Classification Of Fruit Tree Components And Estimation Of Fruit Parameters Based On Hyperspectral LiDAR

To achieve precise and intelligent management of fruit trees,growers need to record information about the structure of leaves,fruits,and wood,etc,which can monitor the overall growth process of the trees.However,it very difficult to identify and monitor the tree’s components with traditional sensors for the complex spatial structure of fruit trees and interference factors,such as light and weather.Therefore,there is an urgent need for a novel sensor for fruit tree components classification and fruit parameter estimation.This thesis utilizes hyperspectral LiDAR(HSL)point cloud data collected from fruit trees to select spectral domain feature parameters.Subsequently,random forest(RF)are used as a preliminary classification method,and combined with edge point replanning optimization based on spatial distance to enhance the accuracy of the preliminary classification.This ultimately achieves the classification of fruit tree components.To address the problem of difficult estimation of fruit parameters,after the fruit tree component classification,fruit point cloud data are separated by a threshold.Affinity propagation(AP)clustering is used to segment fruit point cloud clusters.The least squares method is then used to fit individual fruit point clouds to ellipsoids model,which can accurately estimate the hanging fruit weight and fruit size.The main contributions of this thesis are as follows:(1)To classify the wood,leaves,and fruits of trees,a classification method based on edge point replanning was proposed and then 3D reconstruction of the classification results was completed.Based on the selection of nine spectral domain feature parameters,random forest,support vector machine(SVM),and backpropagation neural network(BPNN)classifiers were used to classify four types of samples(wood,leaf,unripe fruit,and ripe fruit).By analyzing the characteristics of edge misclassification points,spatial distance constraints were introduced to replan the category of edge misclassification points.The results show that after edge point re-planning,the overall classification accuracy of the samples reach 96.5%,which increase 6.55%,14.45%,and 9.4%compared to RF,SVM,and BPNN classifiers,respectively.(2)Based on the classification results of fruit trees,fruit point cloud was extracted and distinguished by label values.This study utilized spatial constraints of fruit points,based on the convex hull structure of point cloud,and applied the AP propagation clustering method to segment adhering and occluded fruits,ultimately achieving statistical analysis of fruit hanging quantity on fruit trees.The recall rate,precision rate,R~2 value,and root mean square error(RSME)for single fruit segmentation were93.33%,96.55%,94.91%,and 2.34%,respectively.(3)The single fruit points obtained by AP clustering were used as input,and the elliptical fitting was performed by the least squares method on the surface points to render the elliptical model of fruit points.The actual size of the fruit was then estimated by calculating the parameters of the ellipsoid.The results show that theR~2 value between the actual and estimated fruit sizes is 0.91,and the RSME is0.017.Figure[21]Table[7]Reference[61]...