The NWS program for radar mosaic development includes plans for
reflectivity mosaics to be issued at 2 km spatial resolution at 5 minute
intervals. These mosaics will also feature at least 16 reflectivity
levels, rather than the 7 levels included in the Radar Coded Message.
An unedited version of these mosaics will be made available to users.
Such mosaics are useful in identifying nonprecipitation features that are
still of meteorological significance; outflow boundaries, for example,
are often indicated by concentrations of flying insects.
We are presently refining methods of editing nonprecipitation features from
these high-resolution mosaics by automated processes. These mosaics
feature higher spatial resolution, greater precision, and a larger dynamic
range than do those based on RCM's. These characteristics enabled us to
attempt to apply feature-recognition methods in addition to the simpler
meteorological checks used for the RCM mosaic. However, the base
reflectivity products from which the HR mosaics are created contain more
nonprecipitation echoes than do the RCM's. Ground clutter appears in most
base reflectivity products, and nonprecipitation targets are represented
down to 5 dBZ, causing insect and bird returns to cover significantly more
area than they do in RCM's.
The radar images contain 16 data levels (or 'gray levels' in some
terminology). The categories are < 5 dBZ, 5-9 dBZ, 10-14 dBZ, ..., 70-74
dBZ, > 74 dBZ. Nonprecipitation features are often easily distinguished
from precipitation by their location relative to radar sites, the spectrum
of echo intensities, and texture characteristics.
Therefore, attempts were made to apply texture measures and
feature-recognition principles to the unedited mosaics, as has been done
with cloud-typing algorithms applied to satellite data. Several
researchers manually identified echo types (birds, ground clutter,
anomalous propagation, stratiform and convective precipitation) within a
series of mosaics produced between April and July 2001. Each case was
characterized by an area of 30x30 km containing no echoes other than the
dominant type. A set of 32 texture and spectrum features were then derived
from the digital image within the region, as follows:
1-11)= PERCENTAGE AREA COVERAGE BY ECHO CATEGORIES <5,5-9,10-14,...,50+
12-21)= PERCENTAGE AREA COVERAGE BY ECHO CATEGORIES 5-9,10-14,...,50+
WITHIN THE FRACTION OF THE AREA COVERED BY > 5 DBZ
22) = TEXTURE (a measure of mean box-to-box differences in echo level)
23) = MEAN ECHO LEVEL (DBZ)
24) = STANDARD DEVIATION OF ECHOES
25) = PERCENTAGE OF ECHOES GE 5 DBZ THAT ARE IN THE RANGE 5-29
26) = MEAN DIFFERENCE
27) = STANDARD DEVIATION OF GRAY-LEVEL DIFFERENCE VECTOR TEXTURE
28) = LOCAL HOMOGENEITY
29) = CONTRAST
30) = CLUSTER PROMINENCE
31) = GRAY-LEVEL DISTRIBUTION (NOT AN IMPORTANT PREDICTOR)
32) = A REGRESSION FORMULA DIFFERENTIATING
PRECIPITATION/NONPRECIPITATION, USING ITEMS 14,24 AND 27 OF THIS LIST
Linear regression equations were derived to see what combination was most
effective in determining echo type. A combination of predictors 2,6,14,24
and 27 could correctly identify echo type 95 percent of the time within the
dependent data sample.
While this method alone was very effective, it was considered more
important to leave all the rain in the image than removing all of the
non-rain echoes. Another method using the texture parameters was
developed. If either of these methods determined that 30X30 km area was
rain, the area would not be removed. A BP network was created that used
all the above parameters. While the results were similar to the
regression, it did identify some isolated rain areas that the linear
regression equations missed.
To use these methods to radar image editing, a 30X30 km floating box region
is logically moved across a radar image at 5-km intervals. The various
statistical parameters above are calculated for the area within the
floating box. If the area is considered to be a nonprecipitation echo by
both methods, a circular region of echoes is removed from the center of the
30-km box by resetting the echo level to zero.
This 'radar-only' method is fairly effective and is very simple to apply,
being based only on the information within the single radar mosaic itself.
However, enough incorrect editing results are generally detected that we
believe it adviseabladvisable ancillary information to the editing
procedure whenever possible.
A 'precipitation area' mask for the conterminous United States is now
routinely derived from information collected during operational editing of
the 10-km RCM mosaic. This information includes lightning-strike
locations, surface reports of precipitation, and the number of radars
detecting 15-dBZ reflectivity over a given place. Echoes detected by three
or more network radars (or by two or more radars during late autumn and
winter) generally indicate an echo feature with such significant vertical
development that it is almost certainly precipitation. Any combination of
lightning, surface observations, or multiple-radar detections are
considered to verify precipitation within a region of 30-km radius. These
areas are ignored during other editing procedures.
Finally, satellite and humidity information are applied to identifying
areas that are very unlikely to experience precipitation. This method was
first developed for quality control of the 10-km RCM mosaic. Satellite IR
temperatures and various other statistical predictors derived from Aviation
Model forecasts were correlated with the results of manual editing of radar
images. Through screening regression several equations were developed that
relate these predictors to the probability that a human analyst would judge
an echo area to be precipitation. When this probability is sufficiently
low (generally < 20%) echoes appearing in the unedited mosaic are removed.
|
