Geoscience Reference
In-Depth Information
the net SNR reduction is proportional to
L
2
. Practical
L
is about 3 to 6, so the decrease is not
catastrophic considering that weather SNRs are mostly larger than 20 dB. Another issue
concerning whitening is the shape of the range weighting function compared to the matched
filter. The two weighting functions have the same range extent but the one from whitening
has rectangular shape smearing slightly its range resolution.
Increasing the number of independent samples when it is advantageous and gradually
reverting to the matched filter has also been proposed (Torres et al. 2004b) and implemented
(Curtis & Torres, 2011) on the National Weather Radar Testbed (NWRT), a phased array
radar antenna powered by a WSR-88D transmitter (Zrnic et al., 2007). The processing is
called adaptive pseudowhitening. It requires initial estimates of SNR and spectrum width.
Vivid example contrasting adaptive pseudowhitening to standard processing illustrates the
much smoother fields obtained with the former (see Fig.12, and caption). The gradient of
Doppler velocities (indicated with an arrow) is at the interface of the storms outflow and the
environmental flow. This type of discontinuity is the key feature detected by algorithms for
locating gust fronts and quantifying wind shear across the boundary; such information is
extremely useful for air traffic management and safety at airports.
In contrast to whitening techniques pulse compression does not degrade the SNR (Doviak
and Zrnic, 2006) but is not considered due to excessive bandwidth and current hardware
constraints. A very simple alternative to speed volume coverage at lowest elevations (where
tornadoes are observed) is a VCP with adaptive top elevation angle based on radar
measurements (Chrisman et al., 2009). It will soon be added to the VCPs on the network.
4.3 Clutter detection and filtering
A novel way to recognize and filter ground clutter is planned. Its acronym CLEAN-AP
stands for clutter environment analysis using adaptive processing (Warde & Torres, 2009).
The essence of the technique is spectral analysis (decomposition) of the autocorrelation at
lag 1 and use of its phase at and near zero Doppler shift. The conventional estimate
M
1
1
ˆ
2
j
2/
kM
R
(1)
Z k
( )
e
,
(10)
2
M
0
where
Z
(
k
) is the discrete Fourier transform of the returned signal, is biased (indicated by
subscript b) and can be unbiased as in (7). Another way to avoid the bias is by computing
two Fourier transform as proposed by (Warde & Torres 2009). One,
Z
0
(
k
) is the complex
spectrum of
d
(
m
)
V
(
m
),
d
(
m
)=window function, and the other
Z
1
(
k
) is the spectrum of
d
(
m
)
V
(
m
+1) from the sequence shifted in time by one unit (
T
s
). Then the unbiased estimate is
M
1
ˆ
(1)
*
2
R
Z
( )
k Z
( ) /
k
M
.
(11)
0
1
k
0
*
Individual terms
Sk ZkZ
constitute the spectral density (over Doppler index
k
) of
the lag 1 autocorrelation function. Thus the autocorrelation spectral density is estimated in
CLEAN-AP from the cross spectrum.
()
() ()
1
0
1
