Estimation Of Individual Tree Parameters Of Larch Plantations Based On UAV-LiDAR And Error-in-variable Regression

Individual tree parameters(Diameter at Breast Height(DBH),Tree Height(H)and Crown Width(CW),etc.)are background data reflecting forest spatial structure,growth health status and ecosystem stability.The accuracy of individual tree parameters estimation is very important.Unmanned Aerial Vehicle LiDAR(UAV-LiDAR)has been widely used in the research of estimating individual tree parameters because it is convenient to deploy and can accurately measure individual tree structure information based on its high density point cloud.However,the top-down scanning method cannot directly extract individual tree DBH,and in dense forests(plant density>725 trees/hm~2),due to the cover and overlap between trees,there are also large errors in the estimation of crown width parameters.Therefore,it is necessary to estimate the individual tree DBH by the tree height extracted by individual tree detection and correct the crown width extracted by individual tree detection.However,most studies ignored the error problem of independent variables(tree height and crown width extracted by individual tree detection based on UAV-LiDAR point cloud)when constructing the DBH estimation model and crown correction model based on UAV-LiDAR point cloud detection tree height and crown width.The statistical model actually violates the basic assumption that there was no measurement error in independent variables.In addition,there were few comparative studies considering different age groups combined with different Error-in-Variable(EIV)methods to perform DBH estimation and crown width correction.Therefore,in this study,Larix olgensis Henry from Maoershan Experimental Forest Farm of Northeast Forestry University was taken as an example,the UAV-LiDAR data of Maoershan Experimental Forest Farm of Northeast Forestry University and the field measured data of 13larch plantation plots in four age groups(young forest,middle-aged forest,near-mature forest and mature forest)were used as data sources,and two individual tree segmentation algorithms(Point Cloud Segmentation(PCS)and Region Hierarchical Cross-Sectional Analysis,(RHCSA))for individual tree segmentation,and extract the tree height and crown width parameters of the individual tree point cloud segmented based on the two algorithms.Compare the influence of different individual tree segmentation algorithms on the extraction of individual tree parameters,and select a group of individual tree parameters with high detection accuracy,combined with ordinary Least Squares(OLS)and three error-in-variable regression(Standard Major Axis(SMA),Ranged Major Axis,(RMA)and Maximum Likelihood Estimate(MLE)),and the artificial larch DBH estimation model and crown width correction model were constructed in the age groups(young forest,middle-aged forest,near-mature forest and mature forest),and the goodness-of-fit and prediction accuracy of the model constructed by different regression methods were evaluated,and the effects of tree height and crown width extracted by individual tree detection based on UAV-LiDAR point cloud on DBH estimation and CW correction were explored and calibrated.The results of the study showed that:(1)Among the two individual tree segmentation algorithms,the individual tree segmentation effect based on PCS algorithm is relatively good(compared with RHCSA,the overall recall rate is 0.12 higher,F-score is 0.06 higher),and more 1:1 matching trees are obtained after segmentation(PCS algorithm 1:1 matching 875 individual trees,RHCSA algorithm 1:1 matching 736 individual trees).Compared with RHCSA,the recall rate of each age group was higher by 0.06～0.17,and the F-score was higher by 0.04～0.08.(2)In terms of the extraction accuracy of individual tree parameters(tree height and crown width),the detection accuracy of individual tree height is similar between PCS algorithm and RHCSA algorithm(r RMSE difference is less than 0.3%).In terms of crown detection accuracy,the RHCSA algorithm has a higher accuracy than the PCS algorithm(the r RMSE range based on the PCS method is 24.63%～52.20%,and the r RMSE range based on the RHCSA method is:20.69%～43.81%).Therefore,in general,the precision of individual tree segmentation based on RHCSA algorithm to extract individual tree parameters is relatively high.(3)In terms of DBH estimation model prediction,the regression of the three error-in-variable methods was better than OLS,among which RMA predicted the best,compared with the OLS regression construction model,the RMSE of the four age groups inverted individual tree DBH decreased by 0.64cm～1.02cm,and the r RMSE decreased by 2.94%～4.25%.(4)In terms of crown width correction model prediction,the regression of the three error-in-variable models were all better than OLS.From the perspective of the four age groups,the RMA prediction effect was the best.Compared with the OLS regression model,the RMSE of the four age groups was reduced by 0.06m～0.16m,and the r RMSE of the four age groups was reduced by 1.44%～3.69%.When model assumptions are satisfied,error-in-variable regressions performs better than OLS in predicting response variables,and is also an ideal method to estimate unbiased model coefficients.In this study,RMA method performs the best.The artificial Larch individual tree parameter estimation model established in this study has high prediction accuracy and all errors are kept within a reasonable range,which can realize the purpose of using UAV-LiDAR to estimate individual tree parameters of large-scale forest efficiently and conveniently,and can be popularized in practice.