【作者】 徐达丽；

【导师】 陈广胜；

【作者基本信息】 东北林业大学 ， 林业信息工程， 2023， 博士

【摘要】 树木结构参数是森林调查的重要内容,主要包括单木和林分高度、胸径、木材体积、生物量等,其中最为关键的是树高和胸径。人工测量方式通常存在高耗时、劳动密集性强、具有破坏性,而且成本高昂等问题。近年来,利用机载和地面激光雷达扫描收集森林结构的详细三维信息,进而估计重要的树木和林分属性已经逐渐成为森林资源调查的重要手段,已有多个国家使用基于多种平台的激光扫描技术进行森林调查。目前,无人机激光扫描产生的点云密度可以达到每平方米数千点的水平,实现了对林分和单木的三维结构的高质量表示。与地面激光雷达扫描相比,无人机激光扫描允许在更大的森林区域内收集数据,但在区域底部地面激光雷达数据具有更高的精细空间分辨率。因此,结合两者优势开展大范围林木结构参数自动测量方法的研究具有重要意义,可以为森林信息的智能化、自动化获取提供有效手段。本文为了解决大范围林地树木胸径和树高的自动化高精度测量问题,提出了一种使用无人机点云进行大范围林木结构参数自动测量,并使用单木地面激光雷达点云测量结果对其校准的高精度森林信息提取方法。主要完成了以下工作。(1)为了获得从点云数据测量树高所需的地面基准,本文提出了一种从森林无人机点云自动生成地面数字高程模型的方法。首先提出了一种多阈值自动迭代布模拟滤波算法,解决了由于布模拟滤波算法异常终止导致的地面点云不完整问题,实现了无需人工调参的地面点云自动化分割。在此基础上,提出了一种基于法向量空间的最低点滤波算法,对无人机高密度地面点云进行降噪,以便获取更为真实的地面点。最终结合Delaunay三角剖分实现了地面数字高程模型的自动化精确提取。(2)为了测量单木的胸径和树高,本文使用深度学习网络建立了包括单木目标区域检测、单木点云提取和单木点云补全三个部分的单木点云分割模型。单木目标区域检测阶段使用两种不同的投影策略生成森林点云鸟瞰图,并分别训练了两个具有不同侧重的Faster-RCNN网络,用于划分具有不同主干特征的单木点云鸟瞰图投影区域。在单木点云提取阶段,提出了基于体素特征的FCN网络和基于点特征的PointNet网络进行单木点云滤波,从目标投影区域对应的点云中分类出枝干点和叶点,并滤除非枝叶点,实现了多批次智能点云分割流程。在单木点云补全阶段,为解决分割出的单木点云可能存在的欠分割和过分割问题,提出了一个深度学习网络对获得的单木点云进行后处理,以进一步提高单木点云分割效果。(3)为了解决无人机和地基树木点云间的匹配问题,提出了一种树木拓扑结构特征一致性判定策略。对单木点云进行树木骨架提取,使用图论中有向根树方式表达单木结构信息,通过对构建的有向根树进行结点特征提取和匹配实现树木搜索。这使得只需要关注树木结构也就是枝干数量和拓扑关系一致性即可实现树木的匹配,有效解决多源树木点云搜索难题。为实现联合运用无人机和地基两种点云进行自动森林参数测算提供了有效方法。(4)为了实现“输入点云数据,输出测量结果”的端到端大范围森林参数测量目标,提出了一个完整的自动计算框架,将大范围森林定量结构参数计算问题分解为三个相互关联的子问题:基于无人机点云的树高和冠幅测量、基于两种单木点云的胸径测量和建立胸径回归方程估算无法直接从点云中测量的胸径。对于从无人机点云中分割得到的单木点云,如果具有可测量的主干点云则可直接获得其对应树木的树高、冠幅和胸径。对于从地面点云中分割得到的单木点云,精确计算对应树木的胸径,并用于校准从无人机点云获得的对应单木胸径,从而得到若干组完整的单木高精度定量结构参数。将这组高精度参数和其他从无人机点云测得的参数作为样本即可建立更加准确的胸径回归方程,用于估算从点云中无法直接测量胸径的树木参数。在针叶人工林和阔叶人工林两个样地上进行的独立实验证明了本文工作的有效性。更多还原

