UC San Diego

Developing automated methods to efficiently process large volumes of point cloud data remains a grand challenge for 3D plant phenotyping applications. Phenotyping is one of the key pre-cursors towards developing more efficient breeding strategies and improvement of crop yields. Using 3D point clouds for plant phenotyping helps to eliminate problems that are usually encountered by performing phenotyping with images. However, point clouds come with their own set of problems, which include registration errors, missing data, sensor noise, and scalability. Point clouds of plants also typically have problems related to ``natural'' noise caused by small features such as trichomes or natural shape irregularities; further, handling the diversity of organic structures observed across plant species is challenging. All these nuances need to be considered to ensure quality extraction of phenotyping traits.

Here, we develop a collection of graph-theoretic and machine learning methods to tackle three primary challenges in plant phenotyping: leaf/branch classification, leaf counting, and branch skeletonization. For classification, we assess and validate the accuracy of our methods on a dataset of 54 3D shoot architectures representing multiple species, growth conditions, and developmental time-points. Using deep learning, we achieved 97.8% accuracy on classifying leaves versus branches; critically, we also demonstrate the robustness of our method to growth conditions and species that have not been trained on, which is important in practical applications but is often untested. For leaf counting, we developed an enhanced region growing algorithm to reduce over-segmentation; this method achieved 86.6% accuracy, outperforming prior methods developed for this problem. Finally, for branch skeletonization, we developed an enhanced tip detection technique that ran an order of magnitude faster and generated more precise skeleton architectures compared to prior methods. In addition, we improved the quality of angles produced by skeletonization. Overall, our improvements enable higher throughput and accurate extraction of phenotypic properties from 3D point cloud data. Finally, all the algorithmic methods presented in this thesis were packaged in a stand alone GUI application, called Plant3D (P3D).