Development of new means for advanced analysis of crop plant shoot architectonics and other phenotype attributes

为了养活迅速增长的人口（预计到2050年将达到97亿），必须加快改善粮食生产。面对全球环境变化，这给农业带来了巨大的挑战。植物茎结构是植物地上部分的三维组织。事实证明，植物构型的改变是提高作物产量的一种很有前途的策略。选择具有理想植株结构的新品种需要准确和广泛的表型性状表征。在该项目中，大豆（Glycine max）将作为模型植物，使用高分辨率地面光探测和测距（LiDAR）传感器获得3D点云。所得数据将用于开发新的分析工具，利用人工智能和计算机视觉在空间和时间上对植物茎部结构进行三维综合表型分析。这些工具可以提取新的高级特征，如3D器官分布和拓扑模式，这些特征是目前使用2D图像的表型方法无法获得的。这个项目有三个重要目标。第一个目标是设计3D深度学习网络来分割点云中的单个器官。三维植物结构的准确分割是单个器官定量表型的关键先决条件。第二个目标是通过重建三维植物模型来表征三维植物结构。本研究将：(1)研究单个器官因遮挡问题在点云中缺失部分的恢复，建立定量结构模型；(2)从整个植物到器官水平提取感兴趣的表型性状。第三个目标是跟踪植物建筑的时空行为。该计划的结果将为植物育种者提供工具，以提高植物育种进展的效率，这些工具将帮助植物科学家探索植物结构如何生长和适应不断变化的环境的原则；此外，他们还将帮助加拿大农民（直接或通过专门的服务提供商）优化作物管理实践，在不影响产量的情况下减少水和能源的使用，提高他们的盈利能力，更好地控制环境足迹。这些工具可以推广到其他作物，如玉米经过轻微修改。该项目代表了一项新的跨学科研究，将人工智能和植物科学相结合，有助于解决几个通用的3D计算机视觉挑战，如3D物体重建、分割和点云中的3D物体跟踪。在教育方面，该项目可以帮助教育和培训加拿大急需的计算植物科学研究人员。在未来，机器人辅助的高通量表型系统将与这些工具相结合，用于田间条件下植物茎结构的自动分析。

To feed the rapidly growing human population, which is projected to be 9.7 billion by 2050, the improvement of food production has to be accelerated. It poses tremendous challenges to agriculture in the face of global environmental change. Plant shoot architecture is the 3D organization of above-ground parts of a plant. It has been proved that modification of plant architecture is a promising strategy to improve crop yield. The selection of new varieties with ideal plant architecture requires accurate and extensive characterization of phenotypic traits. In this program, soybean (Glycine max) will be used as the model plant to obtain 3D point clouds using a high-resolution terrestrial light detection and ranging (LiDAR) sensor. The resulting data will be used to develop novel analytical tools using artificial intelligence and computer vision for 3D comprehensive phenotyping of plant shoot architecture in a spatial and temporal manner. The tools can extract new advanced traits such as 3D organ distribution and topological pattern which cannot be obtained using current phenotyping methods from 2D images. This project has three significant aims. The first aim is to design 3D deep learning networks to segment individual organs in point clouds. Accurate segmentation of 3D plant architecture is a critical prerequisite for quantitative phenotyping of individual organs. The second aim is to characterize 3D plant architecture by reconstructing 3D plant models. This aim will: (1) study recovering missing parts in point clouds for individual organs due to occlusion problems and build quantitative structure models; (2) extract phenotypic traits of interest from the whole plant to organ levels. The third aim is to track spatial-temporal behaviors of plant architecture. The results of this program will provide plant breeders tools to increase the efficiency of plant breeding progress, and the tools will help plant scientists explore principles of how plant architecture grows and adapts to changing environments; Also, they will assist Canadian farmers (directly or through dedicated service providers) optimize crop management practice that will reduce water and energy use without compromising yield, increasing their profitability and better control the environmental footprint. The tools can be generalized to other crops such as maize after minor modification. The project represents a new interdisciplinary research integrating artificial intelligence and plant science and contributes to addressing several general 3D computer vision challenges such as 3D object reconstruction, segmentation, and 3D object tracking in point clouds. Educationally, the project can help educate and train computational plant science researchers who are critically needed in Canada. In the future, robot-assisted high throughput phenotyping systems combining with the tools will be developed for automatic analysis for plant shoot architectures under field conditions.