|
publications > open file report > OFR 2007-1034 > model development approach
Initial Everglades Depth Estimation Network (EDEN) digital elevation model research and development
Model Development Approach
The development of DEMs for use in EDEN
applications, as well as other hydrologic and ecologic modeling and adaptive
management, has been an iterative process. Prior to the EDEN project a DEM for
an area beyond those of EDEN modeling (i.e., including coastal regions
influenced by tides) was produced using the ESRI ARCGIS topogrid algorithm1 (figure 5). This algorithm relies on spline interpolation that is modified to
produce a "hydrologically correct" DEM. While visually pleasing and sufficient
for regional scale analysis, this model is not suitable for the sub-regional
and finer-scale quantitative analyses envisioned for EDEN outputs. For example while
the spline approach honors the individual HAED values (i.e., spline surface
elevations are exactly those of the input points at the input point locations),
topogrid can generate false peaks and pits along regions where drastic changes
in elevations occur and channels are not supported by actual ground
measurements. Figure 6 depicts a small area in which water depth estimates have
been created using the DEM produced by topogrid. While the dendritic drainage
pattern depicted may seem plausible, it is not supported by field measurements
and suggests resolution in the data that do not exist. Also when applied to the
entire HAED dataset at once, the spline process fails to adequately represent
topographic breaks that occur along the boundaries of Water Conservation Areas
where levees, canals, and service roads interrupt the natural gradients
presumably present prior to development of the water control infrastructure.
To create a more realistic region-wide elevation
model for EDEN purposes, the elevation data were segregated by Water
Conservation Areas and National Park boundaries so that local trends could be
isolated, sub-region specific interpolation models could be developed, and
realistic breaks in elevation along sub-region boundaries could be imbedded in a
final, region-wide DEM. For each EDEN sub-area (figure 7), several surfacing
algorithms that are more conservative than topogrid (when interpolating between
known elevations) were evaluated. Outputs from these different methods were evaluated
through three approaches. First, for each sub-region 15 percent of the points were
withheld from the model development for their respective area before numerous
interpolation methods and parameters within interpolation methods were
specified using the remaining 85 percent of HAED points. The withheld points
were then used as a "check" of simulated elevation values by comparing
generated surfaces against their values. Next all HAED were included in
sub-area model development and cross-validation was applied. In this process
the software iteratively compares modeled surfaces to those of the input points
used to create the surface2. Water
depths are created by subtracting the generated DEM from water surfaces that
were interpolated from EDEN gage data; these depths are compared against estimated
depths gathered by various principal investigators (i.e., "PI data") during
field campaigns. Based on these evaluations and consideration of the utility of
other diagnostic surfaces that are created as by-products of various
interpolation processes, a "best" model is selected for each sub-area. Finally selected
sub-area models are combined to create an EDEN regional DEM. The steps used in
surface modeling can be summarized as follows:
1) Subset
the HAED by EDEN sub-areas.
2) Randomly
extract 15 percent of the HAED points from the set of observations associated
with each EDEN sub-area.
3) Create
numerous models for each sub-area using different surface interpolation
methods.
4) Compare
elevations interpolated using each method against those of the data points
withheld during model development.
5) Based
on error analysis select the "best" method for each EDEN sub-area.
6) Given
the best method selected use ALL available HAED points to generate numerous
elevation models for each sub-area by varying within-method modeling
parameters.
7) Create
depth layers for specific wet and dry days by subtracting modeled ground elevation
surfaces from water height surfaces interpolated from recorded EDEN gage data.
8) Compare
modeled water depths against field-estimated water depths.
9) Select
the best performing elevation model for each sub-area.
10) Combine the chosen
sub-area models to create one single EDEN elevation model.
11) As new HAED and field
measurements of water depth become available, return to Step 1.
|
Figure 5. A regional DEM created using the TOPOGRID
algorithm. While visually appealing and useful for region-wide analysis, this
DEM is not suitable for higher-resolution EDEN applications. |
|
|
Figure 6. Modeled water depth for a small area of Water Conservation Area 3 (indicated with a star symbol on figure 7) produced
by subtracting the DEM produced using the TOPOGRID algorithm from the water surface
elevation model created from EDEN water level data for May 27, 2004. This DEM
suggests a channel network that cannot be validated with available data. More
conservative results from other approaches are depicted in the figures 8 and 9. [larger image] |
|
|
Figure 7. A location map depicting the Greater Everglades region, EDEN DEM processing sub-regions and the location of method comparison referenced in figures 6, 8, and 9 (shown by star). [larger image] |
|
< Model Input | Results >
1 Use of
product and trade names is for illustrative and informational purposes only and
does not represent an endorsement by the U.S. Government.
2 Cross-validation only applies to techniques that do not honor the actual
elevation value at the observed point in creating the estimated surface.
|