NED development included procedures to correct and/or minimize the artifacts
found in source DEM files.
DEM's produced by manual profiling techniques
were evaluated for data quality and filtered if necessary.
A "mean
profile filtering" algorithm was used to reduce the banding artifact
observed in profile DEM's. The algorithm isolates the elevation differences
that cause the banding, and then subtracts them from the DEM.
The first step is to smooth each row in the DEM independently using a large
Gaussian-weighted convolution kernel, up to 101 pixels wide. After this
operation, each column consists of a weighted mean of many columns, and
constitutes a "mean profile" in the y dimension. The banding
artifacts are caused by a systematic operator bias in operating the
profiler, and alternate sign from band to band. Since they lie along the
x-axis, they are not significantly attenuated by the smoothing operation,
and are visibly present in the y-axis mean profiles. By gently smoothing
each mean profile with a small convolution window, about the size of a band
(7 pixels works well), an array may be constructed of the residuals obtained
by subtracting each mean profile from its lightly smoothed version. The
resulting profile effectively models the banding. By subtracting these
residuals from the original DEM, a largely band-free DEM is obtained. The
magnitude of the residuals is under two meters, with an RMS of one meter
typical. The resulting data is saved and replaces the input DEM file. This
same method works well in removing patch boundary artifacts from DEM
produced by the GPM II, though it must be run in two directions. (courtesy of the USGS) |
