Terrestrial laser scanning of conifer tree crowns is challenged by occlusion problems causing sparse point clouds for many trees. Automatic segmentation of conifer tree crowns from sparse point clouds is a task that has only recently been addressed and not solved in a way that all trees can be segmented automatically without assignment errors. We developed a new segmentation algorithm that is based on region growing from seeds in voxelized 3D laser point clouds. In our data, field measured tree positions and diameters were available as input data to estimate crown cores as seeds for the region growing. In other applications, these seeds can be derived from the laser point cloud. Segmentation success was judged visually in the 3D voxel clouds for 1294 tree crowns of Norway spruce and Scots pine on 24 plots in six mixed species stands. Only about half of the tree crowns had only minor or no segmentation errors allowing to fit concentric crown models. Segmentation errors were most often caused by unsegmented neighbors at the edge of the sample plots. Wrong assignments of crown parts were also more frequent in dense groups of trees and for understory trees. For some trees, point clouds were too sparse to describe the crown. Segmentation success rates were considerably higher for dominant trees in the plot center. Despite the incomplete automatic segmentation of tree crowns, metrics describing crown size and crown shape could be derived for a large number of sample trees. A description of the irregular shape of tree crowns was not possible for most trees due to the sparse point clouds in the upper crown of most trees.
Read full abstract