paul r. harasti ucar visiting scientist at the naval research laboratory, monterey, ca 93943-5502,...

PAUL R. HARASTIUCAR Visiting Scientist at the Naval Research Laboratory, Monterey, CA 93943-5502, USA.

(harasti@nrlmry.navy.mil)

DAVID J. SMALLEY AND MARK WEBERMIT Lincoln Laboratory, Lexington, MA, USA

CATHY J. KESSINGERNational Center for Atmospheric Research, Boulder, CO, USA

QIN XUNational Severe Storms Laboratory, Norman, OK, USA.

TED L. TSUI, JOHN COOK AND QINGYUN ZHAO  Naval Research Laboratory, Monterey, CA, USA

3.13 RADAR DATA QUALITY CONTROL FOR THE NAVAL RESEARCH

LABORATORY NOWCAST SYSTEM

Outline

1. Radar data quality control (QC) problems2. NRL radar data flowchart3. MIT LL Data Quality Assurance (DQA)4. NCAR Radar Echo Classifier (REC)5. NSSL Radar Data QC6. Principal Component Analysis (PCA) QC7. Current NRL QC approach8. Summary9. Future Work10. References

1. Radar Data Quality Problems Noisy fields (e.g., due to small Nyquist velocity)

Irregular variations due to scan mode switches

Aliased data or unsuccessfully de-aliased data

Contamination by migrating birds

Ground clutter observed during normal propagation (NP)

Ground clutter observed during anomalous propagation (AP)

Clutter caused by moving ground and airborne vehicles

Sea Clutter

Constant power function (CPF) artifacts (calibration patterns, hardware malfunctions, and sun strobes)

Clear Air/Insects (if reflectivity is assumed from precipitation)

2. N R L R adar D ata F low chart

R ad ar P rod u c t G en era tor fo r N O W C A ST- low -tilt Z and V - echo tops - N C A R S torm T racker- 2D com pos ite Z - hou rly prec ip ita tion - 3D M ultip le -rada r W inds- V A D w inds - M IT /LL Sto rm T racke r - 3D R ada r Da ta M osa ic

C O A M P S-O S ® Da ta A ssim ila tion- 3 .5D V A R w inds and the rm odynam ic re tr ieval- A D A S 3D C loud A na lys is S ystem

A d ditio na l or A lte rna tive Q ua lity C on tro l (Un d er C on s ideratio n)- M IT /LL D ata Q ua lity A ssu rance , N S S L R ada r D a ta Q C a lgo rithm ,N C A R R ada r E cho C lass ifier, P rinc ipa l Com ponen t A na lys is m e thod

N R L -D evelo p ed C lutte r R em o va l an d D e -Alias in g Alg o rith m s- C lu tte r assum ed w he re |V | < 1 .5 m /s a round each V A D c irc le if their num berexceed the theo re tica l lim it based on a unifo rm w ind assum p tion . Correspond ing Z da ta re jec ted as w e ll.- G a te -by-ga te V de -a liasing us ing B a rgen and B row n (1980 ) m e thod w ith an E nv ironm en ta l W ind Tab le R efe rence

T h resh o ld in g to R em o ve N o ise and N earb y C lu tte r - B o th Z and V Da tu m R em o ved If :- S > 10 m /s- S igna l-to -N o ise ra tio < 10 db ( if a va ilab le )- R ange < 5 km fo r rada r tilts le ss than 4 deg rees (fo r land -based radars on ly)

R aw R ad ar D a ta{ R e fle c tiv ity (Z ) , R a d ia l V e lo c ity (V ) a n d S p ec trum W id th (S )

fro m W S R -88 D (NO A A ) a n d S u p p lem e n ta l W e ath e r Ra d a r (S W R a t U S N a vy sh o re s ite s)a n d fu tu re TP S -76 (tran sp o rta b le o n U S M a rin e C orp va ns ), S P S -4 8E (o n U S Na vy sh ip s ), a n d po ss ib ly S P Y -1 (on U S N a vy sh ip s)

COAMPS ® and COAMPS-OS ® are registered trademarks of the Naval Research Laboratory (see Geiszler et al. (2004))

• Originally developed for the FAA for use with NEXRAD reflectivity data only, but currently being adapted for use with US military radar reflectivity, radial velocity, and spectral width

• Uses radar reflectivity, radial velocity and spectrum width to identify and remove CPF artifacts and AP clutter

3. MIT LL DQA Algorithm – Overview(Smalley and Bennett 2001;2002 and Smalley et al. 2003)

3.1 MIT LL DQA Algorithm - Concept

Applied in Two Sequential Stages: (1) CPF Artifact detector

(2) AP detector, if minimal CPF artifacts detected

CPF Detector: Constant power signal (reflectivity radius2)

over a sufficiently continuous portion of a single radial

Mixed CPF artifacts + weather in any single radial cannot be removed

3.1 MIT LL DQA Algorithm - Concept

AP Detector: 3-tiered approach:

(1) Identify gates with high reflectivity coincident with very small radial velocity and spectrum width for each single radial

(2)The detected AP gate is allowed to bloom radially to adjacent gates if those gates are sufficiently close to but not quite within the bounds of the basic test

(3) Scatter filter is applied over the entire tilt of data; i.e., the sufficiency of AP neighbors from (1) and (2) is assessed and like-status assigned to the target central gate within the filter

3.2 MIT LL DQA Algorithm: CPF Artifact Example from KMLB WSR-88D

Reflectivity Before QC

Reflectivity After QC

Sun strobe

3.2 MIT LL DQA Algorithm: AP Clutter Example from KAMA WSR-88D

Reflectivity Before QC

Reflectivity After QC

AP Clutter

4. NCAR REC Algorithm – Overview(Kessinger et al. 2003)

• Current algorithm tailored for NEXRAD data, and used on the WSR-88D Open Radar Product Generator system to improve radar-derived rainfall estimates and other products used by forecasters

• Developed and “truthed” using WSR-88D and NCAR S-Pol radar data

• Uses reflectivity, radial velocity and spectrum width in fuzzy logic detection algorithms to make echo-type classifications

• Four separate algorithms to detect AP Clutter, Precipitation, Insect-Clear-Air, and Sea Clutter

4.1 NCAR REC Algorithm - Concept

General schematic of the algorithms within the radar echo classifier. The steps of the process include: ingesting the base data for reflectivity (Z), radial velocity (V), and spectrum width (W), generation of features that are derived from the base data fields, use of a fuzzy logic engine to determine the initial interest output, application of the appropriate threshold (T), and the final output product for the type of radar echo being considered .

T

Final Product

INPUT

Z V W

FEATUREGENERATION

Mean MedianSTDevTextureSpinSign

FUZZY LOGIC ENGINE

w

w

w

wApply

membership functions

Apply weights

wf

w

Compute Interest Field

4.1 NCAR REC Algorithm - Concept

d)c)

b)a)

Example histogram plots of two of the feature fields used by REC. The fraction of range gates within clutter (left column) and within precipitation plus clear air return from insects (right column) are shown for a) and b) texture of the reflectivity (TDBZ; dBZ2), and c) and d) the mean velocity field (MVE; m s-1). Data were derived from one scan of the Dodge City, KS, WSR-88D at 0.5 degree elevation and the truth field. The corresponding membership functions for (a)-(c) are shown to the right (example for (c) is the median, not mean).

Clutter Precipitation plus clear air

0 45 1000

1

0

a) Texture of the Reflectivity (TDBZ)

b) Texture of the Reflectivity (TDBZ)

0 30 1000

1

0

Corresponding Membership Functions for

-50 -2.3 0 2.3 50

1

0

c) Median Radial Velocity (MDVE – similar to mean)

4.2 NCAR REC

Algorithm - Example

a)

d)c)

b)

S-Pol data from the IHOP field experiment on 16 June 2002 at 0000 UTC. Fields shown include: a) reflectivity (dBZ), b) radial velocity (m s-1) with values near zero shaded cyan, c) thresholded APDA (green) and d) thresholded PDA (gold). The red arrow in d) denotes a region of clear air return that is incorrectly classified as precipitation. The 0.0-degree elevation angle is shown. Range rings are at 30 km intervals.

5. NSSL Radar Data QC – Overview(Gong et al. 2003, Liu et al. 2003;2005, Zhang et al. 2005)

Two-part package designed for NEXRAD: 1. Gate-by-gate de-aliasing

Three-step algorithm of Gong et al. (2003)

2. Tilt-by-tilt QC Uses feature fields/QC parameters derived from reflectivity

and radial velocity, similar to NCAR REC. Thresholds for these fields are determined based on accumulated statistics (probability distribution functions) and classified for the different WSR-88D Volume Coverage Patterns (scan types)

Noise is detected using thresholds of statistics parameters Fuzzy logic of statistics parameters used to detect AP clutter

from both stationary clutter and moving vehicles Employs Bayes Conditional Probability Theorem to

determine the two probabilities that either the radar data is, or is not, contaminated by birds

5.1 NSSL Radar Data QC: 3-Step De-aliasing – Concept by Example: WSR-88D KTLX

Aliased Radial Velocity

Reference check

Dealiased Radial Velocity

Step 1: Modified (gradient) VAD winds, derived from aliased data, used as reference for first de-aliasing sweep

Step 2: Traditional VAD winds, derived after step 1, used as new reference; remaining data jump points confined to small areas

Step 3: Reference check area in Step 2 used for a continuity check from all directions around flagged areas to de-alias data within these areas

5.2 NSSL Radar Data QC:Noise and Bird Removal- ConceptExample of some of the NSSL QC Parameters:(1) Percentage of along-beam sign changes of radial velocities (SN).

(2) Along-beam standard deviation of radial velocities (STD).

(3) Percentage of along-beam perturbation radial-velocity sign changes (VSC).

(4) Mean reflectivity (MRF).

(5) Valid radial-velocity data coverage (VDC). Large SN (> 15%) and/or STD (> 3 m s-1) in (1)-(2) indicate

noisy data fields (Liu et al. 2003) Large SN, MRF and VDC in (3)-(5) indicate a high

probability of contamination by migration birds (Zhang et al. 2005)

5.2 Migrating Bird Contamination

Example: Reflectivity Mosaic

Most of the circular-shaped WSR-88D echoes shown above are migrating birds (some indicated by yellow arrows). Birds can be a very wide-spread problem, depending on the time of day and year (see http://www.npwrc.usgs.gov/resource/othrdata/migratio/migratio.htm)

5.2 NSSL Radar Data QC: Estimated PDFs for Birds

-2 0 2 4 6 8 10 12 140.00

0.05

0.10

0.15

0.20(a)

PD

F

MRF (dBZ)

p(x1|A)

p(x1|B)

10 15 20 25 30 35 40 45 50 55 60 65 700.00

0.05

0.10

0.15

0.20 p(x2|A)

p(x2|B)

PD

F

VDC (%)

(b)

20 22 24 26 28 30 32 34 36 38 40 420.00

0.05

0.10

0.15

0.20

0.25 p(x3|A)

p(x3|B)

(c)

PD

F

VSC (%)

Conditional Probabilities

A: not contamined by birds

B: contaminated by birds

x1=MRF; x2=VDC; x3=VSC

5.2 Flowchart of NSSL Real-time Migrating

Bird IdentificationRaw data

Calculate QC parameters

Bayes identification and calculate posterior probability

Night?

yes

Bird echo Next QC step

P(|xi) >0.5

yes

no

no

5.3 NSSL QC Verification:

Migrating Bird QC Statistics Verification

Parameter Type

Multi-parameter (combined MRF, VDC,

and VSC) Statistic

Hit Rate 94.6%

False Alarm Rate 37.2%

6. PCA QC – Concept (Harasti 2000, Harasti and List 2005)

(1)Create the data matrix D whose element is the radial velocity datum at the ith range gate and jth azimuth position (S2 mode).

(2) is D centered along the range coordinate.

rD

ijD

6.1 PCA QC – Concept by Example:

•Hurricane Bret (1999)

Above: WSR-88D Reflectivity in units of dBZ from Corpus Christi, TX (KCRP) with position of Brownsville, TX (KBRO) also shown. Right panel: Radial velocity from KCRP (top) and KBRO (bottom).

6.1 PCA QC – Concept by Example• PCA Synthesis of Doppler Velocity

Tr EPD kk

P and E are matrices containing the principal components and eigenvectors, respectively, of the covariance matrix of Dr.

k is determined by the “broken-stick model” approach (Jolliffe 1986); e.g., 2 < k < 10 for Bret.

For each PPI, make the following approximation for some signal-to-noise eigenvector cut-off value k:

PCA variance summary of the first PPI. There are up to 14 PPIs within the WSR-88D radar volume scans from KBRO and KCRP

KBRO Results: Difference in R2 (Square of the Multiple Correlation Coefficient) of VAD Model Fit between using PCA-edits and not using PCA-edits for all VAD Circles at all ranges within the 0.5 PPI Scan. These results show that the PCA Synthesis removes a significant amount of clutter and noise at close ranges.

6.1 PCA QC – Concept by Example•Procedure:Use the PCA Synthesis approximation of each VAD circle to edit outliers using a 2 standard deviation (SD) threshold – remove corresponding reflectivity data as well if the radar moments are not split on different PPIs.

KBRO VAD Circle Radius (km)

21 20

31 34

55 50

KCRP6.1 Radial velocity (vertical axis, ms-1) versus azimuth angle examples of PCA QC edited ground clutter and noise shown as unfilled red squares. High- order (up to ~10 wavenumbers), truncated Fourier Series Fit shown as blue line. Retained data points shown by blue-filled red squares. These results are from the study by Harasti and List (2001) who required these Fourier coefficients for their hurricane wind retrieval method.

7. Current NRL QC Approach

:

cos

30.15 for clutter

20

r

H

Uniform Wind Assumption

V

V

kttypical

kt

For all VAD circles whose fraction of Vr < 3 kt exceeds the fraction expected, reject all Vr and corresponding Z values where Vr < 3 kt

may be optionally adjusted for different VH conditions at each altitude as indicated by COAMPS-OS® SkewT data.

Vr = Radial Velocity

VH = Horizontal Wind Speed

= Horizontal Wind Direction

=Azimuth Angle of Radar Beam

1cos2

2

Fraction Expected

COAMPS ® and COAMPS_OS ® are registered trademarks of the Naval Research Laboratory

7.1 NRL Ground Clutter Removal Algorithm - Concept

7.2 NRL Radial Velocity De-aliasing Algorithm – Concept

The Bargen and Brown (1980) technique, Algorithm B, is applied gate-by-gate. A reference wind is required at the first gate.

Reference wind at each altitude provided by the Gradient Velocity Azimuth Display method (GVAD - Gao et al. (2004) )

If GVAD winds are not available due to lack of data then SkewT winds from COAMPS-OS® are utilized instead.

COAMPS ® and COAMPS_OS ® are registered trademarks of the Naval Research Laboratory

7.3 NRL-QC Example: SPY-1/TEP data onboard a US Navy ship near Jacksonville, FL

Reflectivity Before QC Reflectivity After QC

7.3 NRL-QC Example: Hurricane Charley (2004), WSR-88D, Key West, FL

Raw Radial Velocity De-aliased Radial Velocity utilizing the GVAD reference winds only.

7.3 NRL-QC Example: SWR, Point Loma, CA AP ground clutter and sea clutter

Second Trip Echo from distant mountains

Reflectivity Before QC Reflectivity After QC

7.3 NRL-QC Example: SWR, Fallon, NV

Radial Velocity Before QCNote: Badly aliased radial velocity due to low, 13.3 m/s Nyquist velocity of the SWR.

Radial Velocity After QCNote: All radial velocities were successfully dealiased. This was the first demonstration of the new NRL dealiasing software that combines both GVAD winds and COAMPS-OS® SkewT winds to form a dual- reference Environmental Wind Table for Dealiasing via the Bargen and Brown (1980) technique. Unlike VAD winds, GVAD Winds are not susceptible to radial velocity aliasing.

COAMPS ® and COAMPS_OS ® are registered trademarks of the Naval Research Laboratory

7.3 NRL-QC Example: SWR, Fallon, NV

Reflectivity Before QC Reflectivity After QC

Ground Clutter

7.4 NRL – NOWCAST Example: The US Naval Air Station, Fallon-Area (red and blue polygon regions) showing real-time surface observations, satellite data, and NRL QC radar data (composite reflectivity and VAD winds from KNFL (Fallon, NV, SWR), KRGX (Reno, NV, WSR-88D) and KLRX (Elko,NV, WSR-88D).

8. Summary

Method Ground Clutter

Precip-

itation

Insects -Clear Air

Sea Clutter

Noise Birds CPF Artifacts

MIT/LL

DQAX X

NCAR

RECX X X X

NSSL

QCX X X X

PCA

QCX X X

NRL

QCX X

The target classifications of each of the aforementioned radar data QC algorithms are summarized in the following table:

9. Future Work Prepare all the aforementioned radar data quality control algorithms for use with both WSR-88D and US Military Radar data received at NRL

Test all of the radar data quality control algorithms on a series of case studies and accumulate performance statistics according to each algorithm’s target classifications

Determine the optimal combination, in a layered sequence, of the QC algorithms (either complete, partial or no components) that optimizes the quality of the radar data for NOWCAST and COAMPS-OS®.

