Study On The Individual Tree Segmentation And Forest Parameters Extraction By Terrestrial Laser Scanning

Forests are the principal of terrestrial ecosystem,which is the vital guarantee to mitigate climate change and maintain ecological balance.Individual trees are the essential units of forests,and it is important to obtain forest parameters quickly and accurately for carbon stock estimation,grasp the status of forest resources and learn the dynamic change pattern.The traditional ways of collecting forest parameters mainly rely on simple measuring tools for field surveys,which are inefficient,time-consuming and limited in terms of forest parameters collected.Therefore,it is hard to adapt the needs of dynamic monitoring of forest resources in the new situation.Terrestrial laser scanning(TLS)technology can acquire high-precision and dense three-dimensional point clouds of forest trees,providing a new solution for the refined forest resource surveys.However,the complex scenes and large amount of TLS point clouds bring many challenges for carrying out data processing and forest parameters extraction.This study is conducted to investigate the data processing methods and key technologies based on TLS data in forest resource sample plot survey in terms of sample plot sampling,data acquisition,LiDAR point cloud pre-processing,individual tree segmentations,precise extraction of parameters,error analysis and parameters correction.The research results provide new approaches and theoretical methodological support to further enhance the application of TLS technology in forest resource surveys.The main research findings are as follows:(1)A 3D Normal Vector features of Voxel(NVV)based individual tree segmentations method improved the accuracy of individual tree point cloud segmentations in subtropical complex forest stands.The point cloud segmentations method of NVV proposed in the study was based on the variability of normal vector distribution features and clustering results within trunk and canopy voxels,consisting of four steps:normal vector feature calculation,voxel construction and classification,tree seed point detection,as well as trunk growth and canopy separation,achieving individual tree segmentations in forest stand point cloud through a programmed interface.Furthermore,the advantages of the NVV-based algorithm were analyzed in comparison with the classical method based on ecological basis and classical metabolic ecological theory,namely comparative shortest-path algorithm(CSP),in terms of segmentation accuracy and the single-tree point clouds details.The results showed that the recall(r)was 89.41%,the precision(p)was 93.69%and the overall accuracy(F)was 0.915,which showed a significant improvement in precision and overall accuracy compared with CSP.The accuracy of individual tree segmentation decreased with the increase of forest type complexity,and the overall accuracy F in complex natural mixed forests were less than 0.85.There still exist limitations in segmenting individual tree point clouds.(2)The TLS point cloud of single tree is able to obtain individual trees structural parameters(diameter at breast height,tree height and stem volume)with high confidence.In terms of diameter at breast height(DBH)parameter estimation,the accuracy of four typical DBH extraction methods namely the range method,convex hull algorithm,orthogonal distance regression circle fitting and least squares circle fitting algorithm,for Eucalyptus and Chinese fir species was firstly compared.In stem volume extraction,the performance of TLS point clouds in extracting the diameter of stem at different heights was demonstrated by using analytical tree of Chinese fir,constructing stem taper equations for Chinese fir and eucalyptus respectively,and calculating the stem volume with the accuracy evaluation by using the measuremental method of section with the measured data.The results showed that the extracted tree height errors were mostly at a relatively low level,with the coefficient of determination R2,root mean square error(RMSE)and relative root mean square error(RMSE%)of 0.86,1.77 m and 14.13%,respectively,and the Bias,Bias%and SD of-0.55 m,-4.39%and 1.20 m,respectively.The linear fit between the measured DBH and extracted DBH based on the least squares circle fitting method was the best,with the coefficient of determination R2,RMSE and RMSE%of 0.97,0.71 cm and 5.70%,respectively,and Bias,Bias%and SD of-0.55 m,-4.41%,and 1.20 m respectively,for the accuracy of the extracted breast diameter parameters 1.20 m.The extraction accuracy of stem volume mainly depended on the extraction accuracy of tree height and DBH,as well as the modeling accuracy of stem taper equation.(3)The differences in forest types had significant effects on individual tree point cloud segmentation and tree height extraction results.The study constructed four vegetation feature,leaf area index(LAI),canopy cover(CC),tree density(TD)and tree height variation coefficient(CVth),based on TLS data of different forest types,and analyzed the correlation between vegetation features and individual tree segmentation accuracy using multiple linear regression and Pearson correlation coefficient,as well as the influence trend of different classes of vegetation features on individual tree segmentation.In addition,the extraction accuracy of the DBH and tree height were analyzed from the differences of forest types,as well as the point cloud quality,segmentation algorithm,and stand characteristics to explore the error sources of parameter extraction.The results showed that the vegetation features had different degrees of influence on the individual tree segmentation results in different forest types,and the trends of the influence of different classes of vegetation features on the individual tree segmentation in different forest types varied greatly,and there was a negative correlation between LAI and CC and the segmentation accuracy F,and CFth was significantly correlated with the segmentation accuracy in all forest types.While the difference of forest types had more influence on the accuracy of tree height extraction,and the extraction accuracy of tree height parameters was higher in four forest types of low-density,plantation,coniferous and flat slopes,while the extraction of high-precision tree height parameters in broad-leaved forests,medium or high-density,and slope forest types needed further research.(4)The improved treetop displacement model can effectively improve the extraction accuracy of tree height.In this study,the extracted tree height was corrected by a single tree topographic slope factor for stands.The accuracy evaluation of the corrected tree height was carried out to verify the tree height correction ability of the introduced treetop displacement model.The results showed that there was a significant displacement of treetop positions before and after point cloud terrain normalization,and the displacement in the vertical direction led to the overestimation of tree height parameters extracted from trees located in steep terrain stands,and the trend was exponentially increasing with the increase of terrain slope.The modified treetop displacement model was used to correct the tree height parameters,which successfully avoided the influence of canopy angle extraction error on tree height extraction,and effectively improved the extraction accuracy of tree height based on normalized TLS point cloud.The coefficient of determination R2 between the corrected tree height parameters and the measured values was up to 0.98,and the RMSE and Bias was 0.29 m and 0.37 m,respectively.