Tree species classification and estimation of stem volume and DBH based on single tree extraction by exploiting airborne full-waveform LiDAR data

The paper highlights recent results of forest structure analysis at single tree level based on analyzing airborne full waveform LiDAR data. Single trees are automatically detected by a 3D segmentation technique applied directly to laser point clouds, which uses the normalized cut segmentation combined with a stem detection method. A subsequent classification identifies tree species using salient features that are defined on single 3D tree segments and utilize the additional information extracted from the reflected laser signal by the waveform decomposition. The stem volume and diameter at breast height (DBH) are estimated by a multiple linear regression analysis which uses tree shape parameters derived from the 3D model of the trees. Experiments were conducted in the Bavarian Forest National Park with full waveform LiDAR data. The data were captured with the Riegl LMS Q-560 system at a point density of 25 points/m2 under leaf-off and leaf-on conditions. The analysis of waveform data in the tree structure shows that the intensity and pulse width discriminate stem points, crown points and ground points significantly. The unsupervised classification of deciduous and coniferous trees is in the best case 93%. If a supervised classification is applied the accuracy is slightly increased to 95%. Concerning stem volume estimation, in the case of coniferous trees the study shows a low RMSE of about 0.46 m3 to 0.43 m3 both for the watershed segmentation and the new normalized cut segmentation. In the case of deciduous trees RMSE has increased by 14% in leaf off condition and by 4% in leaf on condition for the normalized cut segmentation. A similar trend can be confirmed for DBH estimation as well, even demonstrating a larger benefit from 3D segmentation. The study results proved that the 3D segmentation approach is not only capable of detecting more small trees in the lower forest layer but also can allow to derive more promising features of single trees used for yielding better performance in species classification and estimation of forest structural parameters, especially for deciduous trees.