一种基于UAV-LiDAR点云的多尺度单木分割

A Multi-Scale Method for 3D Individual Tree Extraction Using UAV-LiDAR

小型UAV-LiDAR (Unmanned Aerial Vehicle-Light Detecting and Ranging)可以快速获取林区三维数据,已经被广泛运用于林业调查。并且通过单木分割可以在激光点云中快速获取单株植被,进而获取单木几何参数。本文利用UAV-LiDAR进行数据采集,针对现有单木分割算法邻域尺度不易确定、复杂场景下单一尺度无法满足等问题设计由粗到精多尺度的单木分割策略,避免对邻域大小等参数的依赖。此外以往区域合并方法主要基于分割点云的水平投影的曲率、大小等,模糊了高程方向信息,由此本文基于点云三维分布特征设计合并准则。实验表明,本文方法在更有效的分割树冠的同时极大的抑制过分割的产生,测区内F得分达到了89%,较原方法提高7%左右。

The UAV-LiDAR (Unmanned Aerial Vehicle-Light Detecting and Ranging) can quickly produce 3D point clouds in forest industry, and has been widely used for forestry characterization to facilitate forest ecological and management studies. The geometric properties of individual tree can be easily got from the point clouds, which is handled by the method of individual tree extraction. In this paper, the original point cloud is acquired by low-cost UAV-LiDAR. Because of the problems of the existing method, such as the difficulty in determining the scale of neighborhood and the situation of complex scenes, which is hardly processed by the method with fixed scale, we design a coarse-to-fine multi-scale method. In addition, the methods of previous regional merging are based on the curvature and size of the horizontal projection of segmented point clouds, which blur the information in elevation direction. Therefore, we design a new one, based on the 3D distribution characteristics of point clouds. The experiments show that the framework identifies individual trees more precisely, and greatly inhibits the generation of over-segmentation. The result F-score reaches 89%, which is about 7% more than the original method.