Point Cloud Data Mining With HD Map Priors for Making Synthetic Forest Datasets

Airborne laser scanning has proven to be an indispensable tool in surveying outdoor areas due to its efficient land coverage and unimpeded access to difficult-to-reach areas. The spatial information in 3D point clouds, the data produced in airborne laser scanning surveys, can be leveraged to glean insights about the target area that would usually be unavailable in 2D-based remote sensing, such as satellite imaging. In forest point clouds, determining the locations and extents of individual trees, such as individual tree segmentation, may lend itself to forest inventory management and hazard prevention applications. However, substantial amounts of annotated and diverse data are necessary to develop classical and Machine Learning algorithms for Individual Tree Segmentation. While 3D point clouds are already notorious for how difficult it is to annotate them, high-altitude airborne laser scans make the task even more difficult due to their lower point density. This makes the structure of individual trees harder to discern, and prior information may be required to determine tree extents. This study proposes an automated approach to extracting individual trees from urban point clouds to construct synthetic datasets, combining point cloud clustering, geometrical features, and crowdsourced geospatial information of tree locations. The proposed method produces a new Individual Tree Segmentation benchmark dataset for airborne laser scanning applications.