I was pointed to a thread
about smoothing on stormtrack.org tonight. There are a couple of points I've
written about before but want to reiterate again:
1) If you want to see what the Nexrad is reporting then you should use the
unsmoothed display. However, the unsmoothed display, aka. point filtering, is
always the worst reconstruction of
reality.
2) If you want to see the best reconstruction of reality then you should use
smoothing. The smoothed display is always
a better reconstruction of reality than the unsmoothed display.
These are mathematically provable facts.
Unsmoothed displays suffer from the highest amount of aliasing. For example, an
unsmoothed bin of 60dbz will show as a 1km long area of purple, regardless of
the dbz's in the surrounding bins. If the surrounding bins were near 60dbz, that
would be fine. However, if the surrounding bins were 40 dbz then a more accurate
reconstruction of reality would be for the purple area to be *much* smaller than
1km. Smoothing accomplishes this.
GRS currently uses bilinear filtering, which is only one step above point
filtering in terms of quality. Bilinear filtering does not "blur" the data in
any way.
One final clarification: I should have used "interpolation" from the beginning
instead of "smoothing". In technical terms, GRS apps do not "smooth" the data,
they interpolate between sampled data values. In future apps, this distinction
will become more apparent as we go from simple bilinear filtering to
higher-order filtering.
Here's a page showing some
examples of different interpolation techniques in action:
http://photoenlargement.imagener.com/
where "nearest neighbor" is the same as unsmoothed. Note that the biggest
bang-for-your-buck comes from bilinear interpolation. Bicubic interpolation adds
more fine details but its main purpose is to reduce the bilinear interpolation
artifacts. Bilinear artifacts are those prism-like distortions on curved areas
which make them jagged. You can see them on the top of the cat's right eye in
the middle image. These are due to high frequencies in the spatial data
interacting with the bilinear transform.
And a final note about velocity. Velocity displays
are difficult because you're trying to display a vector, a linear magnitude with
a point-sampled direction, with single dot of color. Typically, velocity color
tables do this by displaying the direction as one of two fully saturated hues
(red and green) with the magnitude as the lightness of the hue.
GRS attempts to smooth the velocity by point interpolating the direction and
linearly interpolating in magnitude. This was only partially successful. Higher
order filtering on a standard color may be more successful.
Another approach would be to combine velocity with reflectivity into single 3d
display. Velocity would be a height field of positive and negative values with
reflectivity as its texture. Of course, background maps and other things would
no longer work properly.
Mike - Gibson Ridge Software