Research On Classification Method Of Plant Organs Based On Point Cloud Data

Three-dimensional(3D)Light Detection and Ranging(Li DAR)technology is a high-tech that has been continuously developed in recent years,which can quickly and accurately scan 3D point cloud data.This technology is widely used in many fields and related research.Plant point cloud data can be used for analyzing and monitoring plant structural characteristics,calculating leaf area index(LAI),estimating forest biomass,and reconstructing 3D trees models.The classification of plant organs is the basis of the above-mentioned related research,and has very important research and application value.In order to discuss the classification of stems and leaves in plant point clouds,this paper proposed automatic classification methods and experimented potted plant point cloud data and tree point cloud data for discussion.For potted plant point cloud data,this paper combines mathematical methods and support vector machine(SVM)algorithm to classify three potted plants with different shapes.A method combining three-dimensional convex hull algorithm and projection grid density was proposed to select leaf samples and stem samples as well as construct training sets automatically.In the experiment,three different potted plant point cloud data were used,and the 3D convex hull was constructed according to the spatial distribution characteristics of the potted plant point cloud data,then the turning points of the 3D convex hull were selected as the leaf sample points;in addition,combined with the density characteristics of the potted plant point cloud data,the projection grid density of different plant organs were used for selecting stem sample points automatically.Next,the spherical neighborhoods with radius of r=5mm were constructed based on the samples,and then the inside points of spherical neighborhoods were used as the training sets,which were brought into the support vector machine(SVM)classifier for training and classification.By comparing with the results of random selection method and manual selection method,the proposed method maintains a high accuracy and achieves automation of selection process.The average kappa coefficient of experiments is 0.7997.As for tree point cloud data,the paper used the ordinary least squares method for plane fitting,selected leaf points and wood points automatically by constructing the k-nearest neighbors and comparing the standard deviation of the distance from the neighbor points to the fitting plane.At the same time,this experiment calculated two new local features,which is the curvature change rate of the neighborhood points and the average distance of the k-nearest neighborhood points,and combined with the threedimensional coordinates of the point cloud data for training and classification by using SVM method.The experiments analyzed and processed 10 different sets of tree point cloud data.The results indicated that the average correct classification rate reached 0.9304,and the overall kappa coefficient was 0.7901,which meets the leaf-wood classification accuracy requirements and indicates its future application.