The New NSSL - MDL Partnership: New Multiple- Radar/Senso

1
The New NSSL - MDL Partnership
New Multiple- Radar/Sensor Application RD for
Warning Decision Making
Gregory J. Stumpf
CIMMS / University of Oklahoma NWS
Meteorological Development Laboratory Decision
Assistance Branch Location National Severe
Storms Laboratory, Norman, OK
2
NSSLs Vision in 2003 for NWS Warning Improvement

• Support NWS Science and Technology Infusion Plan
• NSSL pitching idea to NWS to make WDSSII a
  Multi-Sensor Products Generator for AWIPS (to
  supplement ORPG)
• ORPG only produces single-radar products
• Warning test beds (at least one per region) using
  WDSSII to feed products to AWIPS
• Introduce 4D radar analysis concepts as an AWIPS
  pop-up option
• NWS Warning Decision Making team interaction
  training
• Motivate WDM team to aid with the design phase of
  new warning applications and display concepts
• Include WDSSII into WDTB Advanced Warning
  Operations Course as a high-resolution 4D radar
  base-data analysis tool

3
The First Step

• My NSSL position was moved into the NWS
  Meteorological Development Laboratory Decision
  Assistance Branch
• My new boss Dr. Stephan Smith
• My location remained at NSSL in Norman
• Act as a liaison between severe weather research
  and application development at NSSL and NWS
  warning operations program
• Develop AWIPS testbed for new remote-sensing
  technologies and new multiple-sensor warning
  applications

4
NSSLs Mission

• To enhance NOAAs capabilities to provide
  accurate and timely forecasts and warnings of
  hazardous weather events. NSSL accomplishes this
  mission, in partnership with the National Weather
  Service (NWS), through
• a balanced program of research to advance the
  understanding of weather processes
• research to improve forecasting and warning
  techniques
• development of operational applications
• and transfer of understanding, techniques, and
  applications to the NWS.
• NSSL is the sole NOAA agency responsible for the
  RD of new applications and technology to improve
  NWS severe weather warning decision making.

5
Needs Assessment
6
Needs Assessment

• More frequent algorithm updates (not at end of
  volume scan)
• Intermediate products (help to understand final
  output, understand science, build expertise)
• Better data QC (to remove false alarms outside
  storms)
• Multi-radar integration (cones-of-silence, far
  ranges, terrain blockage)
• NSE integration (automated, more frequent
  updates, better spatial resolution, for HDA,
  SCIT-RU, multi-radar integration).

7
The NWS Severe Weather Warning Challenge

• How do operational warning forecasters
  distinguish between severe and non-severe, and
  tornadic and non-tornadic thunderstorms with the
  information they have?

8
The NWS Severe Weather Warning Challenge

• To reduce the uncertainty and improve the
  accuracy of a prediction, a warning forecaster
  will integrate more information about a storm as
  viewed by other radars and other sensors

• Multiple radar data (WSR-88D, TDWR)
• Near-Storm Environment (NSE)
• Surface observations
• Upper Air data
• Lightning data
• Satellite data

• Algorithm guidance
• Trends
• Spotter reports
• Statistical knowledge of past events
• Basic understanding of storm physics

9
Severe Weather Warning Decision Making
Applications

• It makes sense that the NWS severe weather
  detection, diagnosis, and prediction tools also
  integrate multiple-sensor information!
• Multiple-sensor integration is not a new
  concept
• However, the MS concept has yet to be fully
  realized within NWS warning applications (still
  mostly single-radar based)

10
New Severe Weather Algorithm Requirements

• Objectives for new warning application
  development
• Integrate multiple-radar and multiple-sensor
  information
• No longer single-radar specific
• Must input highest resolution data in native
  format
• More accuracy in detection and diagnosis
  (oversampling - more eyes looking at storms).
• Must have rapid-update capability
• Uses virtual volume scan concept
• Better lead time (no more waiting until end of
  volume scan for guidance).
• Must be scientifically sound

11
Legacy WSR-88D Severe Weather Applications

• ORPG Algorithms SCIT, HDA, TDA, MDA, etc.
• Signature detection based on single-radar data.
• Disadvantages of single-radar algorithms
• Products generated at end of volume scan
• Only 5-6 minute updates storm evolution is fast
• Poor sampling within cone-of-silence and at far
  ranges
• Products all keyed to individual radar volume
  scan and radar domain (azimuth/range/elevation)
• No automated tuning for different near-storm
  environments