【Abstract】 The structural parameters of trees are fundamental elements in forest surveys,encompassing tree and stand heights,diameter at breast height(DBH),wood volume.Among these,tree height and DBH stand out as particularly crucial.Conventional manual measurement methods are plagued by issues such as high time consumption,labor intensiveness,destructiveness,and high costs.In recent years,the utilization of aerial and ground-based LiDAR scans to collect detailed 3D forest structure information has emerged as a crucial approach for estimating vital tree and stand attributes.Several countries have employed laser scanning technologies across diverse platforms for forest assessments.Currently,dronebased LiDAR scanning can achieve point cloud densities of thousands of points per square meter,enabling high-quality representation of stand and individual tree three-dimensional structures.While drone-based LiDAR scanning allows data collection over larger forest areas,ground-based LiDAR possesses higher fine spatial resolution at the base level.Therefore,the integration of both approaches to conduct research on automatic measurement methods for forest tree structural parameters across extensive areas holds significant importance,offering an effective means for intelligent and automated forest information acquisition.This paper proposes a method to address the challenge of automating high-precision measurements of tree DBH and height across large forested areas.It introduces a technique that uses drone point clouds for automatic measurement of forest tree structural parameters across extensive areas and calibrates these measurements using single-tree ground-based LiDAR point cloud measurements to extract high-precision forest information.The primary contributions of this study include:1.Proposing a method to automatically generate a digital terrain model from forest drone point clouds to obtain the ground reference necessary for measuring tree height.This involves an iterative bilateral filtering algorithm with multiple thresholds,resolving issues related to incomplete ground point clouds due to abnormal terminations in the filtering algorithm.It enables automated segmentation of ground point clouds without requiring manual parameter adjustments.Additionally,a lowest point filtering algorithm based on normal vector space is introduced to denoise high-density drone ground point clouds,obtaining more authentic ground points.Ultimately,combining Delaunay triangulation achieves the automated and precise extraction of digital terrain models.2.Utilizing deep learning networks to establish a single-tree point cloud segmentation model,including single-tree target area detection,single-tree point cloud extraction,and completion.The detection phase involves two distinct projection strategies to generate forest point cloud top views,training two Faster R-CNN networks with different emphases to delineate areas of single-tree point clouds with diverse trunk features.The extraction phase introduces voxel feature-based FCN networks and point feature-based PointNet networks for single-tree point cloud filtering.This process classifies branch and leaf points from the point clouds corresponding to target projection areas,filtering out non-branch and non-leaf points,enabling a multi-batch intelligent point cloud segmentation workflow.The completion phase presents a deep learning network for post-processing segmented single-tree point clouds,addressing potential under-segmentation and over-segmentation issues to enhance the effectiveness of single-tree point cloud segmentation.3.Proposing a strategy for assessing tree topology feature consistency to address the matching problem between drone and ground-based tree point clouds.This involves extracting tree skeletons from single-tree point clouds,expressing single-tree structural information in a directed root tree manner from graph theory,conducting node feature extraction,and matching based on the constructed directed root tree to achieve tree searches.This method facilitates tree matching solely by focusing on tree structure,i.e.,consistency in branch quantity and topological relationships,effectively solving the challenge of searching multiple-source tree point clouds.It offers an effective approach for jointly utilizing drone and ground-based point clouds to automatically estimate forest parameters.4.Introducing a complete automatic calculation framework aiming for an "input point cloud data,output measurement results" objective for large-scale forest parameter measurement.This framework decomposes the quantitative structural parameter calculation problem into three interrelated sub-problems:tree height and canopy width measurement based on drone point clouds,DBH measurement based on two types of single-tree point clouds,and establishing DBH regression equations to estimate DBH parameters that cannot be directly measured from point clouds.For single-tree point clouds segmented from drone point clouds with measurable trunk point clouds,tree height,canopy width,and DBH are directly obtained.For single-tree point clouds segmented from ground point clouds,precise DBH calculations are performed to calibrate corresponding single-tree DBH obtained from drone point clouds,resulting in several sets of complete,high-precision single-tree quantitative structural parameters.Leveraging these high-precision parameters and other parameters obtained from drone point clouds establishes more accurate DBH regression equations,useful for estimating tree parameters that cannot be directly measured from point clouds.Independent experiments conducted in two sample sites,a coniferous plantation and a broad-leaved plantation,validate the effectiveness of this research.更多还原