SUPERCOMPUTING EXPERIMENTS TOWARDS MULTISCALE MACHINE-ASSISTED LANDFORM IDENTIFICATION FROM LIDAR DATA

We have undertaken a project to detect landforms for soil mapping from large lidar point cloud elevation datasets. The goals of the project are to (a) from the lidar point cloud data, create bare earth digital elevation models (DEM) at any desired project scale and with any desired footprint, and make the process tractable at large scale; (b) detect ridges and gullies, the latter of which are critical for erosion analysis but, until lidar data became available, mappable only in the field; (c) extend the analysis to the range of landforms used in soil survey. That landforms occur across a range of scales and may be superimposed is a particular challenge. For example, an area in Illinois may include dunes, lake plains, and megascale glacial lineations superimposed upon a moraine. DEM construction is based on last returns from lidar point cloud data sets using IDW interpolation. Initial experiments using the open-source PointstoGrid algorithm implemented in GRASS processed the 6440 km² Embarrass watershed of east central Illinois (976 GB of LAS data) in ~8 hr. Instrument and other noise was best removed with a series of Median and Lee filter operations. Implementation of the algorithm in C across 200 parallel processors reduced the entire processing time to 42 minutes.

For landform identification, we are experimenting with image analysis techniques first developed to differentiate ridge and valley morphology of human brains from Magnetic Resonance Imagery. The Difference of Gaussians (DoG) method generates a Gaussian Scale Space representation of the DEM by resampling it at progressively coarser grids, in our case from 3.95 ft to 101.12 ft. Each iteration is saved and subtracted from its previous (finer) scale to generate a new set of images, which represent the DoG at a given scale. Curvature across multiple scales can then be extracted by performing the processing on several of these DoG layers. The process generates ~ 1TB of data per county, including interim products. On the supercomputer, this takes 10 minutes. The next step is to interpret that curvature as multiscalar landforms.