Tree Species Identification Method Based On LiDAR Data

As the main body of renewable natural resources and terrestrial ecosystems on the earth,forests provide abundant material resources for human survival and development and play an important role in maintaining ecological balance.The spatial structure of forest is composed of many single trees.Therefore,the accurate extraction and application of single-wood information is of great significance for the investigation and management of forest resources.In addition,the extraction of single-wood canopy will contribute to tree growth monitoring,tree biomass estimation,tree pest prevention and tree species identification provides an important basis for forest resource inventory and forest above-ground carbon storage estimation.However,the current single-wood canopy extraction and tree species identification techniques mostly rely on some high-cost,time-consuming,laborious field entity surveys and manual interpretation based on low-precision aerial photographs.In recent years,with the rapid development of space technology,the advancement of remote sensing technology has enabled people to access more information about geospatial,resource information and the amount of information is becoming more abundant and the efficiency is getting higher.The research on the refinement of forest vegetation by remote sensing technology provides an effective application method for forestry resource monitoring.The data captured by the ground LiDAR system can express the detailed information of the object,which has an important impact on the geospatial structure survey.This paper takes part of the forestry land in the Xinzhuang campus of Nanjing Forestry University and the rubber experimental base in Chenzhou City,Hainan Province,China as the research area and obtains the light detection and ranging of trees by terrestrial laser scanner(TLS).The data referred to as LiDAR is the remote sensing data source.Several metsequoia,palm,sapindus,bamboo and rubber trees are selected as research objects.Single plant isolation and tree species identification research are carried out.Firstly,the multi-scale virtual grid-based filtering algorithm is used to denoise the LiDAR point cloud data to reduce the error accumulation of the later experimental results.Then,the processed LiDAR point cloud data is compared with the shortest path algorithm for single plant separation operation.Then,three kinds of effective features are proposed for the segmented single wood,68 characteristic parameters are enumerated.The SVM classifier is used to verify and calculate the optimal feature parameter set in the training data set in cross-validation,and finally the test data is based on the optimal feature parameter group.Focus on tree species classification research,the main conclusions in the experimental process are:(1)It is superior to the traditional plane-fitting filtering algorithm,slope-based filtering algorithm,irregular triangulation filtering algorithm and digital morphology filtering algorithm.The multi-scale virtual mesh filtering algorithm used in this study can effectively remove useless LiDAR data.The error points and redundant points in the point cloud data perform well in the filtering process.The final error of the filtering experiment results is 13.9%,the second type error is 8.3%,and the total error is 9.2%.The experimental results prove that the ground is based on the ground.The lidar data uses this algorithm for the effectiveness and reliability of the filtering process.(2)Based on the shortest path algorithm for single wood extraction,the recall rate is 92%,the accuracy rate is 100%,and the F value is 96%.Based on the point cloud segmentation algorithm for single wood extraction,the recall rate is 71%,the accuracy rate is 97%,and the F value is 82%.By comparing the extraction results of these two algorithms,it is found that the former can obtain better single-segmentation results in broad-leaved forests with higher canopy closure.(3)The average classification accuracy of tree species classification based on the optimal feature parameter group of tree relative clustering feature is low(45%);the average classification accuracy of tree species classification based on the optimal feature parameter group of point cloud distribution feature increases(58.8%);the average classification accuracy of tree species classification based on the optimal feature parameter group of tree apparent features is higher(63.8%);the average classification accuracy of tree species classification based on the optimal feature parameter group of integrated three types of features is the highest(82.5%).Because the morphological differences between the metasequoia and other tree species are more obvious,the performance is outstanding in the classification,and the false positive rate is the lowest(6.7%).The experimental results show that the proposed method can achieve high classification accuracy of trees,and it is proved that the method of cross-validation using SVM classifier has high feasibility and provides a powerful tool for obtaining more accurate distribution of forest tree species.