Point Clouds, Voxels, Meshes, and Beyond: Examples for 3D Data Processing in the Environmental Sciences

The availability of 3D data in geology, especially in geomorphology, has increased tremendously in the past years. Airborne lidar data, large-scale UAV surveys using Structure-from-Motion approaches, or virtual outcrops generated from hand-held cameras allow a much finer quantitative description of the Earth’s surface. New analysis techniques for geomorphology are actively developed to explore and take advantage of data structures and higher resolution.

 

This presentation will showcase some recent examples of data collection strategies in the field, but also segmentation and feature generation on dense point clouds and high-resolution orthomosaics. Point clouds often sample objects very densely and form nearly continuous surfaces - 3D coordinates can be converted into a network structure to form meshes, which are spatial data structures containing slope and aspect information for every facet that links three points. The view angle of lidar scanners or cameras during Structure-from-Motion processing can be used to orient normals of points and meshes. As an application example, we use curvature measured along mesh surfaces for 3D segmentation of pebbles and grains. Meshes are also data structures for measuring volumetric differences, for example between pre- and post-event data acquisitions. Textured 3D models are not yet used in the geosciences, but provide opportunities to include spectral and roughness information on meshes.

 

A current challenge in the processing of point clouds is the precise classification of points, such as distinguishing between ground and vegetation points. A common approach is to calculate geometric features in a spherical neighborhood and use these as an input to a classifier such as a random forest or a neural network. Here, we show an alternative approach for deriving point neighborhoods to calculate features for point-cloud classification: Instead of using all points in a neighborhood, we select points based on point attributes such as normal direction, color, or point connectivity. A feature-based classification based on these modified neighborhoods shows improved classification accuracy. 

 

By highlighting two approaches in point-cloud processing - turning point clouds into meshes containing network information and carefully selecting neighborhoods for point-feature calculation - we show the potential that point clouds have in geomorphologic applications.  An open research area is to further improve classifications for environmental point clouds to better monitor and quantify processes in the geosciences.