CLASSIFICATION AND SEGMENTATION OF TERRESTRIAL LIDAR DATA FOR LITHOFACIES ANALYSIS

Terrestrial LiDAR systems (TLS) make it possible to construct three dimensional outcrop models with sub-cm accuracy. Although these geometric data are immediately useful for some applications in many cases information about the extent and distribution of lithostratigraphic units is also required. We describe a preliminary technique for semi-automatically delineating sand and gravel units in fluvial deposits using TLS data.

TLS data consist of irregularly spaced points with positions commonly defined by Cartesian coordinates relative to the scanner. Although irregularly spaced in x, y, z space, points are located on a rectangular lattice defined by angular scanner steps. By treating scans of relatively flat outcrops as multidimensional arrays with columns and rows corresponding to horizontal and vertical steps the problem becomes a more typical image processing problem.

Computer assisted methods such as live-wire segmentation can be very effective for delineating first order units more accurately and rapidly than manual digitizing, but rapidly become unwieldy as the number and complexity of units increases. Therefore we turn to a more automated method, based on differences in surface texture, for identifying inter-unit boundaries. The obvious approach of calculating morphological characteristics of the surface based on TLS geometric data is limited by the accuracy of individual point measurements. Unfortunately, accuracy is typically not fine enough to reliably distinguish between the cobble gravels and sands that were the focus of this study. Texture can, however, be estimated indirectly by examining the variability of reflection intensities across repeat scans. Due to the relationship between angle of incidence and intensity, irregular surfaces exhibit greater variability than smooth surfaces. Various clustering and segmentation algorithms can then be used to delineate units. In our test cases even simple thresholding did a reasonable job of segmenting gravels and sands. Although these initial results are promising, additional work is needed to identify optimal segmentation algorithms and establish the sensitivity of the technique to parameters that vary between scans, e.g. spot size and atmospheric conditions. Use of high-resolution photography combined with TLS data is also being pursued.