Phenotyping Of Structural Traits In Oilseed Rape Plants Based On Multiple Three-dimensional Models

Oilseed rape is an important and major oil crop and economic crop in our country.It is widely used in the production of edible oil,feed and biomass fuel.Selecting and obtaining the ideotype of oilseed rape to improve the light energy use is an important strategy for the designing and breeding of high-yield and high-density oilseed rape varieties.Currently,the genomic data in the crop breeding has shown characteristics of completeness,standardization,and scalability,while the lack of high-quality three dimension（3D）morphological structural phenotype data remains a bottleneck in plant ideotype design.Furthermore,the structure of rapeseed plants is complex and diverse.The plants exhibit a leafy vegetable-like structure at the seedling/early-bolting stages,while they display a slender stem-like structure at podding stage.In addition,the irregular arrangement and distribution of these organs present challenges in accurate,and efficient acquisition and analysis of structural phenotypic traits.Traditional plant phenotyping methods that rely on manual sampling and measurements are characterized by low efficiency,subjectivity,and limited measurement traits.These methods are short in describing the structural changes during the dynamic development process of plants and fail to meet the demands of breeding for large-scale phenotypic traits in plant ideotype design.In this study,the point cloud model and multiple topological models of complete oilseed rape plants at different growth stages were acquired and constructed.Based on multiple 3D models,plant organs were accurately segmented to realize the accurate detection of 3D structural phenotypic traits,and relationship models between these traits and plant biomass/yield were established.The main research contents and conclusions are as follows:（1）In order to overcome the difficulty of reconstructing high-quality 3D point cloud models of oilseed rape plants at different growth stages with significant morphological differences,this study developed 3D point cloud reconstruction and optimization methods for plants at different growth stages.For the plants at seedling/early-bolting stages,multi-view point clouds were collected by a time-of-flight（TOF）based camera.Processing steps including color correction,pose correction,noise reduction,and smoothing were applied,and local points’coordinates were optimized based on neighborhood mesh information of adjacent frames data.The regional part layered phenomenon was effectively eliminated,with the average distances and average angles between neighboring surface patches of local point cloud data improving by 48.33%and 55.94%respectively compared to the pre-optimization results.The high-quality 3D point cloud models were reconstructed,which were suitable for the oilseed rape plants with spiral-shaped leafy vegetable-like morphology.For the plants at podding stage,a laser scanner was used to acquire point clouds.Non-plant data were removed through a combination of height segmentation and shape fitting method,and high-quality 3D point cloud models was reconstructed,which were suitable for the oilseed rape plants with the slender stem/silique morphology.The reconstructed3D point cloud models of the seedling/early-bolting and podding stages provide a high-quality data foundation for analyzing the 3D structural phenotypic traits of oilseed rape plants.（2）In order to solve the difficulty of digital description of 3D topological structure of oilseed rape plants with complex structural at different growth stages,this study developed mesh-skeleton multimodal topology models suitable for plants with leafy vegetable-like morphology at seedling/early-bolting stages and for plants with slender stem-like morphology at podding stage.Based on the leaves point cloud data at seedling/early-bolting stages,a basic mesh using projection triangulation was built,and redundant triangle removal,hole filling,and edge smoothing were performed based on the neighborhood patches’normal and points’distribution information.Smooth 3D mesh models of leaves were constructed,accurately representing the leaf morphology.Based on the complete point cloud of plant at seedling/early-bolting stages,the skeleton points were extracted by dynamically adjusting the neighborhood calculation radius of skeleton points at different positions using the neighboring points’distribution information.Based on the complete point cloud of plant at podding stage,the skeleton points were extracted by a fitting plane constraining neighborhood calculation range.By connecting skeleton points using density clustering and weighted unidirectional graph methods,and correcting erroneous skeleton lines based on the geometric characteristics of skeleton endpoints,3D skeleton models for plants at seedling/early-bolting stages and podding stage plants was constructed.The constructed mesh-skeleton topological models provide more effective structural information for organ recognition and segmentation as well as 3D structural phenotypic parameter analysis of oilseed rape plants.（3）In order to improve the recognition and segmentation accuracy of leaves and siliques of oilseed rape plants at different growth stages,two kinds of approaches were developed in this study to achieve accurate identification and segmentation of organs of plants with complex structures.One is a geometric recognition and segmentation method based on point-cloud-skeleton multi-model geometric characteristics.And the other is an end-to-end deep learning recognition and segmentation method driven by point cloud model,reducing the dependence on the designer’s understanding of plant morphology.Based on the point-cloud-skeleton model of plant at seedling/early-bolting stages,a leaf recognition and segmentation method using the original distribution information of skeleton point neighborhoods was developed,achieving a segmentation accuracy of 96.36%with a standard deviation of 5.62%.The geometric algorithm outperformed traditional methods based solely on point clouds.For plants at podding stage,siliques were segmented using the distribution information of skeleton line lengths,which is suitable for plants with different structure types.And the overall siliques segmentation accuracy reached 90.90%with a standard deviation of 6.07%.Additionally,a dynamic voxelized sparse convolutional deep neural network,DSCU-net,was built based on the silique point cloud model to achieve accurate silique semantic segmentation.The intersection over union（Io U）was 94.13%,and the F₁ score was 96.98%.The comprehensive performance of this method surpassed four benchmark deep learning methods.The above results support the foundation for accurately analyzing the 3D structural phenotypic traits of oilseed rape（4）In order to analyze the 3D structural phenotypic traits of complex plants,structural phenotypic traits analysis methods for oilseed rape plants was developed based on the point-cloud-mesh-skeleton multi-models.The determination coefficients between the analyzed basic structural phenotype traits（such as plant height,leaf number,leaf area,silique number,silique length,silique volume,branching angle,etc.）and manually measured values were all above 0.9,validating the effectiveness and accuracy of proposed methods in analyzing 3D structural phenotypic traits.By establishing correlation models between the structural phenotypes of plants at different growth stages and their biomass and yield,the correlation coefficients were found to be 0.952 and 0.927 for total leaf area versus fresh weight and dry weight of plant at seedlings/early-bolting stages,and 0.935,0.916,and 0.897 for silique number,total length,and total volume versus yield per plant of plants at podding stage,respectively.Those results confirm the potential application of 3D structural phenotypic traits in predicting and estimating biomass and yield in oilseed rape plants.Additionally,by analyzing and describing complex phenotypic traits of canopy structure,such as the distribution parameters of the silique layer and the canopy branching divergence,new digital traits were provided for quantitative analysis of oilseed rape plant types.Above all,based on the algorithms proposed in this study,a software system called3D-Rapeseed Analyzer was designed and developed,providing key technical support for high-throughput digital analysis of plant types and crop breeding based on ideotype design of oilseed rape.