3D segmentation of single trees exploiting full waveform LIDAR data
3D segmentation of single trees exploiting full waveform LIDAR data
复制标题
DOI：
10.1016/j.isprsjprs.2009.04.002
复制
发表时间：
2009-11
影响因子：
12.7
通讯作者：
J. Reitberger;C. Schnörr;P. Krzystek;Uwe Stilla
中科院分区：
文献类型：
--
作者：
J. Reitberger;C. Schnörr;P. Krzystek;Uwe Stilla
本文重点介绍了一种新的激光雷达单树分割方法，并比较了首末脉冲和全波形数据的分割结果。首先，建立传统的基于流域的分割方法，从激光雷达数据中稳健地插值冠层高度模型，并根据冠层高度模型的局部最大值计算出片段中最高树木的可能茎干位置。其次，将该分割方法与一种特殊的词干检测方法相结合。在流域分割的片段中，茎的位置通过树冠基部高度以下的点分层聚类来检测，并使用基于ransac的稳健茎点估计来重建茎。最后，利用归一化切割分割实现了单树的三维分割。这解决了在树冠高度模型下分割小树的问题。其关键思想是在体素空间中对树木区域进行细分，并建立由体素和体素之间的相似性度量组成的二部图。归一化切分割将图分层划分为彼此相似性最小和其成员（=体素）相似性最大的段。通过求解相应的广义特征值问题并对解向量进行适当的二值化，得到了该问题的解。实验在巴伐利亚森林国家公园进行，采用常规的首末脉冲数据和全波形激光雷达数据。第一次/最后一次脉冲数据是在TopoSys的Falcon II系统的一次飞行中以10个点/m2的点密度收集的。用Riegl LMS-Q560扫描仪在点密度为25点/m2（叶片脱落和叶片上）和点密度为10点/m2（叶片上）时捕获完整波形数据。研究结果表明，新的三维分割方法能够检测到低层森林中的小树。到目前为止，如果将基于冠层高度模型的树木分割技术应用于LIDAR数据，这实际上是不可能的。与标准分水岭分割程序相比，茎检测方法和归一化切割分割相结合的分割结果最好，在最佳情况下优于12%。此外，实验表明，使用全波形数据优于使用首末脉冲数据。
This paper highlights a novel segmentation approach for single trees from LIDAR data and compares the results acquired both from first/last pulse and full waveform data. In a first step, a conventional watershed-based segmentation procedure is set up, which robustly interpolates the canopy height model from the LIDAR data and identifies possible stem positions of the tallest trees in the segments calculated from the local maxima of the canopy height model. Secondly, this segmentation approach is combined with a special stem detection method. Stem positions in the segments of the watershed segmentation are detected by hierarchically clustering points below the crown base height and reconstructing the stems with a robust RANSAC-based estimation of the stem points. Finally, a new three-dimensional (3D) segmentation of single trees is implemented using normalized cut segmentation. This tackles the problem of segmenting small trees below the canopy height model. The key idea is to subdivide the tree area in a voxel space and to set up a bipartite graph which is formed by the voxels and similarity measures between the voxels. Normalized cut segmentation divides the graph hierarchically into segments which have a minimum similarity with each other and whose members (= voxels) have a maximum similarity. The solution is found by solving a corresponding generalized eigenvalue problem and an appropriate binarization of the solution vector. Experiments were conducted in the Bavarian Forest National Park with conventional first/last pulse data and full waveform LIDAR data. The first/last pulse data were collected in a flight with the Falcon II system from TopoSys in a leaf-on situation at a point density of 10 points/m2. Full waveform data were captured with the Riegl LMS-Q560 scanner at a point density of 25 points/m2(leaf-off and leaf-on) and at a point density of 10 points/m2(leaf-on). The study results prove that the new 3D segmentation approach is capable of detecting small trees in the lower forest layer. So far, this has been practically impossible if tree segmentation techniques based on the canopy height model were applied to LIDAR data. Compared to a standard watershed segmentation procedure, the combination of the stem detection method and normalized cut segmentation leads to the best segmentation results and is superior in the best case by 12%. Moreover, the experiments show clearly that using full waveform data is superior to using first/last pulse data.