A Leaf Segmentation and Phenotypic Feature Extraction Framework for Multiview Stereo Plant Point Clouds

Abstract

As a plant organ with the largest surface area, leaves are the main place where photosynthesis and respiration take place. High-throughput phenotyping of crop leaves is of great significance for breeding, growth monitoring, and increasing crop yield. Due to the highly complex and diversified plant structures, automated leaf segmentation and phenotypic feature extraction remain to be challenging tasks. In this article, we propose a novel five-stage framework that comprises multiview stereo point cloud reconstruction, preprocessing, stems removal in canopy, leaf segmentation, and leaf phenotypic feature extraction to carry out leaf phenotyping on two types of ornamentals-Maranta arundinacea and Dieffenbachia picta. The phenotypic traits such as the leaf area, leaf length, width, and leaf inclination angle for each single leaf are calculated and compared with ground truths. The experimental results show that the average accuracy of calculated leaf area of the two species reached 96.8% and 97.8%, respectively. The average errors of both the calculated leaf length and width of Maranta arundinacea are less than 4.0%, and for Dieffenbachia picta, the average errors of calculated leaf length and width are both no higher than 4.7%. The average errors of calculated leaf inclination angle for the two plant species are 2.9° and 3.0°, respectively.