower detection accuracy resulting from the algorithm were compared with ground truth measurements. The king flower detection accuracy varied from 98.7% to 65.6% with respect to the flower stages of 20% to 80% blooming. This information can be used to calculate the percentage of the king flowers and the distribution in the tree canopy. Along with horticultural knowledge, the outcome from the study is expected to provide decision-making information for robotic pollination.