Embed Size (px)
Citation preview
Huan Meng1, Wes Adkins2
Jun Dong3, Cezar Kongoli3, Ralph Ferraro1, Banghua Yan1, Limin Zhao4
1NOAA/NESDIS/Center for Satellite Applications and Research2NWS/Juneau, AK Weather Forecast Office
3University of Maryland/ESSIC/Cooperative Institute for Climate and Satellites4NOAA/NESDIS/Office of Satellite and Product Operations
NESDIS Snowfall Rate Product and Assessment
Virtual Alaska Weather Symposium, February 27, [email protected], [email protected]
Product Overview• The NESDIS Snowfall Rate (SFR) product is water
equivalent snowfall rate estimate over global land
ü Passive microwave sensors: ATMS, AMSU/MHS, GMI, SSMIS
ü Eight polar-orbiting satellites: JPSS, NOAA POES, EUMETSAT Metop, NASA GPM, DMSP
• Microwave signals can penetrate clouds, and can infer snowfall rates from the precipitating layer.
• Spatial resolution: variable from 4km x 7km for GMI to 16km at nadir for ATMS and AMSU/MHS
• Sixteen overpasses (snowfall rate estimates) per day on average at a location in mid-latitude and more at high latitude
• The SFR product has been in operation at NOAA/NESDIS since 2012
2
NEXRAD Reflectivity
• Added NOAA-20 ATMS, GPM GMI, and F16/F17 SSMIS to the SFR suite recentlyü Snowfall is highly dynamic so requires adequate temporal coverage
ü The newly added satellites greatly improve the SFR temporal resolution
SFR Suite
GPMMetop-A Metop-BMetop-C
NOAA-18
F18
S-NPP/NOAA-20
{F16NOAA-19
F17
Operation; New development; Future development; Degradation3
• SFR algorithm includes two main components
ü Snowfall detection
ü Snowfall rate estimation
• Snowfall Detection (SD)
ü Statistical approach (Kongoli et al., 2015, 2017)
ü Optimal combination of a satellite-based module and a model-based module
• Snowfall Rate
ü Physical model (Meng et al., 2017; Ferraro et al., 2018)
ü 1DVAR-based retrieval
• All SFR algorithms adopt the same framework but have many differences from sensor to sensor and even among the same type of sensors
Algorithm
4
Snowfall Detection Algorithm• Satellite-based module – sensor/satellite dependent
ü Coupled principal component and logistic regression (LR) model or LR modelü Predictors: Channels at and above 88/89 GHz, or additional temperature sounding channels ü One or two temperature regimesü Training data sets: matching satellite data and ground snowfall observations (QCLCD), i.e.
‘truth’ data
• NWP model-based moduleü LR modelü Input data: RH, V Vel, Cloud Thickness
• Optimal combination of the two modulesü Output is probability of snowfall; use preset
thresholds to determine snowfall
GMI SD Performance
5
Warm Regime Cold Regime
POD (%) 68 56
FAR (%) 15 14
HSS 0.55 0.45
• Additional model-based screening to improve accuracyü Relative humidity, temperature, cloud thickness
Snowfall Rate Algorithm (1/2)• 1D variational method to derive cloud properties
ü Forward simulation of TB’s with a radiative transfer model (RTM) (Yan et al., 2008)
ü Iteration scheme with ΔTBi thresholdsü Ic and De are retrieved when iteration stops
Ic: ice water pathDe: ice particle effective diameter
ei: emissivity at 23.8, 31.4, 89(MHS)/88.2(ATMS), 157/165.5, and 190.31/183±7 GHz for AMSU/MHS and ATMS (similar channels for GMI and SSMIS)TBi: brightness temperature at 23.8, 31.4, 89/88.2, 157/165.5, and 190.31/183±7 GHz for AMSU/MHS and ATMS (similar channels for GMI and SSMIS)
A: Jacobian matrix, derivatives of TBi over Ic, De, and ei
E: error matrix
6
Snowfall Rate Algorithm (2/2)
• Terminal velocity is a function of atmospheric conditions and ice particle properties, Heymsfield and Westbrook (2010):
! = #$%&'(
• Snowfall rate modelAssumption: Ice water content has a linear distribution through cloud column
)*$+ =I-η./0
12&3&'(%45/
6(0789/9; 1 + 8(
>/0
#./0?&@&'3B/
C/0
− 1
0
E(
• Calibration against “truth” data is required to compensate for non-linear IWC distribution through cloud column and correct systematic bias
7
SFR Calibration• S-NPP and NOAA-20 SFR calibration data: MRMS and Stage IV
precipitation analysesü MRMS: Multi-Radar Multi-Sensor precipitation product
ü Stage IV: Uses MRMS precipitation data as input, incorporates gauge/model/satellite data, and applies human quality controls
ü Snowstorm data from two winter seasons (2015-2016)
ü CONUS coverage
• Calibration for other satellites: MRMS; will be calibrated against Stage IV
• Histogram matching (Kidder and Jones, 2007):ü CDF adjustment
ü Lease square method to achieve optimal overall agreement between SFR and Stage IV CDFs
• Cubic adjustment function
8
Validation - Seasonal Average (Jan – Mar, 2017)
mm/hrmm/hrS-NPP SFR
NoCover-
age
• Good agreement in overall pattern between S-NPP SFR and Stage IV seasonal averages• Statistics indicate that SFR performs well against
the Stage IV radar/gauge snowfall analysis• The SFR product can fill in radar gaps in areas
(e.g. western US) where Stage IV does not have coverage
Stage IV
9
Validation – Case Study
10
• A major nor’easter swept through the Mid-Atlantic and the Northeast on March 14-15, 2017
• The ATMS and MHS SFR products captured the evolution of the snowstorm
• The ATMS SFR outperforms the MHS SFR
24-hour snowfall accumulation ending March 15, 2017 12 UTC
NOHRSC Snow Analysis
SFR
CorrCoeff
Bias(mm/hr)
RMS(mm/hr)
ATMS 0.68 0.05 0.71
MHS 0.65 -0.13 1.01
(Courtesy of Patrick Meyers, ESSIC/CICS-MD)
11
New Development - SFR Recalibration
Before Cal Hist. Cal Regress. Cal
Corr. Coeff. 0.51 0.51 0.62
Bia (mm/hr) -0.25 -0.02 0.02
RMSE (mm/hr) 0.67 0.64 0.54
Before Cal
Regress Cal
• A new calibration approach was developed recentlyü Regression function for SFR bias
ü Predictors: satellite measurements, retrieved parameters, GFS atmospheric parameters (e.g. temperature at 650 mb)
• Significant improvement compared to calibration by histogram matching
• Currently only for ATMS (S-NPP and NOAA-20)
11
SFR Advantages• Over data and radar deficient areas like
much of Alaska, the product can aid forecasters in warning decisions by lending confidence. • Potentially increases geographic resolution
for areas of heavy snow within conceptual, observational, and model derived guesses of where the storm is located.• Adds more detail than NEXRAD data by
supplying aggregated data from multiple atmospheric levels in contrast to the single vertical levels of NEXRAD. • Performs best for stratiform snow in non-
shallow clouds and mesoscale and synoptic scale systems (satellite resolution makes a difference!)
12
• Land only product without coverage for coast and ocean
• Measures snowfall rate within the atmosphere and not what is falling to the ground, which may be different because of factors such as wind and evaporation. Highest correlation with radar snowfall rate from 30 minutes later (time shift used for cal/val), potential value for nowcasting.
Known Issues
• For sensors with coarser resolution, i.e. ATMS and AMSU/MHS, the 16 km resolution at nadir means that potential narrow bands from ocean-effect snowfall, local convergence zones or convective regimes may go undetected.
• The current SFR algorithm is not applicable for regions where surface air temperature is 7 F or colder. Many impactful snowstorms in Alaska begin, particularly in inland Alaska and along the Arctic and West Coasts, when temperatures are colder.
• The current snowfall detection is not sensitive to shallow snowfall.
• Algorithm reliance on GFS model data means that model error may impact rain/snow detection
13
• Assessment in operational environment as part of the JPSS Proving Ground activities
• Collaborations among NASA SPoRT, NESDIS, CICS-MD and NWS Weather Forecast Offices
• Participating WFOs from Alaska/Western/Eastern/Southern regions
• Forecasters feedback indicates that the SFR product is useful for weather forecasting
• Some important improvements were made based on user feedback: latency, cold extension, looping capability etc.
