Harnessing AI-powered big data techniques for 3D plant architecture phenotyping and growth pattern modeling

全球人口的持续快速增长和气候变化给农业生产系统带来了巨大挑战。我们必须加快育种进程，培育适应不断变化的环境的新作物品种，以确保粮食安全。基因组革命提供了前所未有的力量来设计新的和先进的作物品种与基因组合在植物育种和选择计划。然而，将基因组信息与表型信息联系起来仍然是费力、昂贵和不精确的。快速、重复测量作物表型参数的高通量植物表型技术是植物育种的主要瓶颈。在过去的十年中，二维成像处理技术被广泛应用于植物表型分析。然而，这些方法很难表征三维表型特征。将3D数据转化为有意义的表型信息仍然是一个瓶颈。在这个项目中，我们建议开发新颖的人工智能大数据分析技术，以3D方式描述和模拟从器官到整个植物尺度的植物茎结构和生长模式。本项目以鹰嘴豆（Cicer arietinum）为模式植物。将从500个遗传广泛的鹰嘴豆品种中选出适应魁北克/加拿大条件的20个品种；对于每个品种，我们将在麦吉尔大学的温室里重复种植三次。将开发一种低成本的多视点摄影测量系统，在20个发育时间点扫描60种植物，建立高分辨率的点云序列数据集。然后，将开发一种新颖的标记高效3D深度学习网络，用于单个植物的端到端实例分割。我们的目标是只使用大约0.5%的点来标记分割模型训练。此外，还将提取器官和整个植物水平的三维表型性状。此外，还将开发一个动态点云建模框架来表征植物建筑的时空生长模式。我们将验证我们的方法，并通过在麦吉尔大学麦克唐纳校区的Emile A. Lods农场进行鹰嘴豆优良品系的育种和发展试验来获得见解。总的来说，该项目的成功实施可以加快植物育种进程，增强对植物如何适应不断变化的环境的原理的理解。此外，开发的点云数据集可以帮助缩小数据差距，从而实现广泛的新研究和应用。在教育方面，我们将帮助培养加拿大和全球计算植物科学急需的跨学科HQP。

The continuous rapid growth in the global population and climate change is resulting in tremendous challenges for agricultural production systems. We have to speed up the breeding process to produce new crop varieties adapted to changing environments to ensure food security. The genomic revolution has provided unprecedented power to engineer new and advanced crop cultivars with gene combinations in plant breeding and selection programs. However, it is still laborious, expensive, and imprecise to relate genomic information to phenotypic information. High throughput plant phenotyping technologies that can rapidly and repeatedly measure phenotypic crop parameters are a major bottleneck in plant breeding programs. Over the past decade, 2D imaging processing technologies have been widely applied for plant phenotyping. However, the methods are hard to characterize 3D phenotypic traits. Converting 3D data into meaningful phenotypic information remains a bottleneck. In this program, we propose to develop novel AI-powered big data analytics technology to characterize and model plant shoot architectures and growth patterns from organ to whole plant scales in 3D.Chickpea (Cicer arietinum) will be used as the model plant in this program. A population of 20 varieties adapted to Quebec/Canadian conditions will be selected from a panel of 500 genetically broad chickpea accessions; For each variety, we will plant three repetitions in a greenhouse at McGill University. A low-cost multi-view photogrammetry system will be developed to scan the 60 plants at 20 developmental time points to build a high-resolution point cloud sequence dataset. Then, a novel labeling-efficient 3D deep learning network will be developed for an end-to-end instance segmentation of individual plants. We aim to use only around 0.5% points to be labeled for the segmentation model training. Also, 3D phenotypic traits at both organ and whole plant levels will be extracted. In addition, a dynamic point cloud modeling framework will be developed to characterize plant architecture spatio-temporal growth patterns. We will validate our methods and gained insights by conducting trials of the breeding and development of elite chickpea lines at the Emile A. Lods farm on the Macdonald campus of McGill University. Overall, the successful implementation of the project can accelerate plant breeding process and enhance the understanding of principles for how plants adapt to changing environments. Also, the developed point cloud dataset can help close the data gaps, enabling a broad set of new research and applications. Educationally, we will help train interdisciplinary HQP who are critically needed in Canada and globally for computational plant science.