Research On The Acquisition Method Of Whole Soybean Plant Phenotype At Maturity Stage Based On Machine Vision

Phenotypic investigation of mature soybean,also known as soybean test,is a process of measuring the characteristics of mature soybean plants,pods and grains.Breeders will guide the selection of breeding materials and analyze the genetic mechanism of target traits based on these measurement results,which is an important part of soybean breeding process.Traditional soybean seed testing needs to be done manually,which is time-consuming and error-prone,and cannot achieve high-throughput measurement.It is difficult to meet the needs of modern breeding research for large-scale breeding and omics analysis.The development of computer technology and artificial intelligence algorithms has made it possible to automate and quantify the soybean seed testing process.However,people ’s initial exploration is to disassemble the whole plant.Although computer vision and deep learning algorithms can replace manual measurement of most indicators,plant disassembly work will also take a lot of time and still cannot achieve rapid quantification.In this study,based on deep learning technology,the automatic acquisition methods of stem-related phenotype,pod number-related phenotype and pod spatial distribution phenotype of soybean were studied for non-dismantled mature soybean whole plants.The following three research results were obtained :1)Automatic acquisition of stem-related phenotypes of mature soybean was realized based on deep target detection and directional search algorithm.In this study,the image of soybean plants at non-disassembly maturity stage collected manually was taken as the target.Firstly,the optimal algorithm YOLOX in target detection was selected as the detection network of stem nodes by comparison,and then the detected stem node coordinates were analyzed and judged by the proposed directional search algorithm to generate the digital skeleton of plants.Finally,the stem-related phenotypes of mature soybean plants were automatically extracted based on the digital skeleton of plants.Finally,the correlation analysis of the stem-related phenotypes such as plant height,node number,node spacing,main stem length,stem curvature and branch angle measured by the algorithm and manual measurement was carried out.The correlation coefficients R were 99.04 %,98.53 %,98.61 %,99.25 %,90.84 % and 93.91 %,respectively,which can basically replace manual measurement to complete the automatic measurement of these phenotypes.2)Based on deep learning and metric learning algorithms,the automatic acquisition of pod number and category-related phenotypes of mature soybean was realized.In this study,the front and back photos of soybean plants in non-disassembly maturity stage were taken as the target.Firstly,the optimal algorithm YOLOX in target detection was selected as the detection network of pods by comparison.The pods identified by target detection were used to form new training samples,and the optimal twin model with SE-Res Net50 as the feature extraction layer was obtained.Then,the prediction correlation of pods was increased to 96.64 % by using the complementary results of the two images.Finally,the phenotypes including the number of one-seed pod,the number of two-seed pod,the number of three-seed pod,the number of four-seed pod,the total number of pods and the total number of seeds were obtained,and the correlation coefficients of 92.65 %,95.17 %,96.90 %,94.93 %,96.64 %and 95.60 % were obtained by comparing with the real values of manual measurement,which basically met the needs of automatic measurement.3)Based on deep learning and twin network,the spatial distribution of soybean pods at mature stage was explored.In this study,the front and back photos of soybean plants in non-disassembly maturity stage were taken as the target.Firstly,by comparing various target detection algorithms,the optimal algorithm YOLOX was selected as the detection network of pod and stem nodes.Then,the SE-Siamese Network was used to integrate the information of the front and back pictures,and the correlation between the predicted value and the real value was increased to 94.14 %.Finally,based on the Euclidean distance,all pods are classified into corresponding nodes,and then a soybean digital plant that can describe the pod distribution is generated.Based on this digital plant,the spatial distribution index of pods such as the spatial distribution compactness of the whole soybean is calculated,and the spatial distribution characteristics of pods among different pod-setting habits are compared to explore the spatial distribution of pods in different varieties.The results of this study will be an important reference for breeder material selection.The method of this study realized the high-precision,automatic and high-throughput acquisition of stem correlation,pod number correlation and pod spatial distribution phenotype of soybean whole plant,and provided important support and guarantee for material selection and genetic analysis of corresponding traits in soybean breeding process.