Research On Soybean Phenotype Acquisition Method Based On Deep Learning And Three-dimensional Reconstruction

The external expression of soybean phenotype and genotype is an important indicator for the selection of new soybean varieties with high yield,high quality and multi-resistance.In the process of crop breeding,phenotypic extraction and analysis of a large number of candidate materials is required,which is an important part of the breeding process.In recent years,with the development of high-throughput sequencing technology,genomics research has made great progress and accumulated massive data.However,crop phenotyping is still at the stage of relying on traditional measurement methods,which are labor-intensive and time-consuming and have low precision.The lack of high-throughput phenotyping techniques results in a severe mismatch between phenotype and genomic data.More importantly,the subjectivity of the observer makes the phenotype measurement lose its objectivity.It can be seen that phenotype acquisition has become a bottleneck that needs to be broken through in the current process of rapid breeding and intelligent breeding.The development of deep learning and three-dimensional reconstruction technology has brought opportunities for automatic phenotype extraction.Designing and developing a model system for soybean phenotype precision,throughput and automatic extraction is the key to breaking through this bottleneck.Based on deep learning and three-dimensional reconstruction technology,this paper developed a new model for the automatic acquisition of soybean phenotypes in the growth and maturity stages of soybean,aiming at the morphological characteristics of various organs in various stages of soybean,and made a preliminary application of soybean flux phenotype technology explore.The main research contents and achievements of the paper are as follows:(1)For the acquisition of soybean phenotype at maturity,a deep learning method for soybean phenotype measurement from two-dimensional machine vision(SPM-IS)is proposed in this paper.At the same time,an algorithm for precise positioning of the bounding rectangle of irregular soybean organs was proposed and integrated into the phenotype measurement model.The experimental results show that after 60,000 iterations,the maximum average precision(m AP)of Mask and Box can reach 95.7%.The correlation coefficients R2 of pod length,pod width,stem length,intact main stem length,grain length and grain width measured by manual measurement and SPM-IS were0.9755,0.9872,0.9692,0.9803,0.9656 and 0.9716,respectively.The correlation coefficients R2 between manual counts of pods,stems and grains and SPM-IS counts were 0.9733,0.9872 and0.9851,respectively.The results show that SPM-IS is a robust measurement and counting algorithm that can reduce labor intensity,increase efficiency,and speed up the soybean breeding process.(2)For soybean plants in the growing period,this study reconstructed the 3D model of soybean in the whole growth period based on multi-eye vision technology,and designed the Soy3D-Net deep network to realize the segmentation and phenotype extraction of 3D soybean virtual plants.Aiming at the bending problems of leaves and stems in the natural growth state of soybean,the Flattening algorithm is introduced,which effectively solves the problem of inaccurate measurement.The correlation coefficients R2 of plant height,main stem,main stem node,leaf width,leaf length,and leaf area of soybean plants measured by hand and Soy3D-Net were calculated to be 0.9562,0.9417,0.9509,0.9229,0.9308,and 0.9312,respectively.The results show that the combination of 3D reconstruction and Soy3D-Net is an effective and accurate method for automated nondestructive measurement of soybean growth phase phenotypes.(3)Soybean introduction line is an important material in soybean breeding and an important source for variety breeding and special material selection.In this study,the SPM-IS phenotype extraction algorithm was used to analyze the introduction of Suinong 14 and wild bean ZYD00006.The stem images of mature plants were used for identification and phenotype extraction,and phenotypes such as main stem nodes,main stem length and node number were obtained.Based on these phenotypic data,12 plants with specific phenotype were screened by cluster analysis algorithm.At the same time,a new phenotype of stem and leaf curvature was proposed to evaluate the curvature of the main stem and leaves of the plant.Through the analysis of the tested varieties,it was found that the curvature of the stem can reflect the strength of the stem.A new phenotype is associated with crop yield and lodging.In this paper,based on deep learning and three-dimensional reconstruction technology,a set of solutions for automatic extraction of soybean maturity and growth phenotypes are provided,which preliminarily realizes the flux and precision extraction of soybean maturity and growth phenotypes.Phenotyping research provides new ideas and provides important technical support for rapid soybean breeding and intelligent breeding.