sensing forest structure from lidar
Nina Amiri is a PhD student in the department of Natural Resources. Her supervisor is prof.dr. A.K. Skidmore from the Faculty of Geo-information Science and Earth Observation.
Remote sensing technology developments increase the possibility of studying the forest structure in detail and support sustainable forest management goals. The usage of remote sensing data from Laser scanning sources has also been remarkably increased for forestry applications, since conventional field inventories are time consuming and expensive. Furthermore, remote sensing-based methods for obtaining accurate and updated forest structure have been under a continuous development. The principal goal of this thesis is to develop methods using remote sensing for obtaining explicit information on forest structure such as regeneration coverage, stem count, segmented and classified tree species. The methods are applied in small test areas and can be extended to larger forest areas. The study areas are chosen from small datasets in southeast Germany, and a small forest region in Austria. A set of indicators of forest structure including regeneration coverage, stem count by segmentation and tree species are selected. A wide range of Lidar data sources are employed, which could provide a high amount of relevant information for forestry applications.
The study is conducted across two temperate forest areas, and consisted of four case studies as follows. First, the regeneration coverage from airborne 3D point cloud using the enhanced 3D segmentation method (mean shift clustering combined with Normalized Cut) is estimated. A general framework is proposed for delineating detectable regeneration structures. To reduce the computational costs for the bipartition of the weighting matrix in Normalized Cuts, we combined the Normalized Cut algorithm with the mean shift clustering. The main advantage of a mean shift is to generate a small number of clusters to represent graph nodes instead of voxels. In the second study, features from high point density Airborne Laser Scanning data are used to reconstruct robust lines representing single tree stems. The components of the stem detection algorithm and the classifier parameters are learned from a training which is a three-step procedure at point level, segment level and object level. The outputs from classifier training are employed for modeling and generalizing single tree stem lines. In the 3rd phase a study is carried out to explore the potential of paraboloid surfaces for segmentation of single coniferous trees, where the static segmentation failed to partition the multiple tree clusters. The main aim is to significantly reduce over/under-segmentation. It can be expected if single trees are identified and characterized more precisely at object-based level by an evolutionary adaptive 3D segmentation. The applied adaptive criterion to the Normalized Cut method was thus concluded to show positive potentials towards solving over/under-segmentation issues on coniferous trees. The 4th case study is focused on the combination of features extracted from airborne multispectral Lidar and aerial imagery for detailed tree species classification. This is done through the segmentation of the 3D point cloud and later projection of the clusters onto the image plane to obtain bounding polygons for each tree crown. Spectral features are derived from pixels inside the bounding polygons. The process consisted of exploring a wide range of feature combinations including a feature selection step to optimize the feature space and to indicate the most relevant ones. Moreover, combining different structural and spectral features from multispectral Lidar yielded more accurate results than fusing multispectral aerial imagery and single wavelength Lidar data. Intensity of multispectral Lidar data (1064 nm) was the most influential feature adding up to 10% to the classification accuracy. The experimental results showed that the Lidar-based features provided the most effective information for forest structure analysis. Using the methods developed in this thesis, the approaches have the potential to be transferred to other sites.
�
