Transcript NCAR Nexrad Support Spring 98 TAC Meeting Salt Lake City, Utah

NEXRAD Data Quality
25 August 2000 Briefing
Boulder, CO
Cathy Kessinger
Scott Ellis
Joe VanAndel
Don Ferraro
Jeff Keeler
Overview
NCAR working with NOAA OSF to improve
data quality of WSR-88D
AP clutter is significant problem
Creates errors in hydrologic algorithms that
estimate rainfall from radar
Other algorithms are effected, too
Leads to errors in interpretation of base data
Very important to remove AP clutter
Slide 2
Ground clutter
due to anomalous
propagation
degrades
the performance
of rainfall estimates
from radar
Currently, it must
be detected by
operators and clutter
filters turned on
manually
Reflectivity
Radial
Velocity
Reflectivity
Precipitation
Reflectivity
AP Clutter
Automation!
Slide 3
AP Clutter Mitigation Scheme
• Automatic clutter filter control
• Radar Echo Classifier
–
–
–
–
–
Uses fuzzy logic techniques
AP Detection Algorithm (APDA)
Precipitation Detection Algorithm (PDA)
Clear Air Detection Algorithm (CADA)
other algorithms, as needed
• Reflectivity compensation of clutter filter bias
• Tracking of clutter filtered regions
Slide 4
Fuzzy logic recognition
Membership
function
Feature
fields
derived
from
base data
w1
REC outputs
Membership
function
w2
Membership
function
w3
Sum
AP clutter
Precipitation
Clear air
Bright band
Sea clutter
etc
Slide 6
Evaluation of REC
• Use statistical indices to measure
performance of algorithms against “truth”
– CSI, POD, FAR computed from
2x2 contingency table
• For NEXRAD cases, truth defined by
human experts (subjective)
• For S-Pol cases, truth defined by Particle
Identification algorithm (objective)
Slide 7
Use of S-Pol data for “truth”
• Advantages:
– Independent determination of truth using
multi-parameter data
– Objective determination of truth (no humans!)
– No temporal & spatial differences in Z,V,W
fields
– Can define ground clutter, precipitation and
clear air (from bugs) echoes
Slide 8
Hydrometeor identification with
polarimetric radar
Z
Zdr
Fdp
rhv
LDR
V,W…
Fuzzy
Fuzzy
logic
inference
engine
Rain
Snow
Hail
Graupel
Ice crystal
SC Liq Water
Clutter
Freezing Level
Slide 9
PID Algorithm
E ch o T y p e
R E C A lg o rith m
E ch o T y p e
G ro u n d clu tter
L ig h t, m o d erate P recip itatio n
& h eav y rain
D etectio n
H ail
A lg o rith m
C lo u d
A P D etectio n
A lg o rith m
C lear A ir D etectio n
A lg o rith m
N o t u sed y et
D rizzle
N o t u sed y et
S u p er-co o led
liq u id d ro p lets
F ly in g b ird s
N o t u sed y et
F ly in g in sects
N o t u sed y et
R E C A lg o rith m
R ain & h ail
m ix tu re
G rau p el &
sm all h ail
D ry an d w et
sn o w
Ice cry stals
Irreg u lar ice
cry stals
Slide 10
Use of S-Pol data for “truth”
• 11 February 1999
• AP, clear air &
precipitation
• Truth:
– green = AP
– gold =
precipitation
– red = clear air
Reflectivity
Spectrum Width
Radial Velocity
Truth
Slide 11
AP Detection Algorithm
• Features derived from base data
–
–
–
–
Mean radial velocity
Standard deviation of radial velocity
Mean spectrum width
“Texture” of the reflectivity (mean squared
difference)
– Vertical difference in reflectivity
– First 4 are computed over a local area;
vertical difference is a gate-to-gate comparison
Slide 12
1
a) M ean R ad ial V elocity
(M V E )
0
-5 0
-2 .3
0
2.3
50
1
• APDA
membership
functions
b ) M ean S p ectru m W id th
(M S W )
0
-3 0
0
3 .2
30
1
c) T extu re of S N R
(T S N R )
0
0
45
1
1000
d ) S tan d ard D eviation of
R ad ial V elocity (S D V E )
0
-3 0
0
1
0 .7
30
e) V ertical D ifferen ce of
R eflectivity (G D Z )
0
-1 0 0
-1 8
0
100
Slide 13
APDA Data Sets
• 60 scans of NEXRAD data that were
truthed by humans
• 151 scans of S-Pol data (Brazil) that were
truthed with the PID
• APDA run with 5 features shown in slide 12
Slide 14
NCAR S-Pol
Reflectivity
Radial Velocity
AP/NP Clutter,
Precipitation,