12
Many single radars provide many different answers
13
Many single radars provide many different answers
The best detection?
14
Multiple Radar Algorithms

• Storms are oversampled, especially in
  cones-of-silence and at far ranges from single
  radars.
• Outputs information in rapid intervals can be as
  fast as individual elevation scan updates using
  virtual volume scans.
• Rapid update also works in single-radar mode if
  coverage or outages dictate.
• Multiple radars and rapid update lead to more
  stable tracks and trends
• Products keyed to 4D earth-relative coordinate
  system (lat, lon, elevation, time).
• Designed to be VCP independent, and can be
  integrated with other gap-filling radar
  platforms (TDWR, ASR, PAR, SMART-R, NETRAD/CASA,
  foreign radars, commercial radars).

15
Multiple radars provide one answer
16
3D Multiple-radar grid applications

• Mosaic multiple radar data to create a 3D
  Cartesian lat/lon/ht grid.
• Uses time-weighting and power-density (distance)
  weighting schemes.
• Intelligently handles terrain blockage,
  interpolation in sparse grid cells
• Can advect older data when running a motion
  estimator.
• Run algorithms on continuously-updating 3D grids
  (virtual volumes) the data are nearly LIVE
• 3D reflectivity field for MaxRef, VIL, echo top,
  LRM, LRA, hail, Cell ID
• 3D velocity derivative fields for vortex
  (rotation) and wind shift (convergence)
  detection.
• Easy to integrate other sensor information (NSE,
  satellite, lightning, etc.) on similar grids.
• e.g., Thermodynamic info for hail diagnosis.

17
Multiple-Radar 3D Reflectivity Mosaic
18
Multiple-Radar 3D Merging

• 3D Grid information
• Create a 3D Lat-Lon-Height grid of 3D voxels
• Current resolution 0.01? x 0.01? x 1 km
• Current domain covers the entire CWAs of OUN,
  FTW, and TSA, plus a buffer
• Radars Level-II data from KTLX, KINX, KSRX,
  KVNX, KICT, KDDC, KAMA, KLBB, KFDR, KFWS, KDYX
  (later KGRK, and even later CONUS!)
• Each 3D grid voxel knows which radars are
  sensing it (this info is cached)
• If terrain blocks a radars view of a voxel, that
  radar is not used for that voxel
• Latest elevation scan of data from any radar is
  used, replacing the previous version (virtual
  volumes).

19
Multiple-Radar 3D Merging

• Data are QCed, to remove non-precipitation echoes
  (e.g., AP)
• Older data are advected forward in time using a
  motion estimator
• Data are interpolated between elevation scans
• For each radar sensing a voxel, the radar info is
  weighted based on a power-density function
  (inversely proportional to distance).
• Internal 3D grid is updating continuously, but
  new product grids are generated every 60 seconds
  (can be faster!).

20
Quality Control Neural Network (QCNN)

• Use multiple-sensor information to segregate
  precipitation echoes from non-precipitation
  echoes
• Non-precipitating clear-air return
• Ground Clutter
• Anomalous Propagation (AP)
• Chaff
• Multiple Sensor Information (two stages)
• Radar (texture statistics from all three moments,
  vertical profiles)
• Radar, satellite, and surface temperature (cloud
  cover)
• Resulting clean precipitation field used as
  input to other applications (MDA, TDA, QPE, LLSD)
• MDA and TDA false alarms are going to be a major
  issue when radars sample clear air return with
  more resolution (new VCPs, TDWR).

21
Quality Control Neural Network (QCNN)

• Uses all three radar moments, and IR satellite
  and surface data to estimate cloud cover

Original dBZ
22
Quality Control Neural Network (QCNN)

• Uses all three radar moments, and IR satellite
  and surface data to estimate cloud cover

Radar-only QCNN
23
Quality Control Neural Network (QCNN)

• Uses all three radar moments, and IR satellite
  and surface data to estimate cloud cover

Cloud Cover (Tsfc Tsat)
24
Quality Control Neural Network (QCNN)

• Uses all three radar moments, and IR satellite
  and surface data to estimate cloud cover

Multiple- sensor QCNN
Kept precip cells
Removed remaining Non-precip returns
25
Quality Control Neural Network (QCNN)
Original dBZ
26
Quality Control Neural Network (QCNN)
Quality Control Neural Network (QCNN)
27
Multi-Scale Storm Segmentation

• A novel method of performing multi-scale
  segmentation of image data (e.g., radar
  reflectivity) using statistical properties within
  the image data itself.
• The method utilizes a K-Means clustering of
  texture vectors computed within the image
  clusters are hierarchical.
• Uses, besides the actual values on the image
  grid, the distribution of values around each grid
  point.

28
2D Motion Estimation

• Uses K-means texture segmentation to extract
  multiple-scale components
• Advects multiple-scale textures
• Growth and Decay component
• Can track and trend individual multiple-scale
  textures
• 2D motion field (u, v) used to advect older data
  in 3D dBZ grid.
• This is a 60-minute loop
• 30-min actual data
• 30-min forecast

29
Multiple-Radar 3D Reflectivity Mosaic
Continuously- Updating Grid
30
Multiple-Radar 3D Reflectivity Mosaic
0120Z - 0130Z, 2002-08-14
Continuously- Updating Grid
cross section
31
Multiple-Radar 3D Reflectivity Mosaic
Continuously- Updating Grid
32
Multiple-Radar 3D Reflectivity Mosaic

• Proposed CONUS Testbed

33
Multiple-Radar 3D Reflectivity Mosaic

• Filling the cones-of-silence
• Single Radar

34
Multiple-Radar 3D Reflectivity Mosaic

• Filling the cones-of-silence
• Multiple radars

35
Multiple Radar Cell ID
36
Single Radar Cell ID

• VIL "hole" as the storm goes through the
  cone-of-silence.

Cone-Of-Silence
Single Radar KINX
37
Multiple Radar Cell ID

• No VIL "hole" as the storm goes through the
  cone-of-silence.
• The VIL maxxes out within the cone-of-silence.
• The upward trend of max VIL is only observed by
  integrating multiple-radars.
• Trend information is smoother (fewer sharp peaks
  and valleys) and is available at more rapid
  intervals (60 seconds versus 5 minutes).
• The data are nearly live.
• Cell tracking tends to be much more stable.
• Time association techniques are employed every 60
  seconds (more rapidly), instead of every 5-6
  minutes (per volume scan) where there is a
  greater likelihood of storm evolution and storm
  centroid "jumping".

Single Radar KINX Multi-Radar KINX, KICT,
KSGF, KSRX, KTLX
38
New Tools for Hail Diagnosis Using Conventional
Radar

• Taking the concept of cell-based HDA to a grid
• Integrate multi-radar and NSE information
• Provide intermediate products
• Provide other popular hail-diagnosis products

39
VIL versus SHI

• Both use vertically integrated dBZ
• dBZ profiles can come from
• A Storm cell
• A 3D grid (integrate hold (lat, lon) constant
• A 3D grid (integrate along a tilt)
• VIL integrates the entire profile, and caps dBZs
  at 56 to remove ice contamination
• Severe Hail Index (SHI) integrates only the
  profile above the melting layer, and excluded dBZ
  below 40, to include ice.

40
Severe Hail Index (SHI)

• No reflectivities below 40 dBZ are used, all
  reflectivities above 50 dBZ are used, and
  reflectivities between 40 and 50 dBZ are linearly
  weighted from 0 to 1 (a proxy to the curve shown
  in Fig. 2).
• Furthermore, only reflectivities (meeting the
  above criteria) above the melting layer are
  considered. Reflectivities between the 0?C and
  -20?C levels are weighted from 0 to 1, and all
  reflectivities (meeting the above criteria) above
  the -20?C level are considered.
• Temperature profile is made available from RUC
  00h analysis grids.
• Maximum Expected Size of Hail (MESH inches)
  0.1 (SHI)0.5

41
Cell versus Grid

• dBZ profile from cell
• Pros Follows max dBZ, inherent tilt
• Cons SCIT frequently misses detection of entire
  dBZ profile
• dBZ profile from 3D grid
• Pros There is always a complete dBZ profile,
  multiple-radar, motion estimation minimizes
  apparent tilt due to fast motion
• Cons Vertical integration may not capture storm
  tilt (BUT we are working on this issue)

42
Tilted Storm Cores

• Future Tilted Integration

VIL Vertical Integration
43
Tilted Storm Cores

• Future Tilted Integration
• Storm Cores projected to location of hail fall,
  not under echo overhangs
• Cleaner image
• Grid can replace cell-based value
• Multi-radar SCIT and/or NSE can be used to
  develop 2D grid of expected storm tilt angle

VIL Tilted Integration
44
Cell versus Grid

• Cell-based VIL or SHI
• Only one value per volume scan
• Always at end of volume scan
• Single-radar
• NSE data is sparse, and must be manually-input
• Multi-radar Grid based VIL or SHI
• Geospatial information where in cell is largest
  hail falling?
• Can accumulate grid over time for hail swaths
• Much easier for event verification (know where to
  make the probing calls).
• Storms are oversampled by multiple-radars,
  especially in cones-of-silence
• Output is essentially live (rapid update).

45
Intermediate and Popular products

• Its nice to have the final answer, but what
  ingredients went into the gridded MESH?
• dBZ relative to temperature altitudes
• Reflectivity at 0?C, Reflectivity at -20?C
• Height of 50 dBZ above -20?C altitude
• Echo tops of various dBZ thresholds
• 50 dBZ Echo Top
• Eliminates the arduous task of using all-tilts
  and data sampling, as well as mental multi-radar
  and NSE integration, to determine these values
  for each and every storm at every time
• What values correspond to what hail sizes? You
  tell us!

46
Rapidly-Updating Gridded Products from 3D Mosaic

• Shown Maximum Expected Hail Size (MEHS)
• Virtual Volume updates for each new elevation
  scan.
• Integrates NSE thermodynamic data from model
• 10-minute loop

47
Gridded Hail Products integrated with NSE data

• Easier to integrate with thermodynamic data from
  mesoscale model grids.
• Automated.
• Better spatial and temporal resolution.

48
Rapidly-Updating Gridded Products from 3D Mosaic

• Gridded data can be accumulated to give hail
  swath.
• Geo-spatial information on hail size versus a
  simple yes/no per cell.
• Geospatial info facilitates improved
  verification.

49
ExamplesMay 20, 2001
50 dBZ Echo Top
Height of 50 dBZ Above -20?C
MESH
1 km MSL Reflectivity
MESH 2hr Swath
Reflectivity at -20?C
Reflectivity at 0?C
50
Vortex Detection and Diagnosis (VDDA)

• Linear-Least Squares Derivatives (LLSD) of
  velocity
• Azimuthal and Radial Shear
• Multi-radar mosaic of 0-4 km shear
• Azimuthal Shear can be accumulated in time.

LLSD Azimuthal Shear
51
Vortex Detection and Diagnosis (VDDA)

• Modeled Rankine Vortex (Northern Hemisphere)

Radial Shear (LSD)
Simulated WSR-88D Velocity
Azimuthal Shear (LSD)
Cyclonic Shear
Divergence
Anticyclonic Shear
52
Linear Least Squares Derivative (LSSD)

• Rotational shear (us) is calculated on a local
  neighborhood surrounding each range gate (a
  range-dependent variable size mask), where
  ? sij Vij wij
• us -----------------
• ? (?sij)2 wij
• Vij is the radial velocity, sij is the azimuthal
  distance from the center of the kernel to the
  point (i,j), and wij is a uniform weight
  function.
• Because us is derived from only the radial
  component of the wind, they are approximations of
  one half the vertical vorticity (half
  vorticity, hereafter), respectively, assuming a
  symmetric wind field.

53
Multiple Radar Azimuthal Shear

• First, QCNN is run to id only precipitation
  echoes, stamp those out of the velocity field,
  and then the resulting field is dilated to
  include a small clear air buffer around storms.
• Azimuthal shear is calculated for each single
  radar (since it is radar coordinate-system
  specific) for every sample volume in the 0-3 km
  MSL layer.
• In addition, the 0.5 degree tilt is always used,
  regardless if it has an altitude above 3 km MSL.
• The maximum value in the vertical column in this
  layer is projected to a 2D polar grid.
• Using the same merging techniques as the dBZ data
  (but for a 2D grid) the azimuthal shear single
  radar grids are combined into a multi-radar grid.
• The maximum positive azimuthal (cyclonic) shear
  over a 6 hour period is plotted to produce
  Rotation Tracks.

54
Vortex Detection and Diagnosis (VDDA)

• Linear-Least Squares Derivatives (LLSD) of
  velocity
• Rotation and Divergence
• May 3 1999 Tornado Paths from shapefile
• Multi-radar mosaic

Six Hour Path of Rotational Shear
55
Rotation Track Usefulness

• Real-time
• A simple diagnostic of the radial velocity data
• Provides, in one image, information about the
  past locations and the past trend of intensity.
• Doesnt suffer from centroid matching failures
  (in 3D and 4D), threshold failures, etc, as does
  the MDA and TDA
• Post-event
• Very useful for verification first guess at
  where strongest rotation tracked send survey
  teams there.
• Eliminates need to manually replay radar data and
  track the mesos.

56
Vortex Detection and Diagnosis (VDDA)
May 9-10 2003 OKC - Six Hour Path of Rotational
Shear
57
Multiple-sensorCG Lightning Prediction

• Uses Radial Basis Function (RBF) for
  initiation, growth, decay.
• Input data include multiple radars (MR),
  lightning density, and mesoscale model analyses
• Self-training using live CG data.
• Also uses WDSSII Motion Estimator to advect
  fields
• For possible NWS Lightning Warnings

MR Composite dBZ
15-min Lightning Density
MR dBZ at -20?C
MR dBZ at 0?C
58
Summary of WDSSII AWIPS Products

• Multi-Radar/Sensor WDSSII to AWIPS Volume Browser
  Products
• MESH
• MESH 2hr Swath
• dBZ at 0C
• dBZ at -20C
• Note Volume Browser treats the grid as a grid
  of points, thus runs an OBAN which results in
  somewhat smoothed data.
• Possible Additional Grids
• 3D Lightning Mapping Array (LMA)
• Vertically-Integrated source density
• Vertically-integrated flash density
• Quantitative Precipitation Estimation and
  Segregation Using Multiple Sensors (QPESUMS)
• Instantaneous Rain Rate
• 1 hr and 24 hr accumulation

• Height of 50 dBZ above -20C
• 50 dBZ Echo Tops
• 6-hour Rotation Tracks

59
AWIPS Examples
MESH
60
AWIPS Examples
dBZ _at_ 0?C
61
AWIPS Examples
dBZ _at_ -20?C
62
Experimental Warning Application Testbeds

• First ever AWIPS development system was installed
  at NSSL
• Ideally, at least one AWIPS severe weather
  warning testbed per region
• SR Norman, plus ?
• CR Boulder, plus ?
• WR Support for hydro apps, plus ?
• ER Sterling, plus ?
• Similar in concept to WDSS and WDSSII testbeds
• Application developer staffing during severe
  weather operations
• Feedback via surveys, etc.
• Proposed Spring 05 activities
• Experimental svr wx grids in D2D (gridded hail
  and hail swath, rotation tracks).
• Proposed Spring 06 activities
• Four-dimensional Stormcell Investigator (FSI)
• More WDSSII grids in AWIPS, perhaps as part of
  SCANprocessor.

63
Warning Decision Support System Integrated
Information (WDSS-II)

• The WATADS replacement!
• Linux OS
• Available for FTP download at http//www.wdssii.or
  g
• A large number of new experimental multi-sensor
  warning applications
• Can run archive cases with Level-II data, from
  single or multiple radars, and with other sensor
  data (e.g., RUC20)
• Innovative 4D display tool for intermediate and
  final application output and case analysis (the
  WDSSII GUI, or wg)
• API support for multi-sensor severe weather and
  flash flood application development
• Peer-to-peer support via an electronic forum
  http//forum.nssl.noaa.gov

64
NSSL Forum

• Open to all in NWS and academia
• http//forum.nssl.noaa.gov
• Follow the Register link
• User name should be format First Last, and
  please include your Location when signing up
• Topics
• WDSSII Support
• Multiple-Sensor Severe Weather Applications
• Quantitative Precipitation Estimation
• 4D Base Radar Data Analysis
• Warning Decision Making Theory

65
Resources
Greg.Stumpf_at_noaa.gov https//secure.nssl.noaa.go
v/projects/feedback/ Go to NSSL Experimental
Products in AWIPS Look for links to papers
on New multi-sensor applications Multi-radar
Merging Hail Diagnosis Linear Least Squares
Derivatives