摘要
为解决传统植物表型测量存在的破坏性、主观性强以及三维测量设备成本高,使用场景受限等问题,提出了一种基于改进后的神经辐射场(neural radiance field,NeRF)方法——L-NeRF的植物表型参数测量方法。通过从多视角收集图像数据,利用L-NeRF技术进行三维重建,非破坏性地获取了植物的关键表型参数,如株高、叶宽和叶长。以墨西哥四叶草(Prosopis juliflora)为研究对象,在室内环境下进行试验,结果表明,该方法在叶宽、叶长与株高的测量上显示出较高的精确度,决定系数分别达到0.9037、0.7827和0.8516。此外,相较于现有的运动结构恢复(structure from motion,SFM)视觉重建方法,在采用L-NeRF技术时,在重建时间上平均减少了约30%,在模型精度上的标准差减少了约45%。试验结果不仅验证了L-NeRF技术在植物表型参数测量中的有效性和实用性,也展示了其在提升研究效率和精度方面的潜力。后续研究将探索收集更多植物品种的数据,以提高该方法的适用性。
Plant phenotyping has been widely used to quantitatively describe the plant's anatomical,ontogenetical,physiological,and biochemical properties.However,traditional measurement can often pose some challenges in recent years,due to the destructive nature,subjectivity,high costs,and operational complexities,particularly with three-dimensional measurement equipment.In this study,an advanced measurement approach was introduced for plant phenotypic parameters using an improved version of the Neural Radiance Field(NeRF),termed L-NeRF.Capturing image data from a multitude of perspectives and harnessing the prowess of L-NeRF technology for three-dimensional reconstruction has unlocked a new paradigm in the non-destructive acquisition of essential plant phenotypic parameters.This method has illuminated the path to obtaining critical measurements such as plant height,leaf width,and leaf length with unprecedented accuracy.The Prosopis juliflora,serving as a case study within a meticulously controlled indoor experimental environment,has yielded results that are a testament to the method's precision.The determination coefficients(R²)for leaf width,leaf length,and plant height are nothing short of remarkable,with values of 0.9037,0.7827,and 0.8516,respectively.These high R²values are a clear indication of the strong correlation between the measured parameters and their true values,highlighting the method's reliability and accuracy.In stark contrast to the traditional Structure from Motion(SFM)visual reconstruction techniques,the L-NeRF technology has demonstrated a significant acceleration in the reconstruction process,reducing the time by an average of 30%.Moreover,it has substantially improved model accuracy,evidenced by a nearly 45%reduction in the standard deviation of model precision.These enhancements are not just quantitative-they represent a qualitative leap forward in the efficiency and precision of plant phenotyping.The experimental outcomes not only validate the efficacy and practicality of L-NeRF technology in measuring plant phenotypic parameters but also highlight its potential to significantly augment research efficiency and precision.The implications of these findings are far-reaching,opening the door to future studies that will explore a wider array of plant species,thereby expanding the method's applicability and versatility.The significance of this research extends beyond the immediate context of the study.It heralds the potential to revolutionize phenotyping in plant biology.By streamlining data acquisition and enhancing the accuracy of measurements,L-NeRF technology is poised to elevate the standards of plant phenotyping research.The pursuit of a more comprehensive dataset will contribute to a deeper understanding of plant phenotypic diversity,paving the way for innovative approaches in agricultural and botanical studies.Furthermore,the application of L-NeRF technology could extend to other domains of biological research,such as zoology,medical imaging,and environmental science.Its ability to provide high-resolution,three-dimensional data sets could prove invaluable in these fields,offering new insights and possibilities for research and application.In conclusion,the advent of L-NeRF technology for 3D reconstruction in plant phenotyping is a significant milestone.It represents a paradigm shift towards more efficient and accurate data acquisition methods,which will undoubtedly contribute to the advancement of scientific knowledge and the development of innovative solutions in agriculture and beyond.The non-destructive plant phenotyping techniques have promising potential for future innovations in the field of agricultural science and technology.The finding can also offer a novel approach to transforming plant phenotypic studies into breeding programs.
作者
李添天
戴倩
谢元澄
LI Tiantian;DAI Qian;XIE Yuancheng(College of Artificial Intelligence,Nanjing Agricultural University,Nanjing 210012,China)
出处
《农业工程学报》
EI
CAS
CSCD
北大核心
2024年第17期188-196,共9页
Transactions of the Chinese Society of Agricultural Engineering
基金
国家自然科学基金资助项目(31872847)。
关键词
植物
点云
方法
植物表型组学
神经辐射场
三维重建
plants
point clouds
method
plant phenomics
neural radiance fields
3D reconstruction