Clear air echoes
S-Pol movie loops:
June 19, 2000
AP Detection Product
Objective Truth
June 22, 2000
Precipitation
Clear Air
Figure shown and movie
loops use the 5 features shown
in slide 12 for AP clutter
AP & NP Clutter
Slide 15
Statistical Performance of the AP Detection Algorithm
S-Pol 151 Scans
Pratte and Ecoff Version
FY-99 Version
Slide 16
AP Detection Algorithm
• 2 reflectivity features added for non-Doppler region
– Both computed over a local area (max range = 430 km)
– Matthias Steiner “spin” variable
•
•
•
•
Reflectivity difference from gate to gate > threshold
Difference > 0, spin > 0; Difference <0, spin <0
Percentage of maximum possible spin changes
Sign =100 for “speckled” fields, =0 for pure gradients
– Tim O’Bannon “sign” variable
• Reflectivity difference from gate to gate
• Accumulate + or -1 depending on sign of difference
• Sign=0 for “speckled” fields, =+1 for pure gradients
– Used in KNQA movie loop (slide 18)
Slide 17
Reflectivity
Radial Velocity
AP Detection Product
Subjective Truth
NEXRAD
AP Clutter,
Precipitation,
Clear air echoes
KNQA movie loop
AP Clutter
Figure shown uses the five
features shown in slide 12 for
AP clutter
KNQA movie loop uses four
reflectivity variables and no
Doppler information for AP
clutter
Clutter Residue
Precipitation
Slide 18
Statistical Performance of the AP Detection Algorithm
NEXRAD 60 Scans
Pratte and Ecoff Version
FY-99 Version
Slide 19
APDA Summary
• Changes to membership functions and the
weighting scheme have improved results, in
general
• Better understanding is needed of the effect
on REC algorithm performance that the
radar system differences between S-Pol and
NEXRAD creates
Slide 20
Precipitation Detection Algorithm
• For FY98, three NEXRAD scans were used
to devise a preliminary algorithm
• For FY98, algorithm detected convective
regions of precipitation, not stratiform
regions
• For FY99, algorithm detects both
convective and stratiform regions
Slide 21
Precipitation Detection Algorithm
• New features and membership functions
used
– FY98 used MVE, MSW, TSNR, MDZ, GDZ
– FY99 uses SDVE, SDSW, TSNR, MDZ, GDZ
• The PDA algorithm was run on 42 scans of
S-Pol data that covered 4 cases
Slide 22
1
a) Standard Deviation of
Radial Velocity (SDVE)
0
0
0.5
50
1.0
1
• FY99 PDA
membership
functions
b) Standard Deviation of
Spectrum Width (SDSW)
0
0
2.0
50
1
c) Texture of the SNR
(TSNR)
0
0
2.5
1000
30
1
d) Mean Reflectivity
(MDZ)
0
-20
35
-5
1
80
e) Vertical Difference of
the Reflectivity (GDZ)
0
-100
-20
0
20
100
Slide 23
Reflectivity
• S-Pol scan with
strong convective
region
• CPDA does better in
stronger region of
convection
• PDA detects all the
precipitation regions
while not detecting
most of the clutter
regions
FY99 PDA
Truth
FY98 CPDA
Slide 24
APCAT Performance Curves
42 S-Pol and
60 NEXRAD
scans
Note improved
performance of
PDA vs CPDA
a) FY99 PDA
S-Pol
c) FY99 PDA
NEXRAD
b) FY98 CPDA
S-Pol
d) FY98 CPDA
NEXRAD
Slide 25
Clear Air Detection Algorithm
• 12 S-Pol scans from 1 case used to devise a
preliminary algorithm
• Features used are TVE, MSW, SDSW,
MDZ and TSNR
Slide 26
1
a) Texture of the Radial
Velocity (TVE)
0
0
10
1
• FY99 CADA
membership
functions
1000
150
b) Mean Spectrum
Width (MSW)
0
-30
0
3
10
30
1
c) Texture of the SNR
(TSNR)
0
0
70
20
1
1000
d) Mean Reflectivity
(MDZ)
0
-30
0
1
30
80
e) Standard Deviation of
Spectrum Width (SDSW)
0
0
2
4
30
Slide 27
Reflectivity
• S-Pol clear air
case with low
radial velocity
values
• Truth field shows
clutter (green),
clear air return
(red) and small
precipitation
echoes NE of
radar (gold)
Spectrum Width
Radial Velocity
Truth
Slide 28
CADA
Truth
• Results shown for
case shown on
previous slide
• CADA performs
well at detecting
the clear air and
does not detect
most of the
clutter return
• The edges of
precipitation
echoes are falsely
detected
Slide 29