Stem Detection From Terrestrial Laser Scanning Data

Forests are a critical resource in the ecological environment,playing a crucial role in mitigating climate change,maintaining ecological balance,and regulating global carbon reserves.As the smallest unit of forest resources,an individual stem serves as the foundation for obtaining various structural parameters of forest trees and is of crucial importance for forest resource investigation and monitoring.Traditional forest resource surveys are time-consuming,labor-intensive,and unable to overcome the subjectivity of human observation.Although optical remote sensing technology can be applied to largescale forest surveys,its ability to retrieve forest three-dimensional structural parameters is limited.Terrestrial Laser Scanning(TLS)technology,which has the advantage of effectively detecting the vertical structure of the forest canopy,has gradually become a new technology for forest surveying.However,the complexity of the forest and differences in point cloud density make it difficult to extract individual stems from the point cloud.This thesis focuses on the extraction method of individual stems based on Terrestrial Laser Scanning data with different levels of forest complexity,providing a theoretical basis and technical means for large-scale,high-precision forest surveying,and further enhancing the application and promotion value of TLS technology in forest surveying.The main research work and achievements of this study are as follows:(1)The voxel features that can be used to extract stem points in forest environments are developed.Point cloud classification can be used to identify stem points from forest TLS point clouds,which serve as the data foundation for stem extraction.Feature extraction and selection are critical components of the point cloud classification process,as the effectiveness of the features directly impacts classification accuracy.Therefore,this study designed 7 voxel features based on the morphological differences and spatial distribution of different classes in forest environments,and combined them with commonly used features to construct local feature descriptions of forest point cloud.Finally,the forest TLS point cloud classification was completed using a Random Forest(RF)classifier,with the stem point classification accuracy as the primary evaluation metric to analyze the importance of different features and determine the optimal feature subset.Using three different forest TLS point cloud datasets with varying levels of complexity,experimental results showed that the addition of voxel features improved stem point classification accuracy(F1)by 7.9%,13.1%,and 9.5%,respectively.(2)Individual stem extraction based on an improved mean shift clustering algorithm is proposed.As the extracted stem points from point cloud classification are still represented as a set of discrete point clouds,they are segmented into individual stem objects through a clustering algorithm.However,considering the variation in forest density and point cloud scanning density,we introduce a density-based mean shift clustering algorithm into the single stem extraction process of TLS point cloud.We optimize the clustering and filtering parameters in the algorithm through a distanceadaptive method to mitigate the impact of density difference on single stem extraction.Moreover,non-stem clusters are further eliminated through cylindrical fitting using Random Sample Consensus(RANSAC),and stem positions are calculated by combining the cylindrical model with the triangulated irregular network(TIN).The experimental results indicate that the improved mean shift clustering algorithm can better resist the interference of point cloud density difference and non-stem points.(3)Comprehensive experiments of single stem extraction.Based on the single stem extraction process and method proposed in this paper,the effect of different feature combination selection methods on single stem extraction results is analyzed by comprehensive experiments,including the feature combination selection methods based on single point classification results and single stem extraction results.Firstly,the effectiveness of voxel features in stem point extraction is verified,and secondly,the feature selection methods that take into account the single stem extraction results are investigated through feature importance measures and compared with the feature combinations used in previous studies.The experimental results demonstrate that improving the stem point extraction accuracy does not necessarily improve the single stem extraction accuracy.Compared with feature selection methods based on single point classification results,single stem extraction-oriented feature selection can achieve similar or higher single wood extraction accuracies with fewer features.The results of this paper can provide suitable and effective feature combination reference for single stem extraction of forest scenes and provide new ideas for feature combination selection.Second,compared with the feature combinations in previous studies,the feature combinations selected by this method effectively reduce the feature dimensionality and improve the single stem extraction accuracy.