3d Segmentation of Full Waveform Lidar Data for Single Tree Detection Using Normalized Cut

The study highlights a novel method to segment single trees in 3D from dense airborne full waveform LIDAR data using the normalized cut segmentation. The key idea is to subdivide the tree area in a voxel space and to setup a bipartite graph which is formed by the voxels and similarity measures between the voxels. The normalized cut segmentation divides the graph hierarchically into segments which have a minimum similarity among each other and whose members (=voxels) have a maximum similarity. The solution is found by solving a corresponding generalized eigenvalue problem and an appropriate binarization of the solution vector. We applied the method to small-footprint full waveform data that have been acquired in the Bavarian Forest National Park with a mean point density of 25 points per m in leaf-off situation. The segmentation procedure is evaluated in different steps. First, a linear discriminant analysis shows that the mean intensity of the voxels derived from the full waveform data contributes significantly to the segmentation of deciduous and coniferous tree segments. Second, a sample-based sensitivity analysis examines the best value of the most important control parameter that stops the division process of the graph. Third, we show examples how the segmentation can cope with even difficult situations. We also discuss examples showing the limits of the current implementation. Finally, we present the detection rate of the new method in controlled tests using reference data. If we compare the new method to a standard watershedbased segmentation approach the overall improvement for all tree layers is 9%. However, the biggest improvement can be achieved in the intermediate layer with 14% and in the lower layer with 16% showing clearly the advantage of the new approach to a 3D segmentation of single trees. * Corresponding author.