Hierarchical Segmentation Of Individual Trees From Side-looking LiDAR Data

Forest resource plays a significant role in ecosystem and human’s living space. Acquiring accurate tree parameters is the main job for forest research, and the prerequisite for agroforestry ecosystem study and virtual reality simulation. Side-looking light detection and ranging (LiDAR) can derive point cloud of stems, canopy surfaces, and branches inside the canopy, by means of side-looking laser scanning mechanism, which is very suitable for individual tree scale scientific research and elaborate parameters extraction. Segmenting individual tree’s point cloud from LiDAR data, is the premise to conduct such parameters extraction and scientific research, which is becoming a research priority of LiDAR technique.Side-looking LiDAR data is spatial-discrete high density point cloud in 3-D space, which owns quite large data volume and lacks topological information. A great many gaps exist in the point cloud due to occlusions. What’s more, huge shape differences can be found between tress with different species, different growing environment or different phenology periods. Therefore, it’s very difficult to automatically split individual tree points from LiDAR data. This thesis developed a hierarchical technique for segmenting individual tree points based on adaptive octree structure, which use adaptive octree leaf nodes as replacement of LiDAR points for conducting a series of segmentation procedures at regional scale, individual tree scale and canopy scale. Each LiDAR point was then labeled its tree number according to the corresponding adaptive octree leaf node, generating individual tree points. Specific research contents are as follows.(1) Adaptive octree indexing for point cloud. Side-looking LiDAR data has much higher point density (hundreds to thousands of points per square meter) compared with airborne LiDAR data, and does not have spatial connection relationship. Therefore, if the original LiDAR points are processed directly, a large amount of computer resources will be occupied by spatial searching, which decreases processing efficiency. Thus, it is essential to build a spatial index for side-looking LiDAR data. In existing point cloud indexing methods, octree has a good behavior both on index building efficiency and searching efficiency. However, lots of fake-fat nodes can be found in octree structure, which decreases nodes position accuracy. To solve this problem, an adaptive octree index is proposed, which changes node splitting rule during index construction. Index constructing experiments for both TLS and MLS data showed that adaptive octree had a relatively lower efficiency on index building, an almost the same efficiency on neighboring searching, and a higher node position accuracy compared with traditional octree structure.(2) Hierarchical technical framework for individual tree points segmentation. Aided by adaptive octree index structure, individual tree points are derived through a top-down segmenting process from regional scale to individual tree scale, and finally to canopy scale from side-looking LiDAR data. For data preprocessing, ground points and building facade points are filtered before hierarchical segmentation. At regional scale, adaptive octree is built on point cloud, and LiDAR points are clustered using leaf nodes’connectivity, in which points are organized as regional distributed point sets. At individual tree scale, adaptive octree is built again to generate nodes horizontal histogram, and tree stems are detected by analyzing local morphological characteristics of the histogram. A Voronoi diagram is then derived with stems location, with which tree core region and canopy overlapping region are labeled, and finish all segmenting procedures except overlapped canopies. At canopy scale, traditional normalized cut (Ncut) method is modified on data structure, segmenting order and node similarity calculation. The modified Ncut method are then used to partition overlapped canopy between neighboring tree, and classify the points into their corresponding trees. Thus, complete individual tree points are segmented successfully.The proposed technical framework was tested on a leaves-off terrestrial LiDAR dataset and a leaves-on mobile LiDAR dataset acquired in different environment. Experiment results showed that the framework was able to segment individual tree points from side-looking LiDAR data effectively, both reaching overall accuracy of 94%. In addition, the stem detection strategy combing histogram’ morphological characteristics and ground connectivity in individual tree scale segmentation is suitable for both high-density and low-density trees area, which broaden the application range of the framework. And the modified node similarity calculation in canopy scale segmentation improved the accuracy at the tip of branches, ensuring good results at local region.