Self-Organizing Map Clustering of Terrestrial Lidar Data within Marshes

Coastal marshes are one of the critical environments affected by sea level rise. To assess their resiliency and potential for adaptation to higher sea levels, it is important to develop techniques to accurately characterize marshes surfaces. Terrestrial laser scanning (TLS), commonly known as terrestrial lidar, is a detailed topographic approach for accurate, dense surface measurement. The dense point cloud provides a 3D representation of the surface, which includes both terrain and non-terrain objects. The extraction of topographic information requires filtering of the data into like-groups or clusters, therefore, methods must be developed to identify structure of the point cloud prior to the creation of an end product.

TLS data were collected at a marsh site along the central Texas Gulf Coast using a Riegl VZ 400 TLS. The site consists of both exposed and vegetated surface regions. To characterize the 3D structure of the TLS point cloud data, octree segmentation is first applied to create a tree data structure of voxels containing the points. Each voxel is recursively decomposed into eight child voxels until the smallest dimension of the voxels reaches a set size. For this research, a multi scale voxel approach was applied with respective sizes of z=10cm and z=30cm to capture the geometry and complexity of the terrain at the leaf scale and at a larger scale. The number of points in each voxel varies based on factors such as distance from the TLS scanner. The flexibility of the multi scale voxels makes this algorithm a promising candidate to locally characterize the geometry of the terrain. Once voxelization is applied, statistical features of the point cloud belonging to each voxel are computed: the mean and standard deviation of the elevations, the standard deviation of the standardized reflection and geometric features of voxels expressed as the first two principal components computed over the voxels’ points coordinates. The voxel features are combined with individual point characteristics: the point elevation, normalized reflectance and waveform deviation. The resulting 13 features, three point characteristics combined with 5x2 voxel based statistics at two different scales, are then used as input for clustering of the points using Self Organizing Map (SOM) clustering algorithms.

The SOM algorithm is implemented on three different platforms, a PC, a High Performance Cluster and the Google Cloud Platform. Computational efficiency is compared as well as performance with different CPU and GPU core configurations. The initial number of clusters, n=8, is based on previous results and the application of the Davies Bouldin criterion and K-means clustering. The set of eight clusters allows to identify anthropogenic (power lines) versus natural features and including distinguishing between three types of tidal flats, and different vegetation types, Black Mangrove, Low marsh transition to high marsh vegetation and Upland vegetation. Results for various SOM configurations will be presented and compared to the previous K-means results. In particular, the topological organization of the SOM clusters will be discussed. The developed method provides a novel approach for multi-scale segmentation of TLS point cloud data in marshes that can be applied to help better characterize terrain, such as for vegetation biomass studies or erosion monitoring.