Scalable individual tree delineation in 3D point clouds

Abstract

Manually monitoring and documenting trees is labour intensive. Lidar provides a possible solution for automatic tree-inventory generation. Existing approaches for segmenting trees from original point cloud data lack scalable and efficient methods that separate individual trees sampled by different laser-scanning systems with sufficient quality under all circumstances. In this study a new algorithm for efficient individual tree delineation from lidar point clouds is presented and validated. The proposed algorithm first resamples the points using cuboid (modified voxel) cells. Consecutively connected cells are accumulated by vertically traversing cell layers. Trees in close proximity are identified, based on a novel cell-adjacency analysis. The scalable performance of this algorithm is validated on airborne, mobile and terrestrial laser-scanning point clouds. Validation against ground truth demonstrates an improvement from 89% to 94% relative to a state-of-the-art method while computation time is similar.