Terrestrial Laser Scanning (TLS) is well suited to acquire high resolution point clouds, which can be used to derive single tree attributes for forestry applications. The processing of TLS point clouds requires the implementation of automated processing chains allowing operational data analysis in 3D. In order to obtain detailed information such as canopy structure, biomass and leaf area index and their respective changes over time the extracted branch structure of the tree is of particular interest. Based on TLS-point clouds a variety of procedures and workflows have been developed in order to extract branch parameters. Most approaches are limited by noisy data and inhomogeneous coverage within the canopy. This is mainly the case under leaf-on conditions were most branch segments are masked or under sampled.
We present an experimental framework for an automated processing chain for structuring and classifying TLS point cloud data into branches and foliage. The aim is the generation of improved input data for further sophisticated processing procedures such as skeletonization. By separating branches from leafs an improved input for branch hierarchy generation is expected. Branch extraction was done for an exemplary tree data set, which was acquired in summer (leaf-on) and winter (leaf-off) conditions respectively. First, segmentation of tree components inside of the TLS point cloud was performed. For each derived segment a corresponding structure tensor was computed, whereby eigenvectors and eigenvalues were derived. Applying a form-index describing the eigenvalue relationship for each segment, a classification into segment shapes has been done.
The quality of the extraction was tested under leaf-on and leaf-off conditions. Therefore, derived branch structures were analysed. It could be shown that major branch structures are derivable in both leaf-on and leaf-off conditions whereby small branches are more difficult to extract under leaf-on conditions. The algorithm shows good results in the reduction of noise from the data. However, large data gaps within the main branch system mainly due to occlusion in dense leaf-on canopy cannot be overcome by the current approach. Concluding, the presented approach cannot solve the problem of data gaps, but sufficiently unmasks branch structures of noisy data sets. It uses branch segments instead of point or voxel neighbourhoods for data representation and is a promising pre-processing approach for following procedures such as skeletonization and tree modelling of dense point clouds.
EARSeL European Association of Remote Sensing Laboratories, Strasbourg, France