Image-based Dynamic Quantification Of Phenotyping For Individual Plant

In the coming decades,humans will face a huge food challenge.In order to increase global food production,crop breeding programs are needed to cultivate crops with desirable traits.Efficient breeding programs needs to combine genotypes and phenotypes to select desirable plant genetic traits.However,traditional plant phenotyping measurement technology is usually carried out manually,which is time-consuming,laborious and very expensive.The method of building three-dimension(3D)plant model based on multi-view geometry algorithm has been widely used in greenhouse,and it can dynamically monitor plant growth process,but at present,this method is rarely used in complex field environment.In this study,the pepper in greenhouse and sugar beet in filed were chose as the research objects.Multi-view geometry algorithm was used to reconstruct the 3D models of pepper and sugar beet in multiple growth stages.After obtaining the high-quality 3D models of pepper and sugar beet in multiple growth stages,an automatic pipeline was developed to process the 3D point cloud of pepper and sugar beet,including preprocessing of point cloud,coordinate correction of point cloud,filtering of point cloud,segmentation of point cloud and extraction of traits(such as plant height,leaf length,leaf area,convex volume,etc.)to monitor the growth of pepper and sugar beet and explore the adaptability and difference of multi-view geometry method between greenhouse and field.In the process of processing of plant point cloud and estimation of phenotypic traits,we propose three new algorithms:in the process of point cloud segmentation,we propose an improved region growing algorithm.Firstly,region growing algorithm is used for initial segmentation,and then Euclidean clustering is used to improve the quality of leaf point cloud after original segmentation;in the process of leaf length estimation,we propose a fast extraction of leaf midrib algorithm to extract relatively flat leaves(such as pepper leaf).In addition,a more general algorithm for estimating leaf midrib is proposed to extract more complex leaf length(such as leaf of sugar beet)quickly and accurately.In the greenhouse experiment,the results showed that R~2 between the estimated value and manual measurement for leaf length and leaf area were all greater than 0.99,and RMSE were 2.55 mm and 1.4cm2,respectively.In the field experiment,the R2 between the estimated value and manual measurement for plant height and leaf length were 0.88 and 0.83 respectively,and RMSE was 5.2cm and 1.77cm respectively.The results show that the multi-view geometry method can reconstruct high-quality 3D plant model in greenhouse and field environment.Combined with the traits extraction method proposed by us,it can accurately monitor the plant growth process,provide the basis for high-throughput phenotyping analysis related to genotype,and provide some support for crop breeding.