TrackPlant3D: 3D organ growth tracking framework for organ-level dynamic phenotyping

Abstract

The extraction of dynamic plant phenotypes is highly important for understanding the process of plant phenotype formation and formulating growth management plans. Although rapid progress has been made in the analysis of the efficiency and throughput of static phenotypes, dynamic growth tracking methods are still a key bottleneck for dynamic phenotyping. The major challenges related to organ growth tracking include the nonrigid deformation of organ morphology during growth, the high frequency of growth events, and a lack of spatiotemporal datasets. Inspired by the phenomenon in which a human naturally compares two similar three-dimensional objects by overlapping and aligning them, this study proposes an automatic organ growth tracking framework—TrackPlant3D—for time-series crop point clouds. The unsupervised framework takes crop point clouds at multiple growth stages with organ instance labels as input and produces point clouds with consistent organ labels as organ-level growth tracking outputs. Compared with the other two state-of-the-art organ tracking methods, TrackPlant3D has better tracking performance and greater adaptability across species. In an experiment involving maize species, the long-term and short-term tracking accuracies of TrackPlant3D both reached 100%. For sorghum, tobacco and tomato crops, the long-term tracking accuracies were 81.25%, 64.13% and 86.75%, respectively, and the short-term tracking accuracies were all greater than 85.00%, demonstrating satisfactory tracking performance. Moreover, TrackPlant3D is also robust against frequent organ growth events and adaptable to different types of segmentation inputs as well as to inputs involving inclination and rotation disturbances. We also demonstrated that the TrackPlant3D framework has the potential for incorporation into a fully automatic dynamic phenotyping pipeline that integrates organ segmentation, organ tracking, and dynamic monitoring of phenotypic traits such as individual leaf length and leaf area. This study may contribute to the development of dynamic phenotyping, digital agriculture, and the factory production of plants.