Modern remote sensing methods such as LIDAR readily generate very
large data sets of high-resolution elevation data.  Several
applications including stream mapping, landslide risk assessment,
hydrological and erosion modeling can benefit from this
high-resolution data, but processing the data sets which can be tens
or hundreds of gigabytes in size poses a number of technical
challenges.  LIDAR point sets must be transformed into a digital
elevation model (DEM) and derived products such as a river network or
watersheds, line of sight information before users can conduct
relevant studies.

We describe our approach as a pipeline consisting of a number of
individual stages.  In the first stage we convert raw LIDAR point sets
to a digital elevation models using the spline approximation method
with substantially modified segmentation procedure to handle hundreds
of millions of points.  The constructed DEM may have some artifacts
due to sampling noise or introduced by the approximation method.  We
therefore remove from the terrain topological noise that would impede
water flow along a river network while preserving large natural
depressions or sinks such as quarries or craters.  
The next stages use the denoised DEM for constructing various derived 
data or terrain analysis tools. For example, we have developed these stages 
for computing flow network and water shade hierarchies.


We designed and implemented the pipeline mentioned above such that the
entire pipeline is scalable to large data sets.  A single non-scalable
stage in the pipeline would create a bottleneck and limit overall
scalability.  The experimental results on
real LIDAR data that show our approach is scalable to data sets
containing hundreds of million of points--over 20GB of raw data.  Our
approach allows users to go from raw data to useful high-level
information with little or no manual intervention; at the same time,
our software is highly modular and each stage can be run individually
if certain intermediate results are desired.