COAMPS ® and COAMPS_OS ® are registered trademarks of the Naval Research Laboratory

10. References

Bargen D. W., and R. C. Brown, 1980: Interactive radar velocity unfolding. Preprints, 19th Conference on Radar Meteorology, Miami Beach, FL, Amer. Meteor. Soc.,278-285.

Gao, J., K. K. Droegemeier, J. Gong, and Q. Xu. 2004: A method for retrieving mean horizontal wind profiles from single-Doppler radar observations contaminated by aliasing. Mon. Wea. Rev., 132, 1399–1409.

Geiszler, D., J. Kent, J. Strahl, J. Cook, G. Love, L. Phegley, J. Schmidt, Q. Zhao, F. Franco, L. Frost, M. Frost, D. Grant, S. Lowder, D. Martinez, and L. N. McDermid, 2004: The Navy’s on-scene weather prediction system, COAMPS-OS®. Preprints, 20th International Conference on Interactive Information and Processing Systems for Meteorology (IIPS), Oceanography, and Hydrology, Seattle, WA, Amer. Meteor. Soc., P19.1.

Gong, J., L. Wang, and Q. Xu, 2003: A three-step dealiasing method for Doppler velocity data quality control. J. Atmos. & Oceanic Technology, 20, 1738-1748.

Harasti, P. R., 2000: Hurricane properties by principal component analysis of Doppler radar data. Ph.D. dissertation, Department of Physics, University of Toronto, 162 pp.

Harasti, P. R., and R. List, 2001: The hurricane-customized extension of the VAD (HEVAD) method: Wind field estimation in the planetary boundary layer of hurricanes. Preprints, 30th Conf. on Radar Meteorology, Munich, Germany, Amer. Meteor. Soc., 463-465.

Harasti, P. R., and R. List, 2005: Principal component analysis of Doppler radar data. Part I: Geometric connections between eigenvectors and the core region of atmospheric vortices. J. Atmos. Sci. (in press).

Jolliffe, I. T., 1986: Principal Component Analysis. Springer-Verlag New York Inc., 271 pp.

Kessinger, C., S. Ellis, and J. Van Andel, 2003: The radar echo classifier: A fuzzy logic algorithm for the WSR-88D. Preprints-CD, 3rd Conference on Artificial Applications to the Environmental Science,, Long Beach, CA, Amer. Meteor. Soc.

Liu, S., P. Zhang, L. Wang, J. Gong, and Q. Xu, 2003: Problems and solutions in real-time Doppler wind retrievals. Preprints, 31th Conference on Radar Meteorology, 6–12 August 2003, Seattle, Washington, Amer. Meteor. Soc., 308-309.

10. References

Liu, S., Q. Xu, and P. Zhang, 2005: Quality control of Doppler velocities contaminated by migrating birds. Part II: Bayes identification and probability tests. Submitted to J. Atmos. Oceanic Technol. (in press).

Smalley, D.J., and B.J. Bennett, 2001: Recommended improvements to the Open RPG AP-Edit algorithm. MIT Lincoln Laboratory Wx Project Memorandum No. 43PM Wx-0081, November 2001. MIT Lincoln Laboratory, Lexington, MA. 37 pp.

Smalley, D.J., and B.J. Bennett, 2002: Using ORPG to enhance NEXRAD products to support FAA Critical Systems. Preprints, 10th Aviation, Range, and Aerospace Meteorology Conference, Portland, OR, Amer. Meteor. Soc., 3.6.

Smalley, D.J., B.J. Bennett and M. L. Pawlak, 2003: New products for the NEXRAD ORPG to support FAA Critical Systems. Preprints, 19th International Interactive Processing Systems Conference, Long Beach, CA, Amer. Meteor. Soc.,14.12.

Zhang, P., S. Liu, and Q. Xu, 2005: Quality control of Doppler velocities contaminated by migrating birds. Part I: Feature extraction and quality control parameters. J. Atmos. Oceanic Technol. (in press).

10. References