Product Assessment
14
Response to User Feedback – Latency Reduction• Weather forecasting requires low latency;
operational polar-orbiting satellite data often do not meet the requirement
• Use Direct Broadcast data to reduce latencyü Retrieve DB CONUS and Alaska L1B data from the
Cooperative Institute for Meteorological Satellite Studies (CIMSS)/University of Wisconsin, Madison and the Geographic Information Network of Alaska (GINA)/University of Alaska
ü Available DB data: NOAA-20, S-NPP, POES and Metopsatellites; (GPM)
ü Produce SFR estimates to less than 30 minutes after satellite observations
ü Webpages:
NRT SFR Image at NASA/SPoRT
15
https://weather.msfc.nasa.gov/cgi-bin/sportPublishData.pl?dataset=snowfallratealaskahttp://cics.umd.edu/sfr
Response to User Feedback – Looping Capability• Product looping is highly desirable for weather forecasting. However, the polar-orbiting
satellite-based SFR product is not suitable for looping due to its coarse temporal resolution
• Achieve snowfall rate looping capability by merging NOAA/NSSL Multi-Radar Multi-Sensor (MRMS) radar precipitation analysis with the SFR product from eight satellites including S-NPP and NOAA-20
• The merged product, mSFR, provides improved spatial (filling radar gaps) and temporal (every 10-min) coverage
16
CMORPH Stage IV Radar CMORPH2
Application in Hydrology - CMORPH 2
• SFR has been integrated in the NOAA CMORPH2 blended precipitation analysis
• Accumulated precipitation (rainfall, snowfall, mixed) from the Stage IV radar observations (middle), the currently operational CMORPH (left) and the second generation CMORPH (CMORPH2, right)
• The operational version CMORPH missed / under-estimated snowfall over most of the regions covered with surface snow
• CMORPH2 is capable of capturing snowfall along the path of Grayson over the east coast
Winter Storm Grayson 3-5 January 2018
(Courtesy of Joyce and Xie)
17
• Recalibrate SFR for all satellites against Stage IV and
implement the new algorithms before next winter
• Further enhancement of the SFR product to meet user needs
ü Snowfall detection especially for light snowfall
ü More sophisticated cloud microphysics in the radiative transfer model
• Continued validation study
• Extend SFR retrieval to over ocean, coast, and sea ice
• For Alaska
ü An Alaska SFR algorithm? Alaska has different climatology from most
CONUS (e.g. max temperature profile criterion)
ü Cold snow extension
ü Calibration/validation data from Alaska?
Ongoing & Future Development
18
Summary• An overland SFR product was developed from satellite observations and has
been in operation at NOAA/NESDIS since 2012
• The SFR product was calibrated and validated against ground observations and radar precipitation estimates
• Continuous development enhances product quality and adds new satellites to the SFR suite to improve temporal coverage
• Forecaster feedback indicates that SFR is a useful tool for weather forecasting particularly in data and radar deficient areas
• Future development can be expected to resolve some of the known issues of the current SFR algorithm
AcknowledgementsJPSS PGRR
JPSS STAR
JPSS PSDI
NASA SPoRT
19
Liquid Equivalent Snowfall Rate (SFR)
Assessment for Alaska
Virtual Alaska Weather SymposiumFebruary 2019Wes Adkins, NWS JuneauHuan Meng, NESDIS
Outline of the SFR Assessment
I. Notable Alaska case studies (both good and bad)
1. Edna Bay and Thorne Bay2. Susitna and Copper River Valleys3. Haines Highway4. King Salmon and New Stuyahok5. Juneau6. Pelican
II. Best Practices for ForecastersIII. Additional Research from WFOs to increase
SFR utility
Case Studies for SFR Assessment
Case Study 1: Edna Bay & Thorne Bay
Thorne Bay
Edna Bay
• Tight baroclinic system moved over southern Panhandle.
• Strong dynamics associated with system.
• Winter storm initially posted for Petersburg.
• A winter weather advisory for 2-4 inches posted for Edna Bay and Thorne Bay
• After moderate to heavy snow reported at Ketchikan, forecasters were prompted to phone spotters.
• The winter weather advisory was upgraded to a winter storm warning for 4-8 inches of snow in Edna Bay and Thorne Bay.
Storm Totals:
Edna Bay: 12 inches (SFR miss)Thorne Bay: 9 inches (SFR miss)Petersburg: 7 inches (SFR success)
Case Study 1: Edna Bay & Thorne Bay
Thorne Bay
Edna Bay
Edna Bay Thorne Bay
Petersburg
Ketchikan
• SFR correctly indicates heavier snowfall rates around Petersburg.
• SFR incorrectly indicates no snowfall in Edna Bay and Thorne Bay when in fact, it was heavier than Petersburg at the time.
We asked developers why?
Storm Total Snow:
Petersburg: 7 inches Edna Bay: 12 inchesThorne Bay: 9 inches
Case Study 1: Edna Bay & Thorne Bay
Thorne Bay
Edna Bay
• The algorithm takes into account that scattering decreases in heavy snowfall.
• The Petersburg area, as expected, registered less scattering than areas around it due to heavy snow.
• But why did Edna Bay & Thorne Bay still feature higher rates of scattering, when they experienced heavier snowfall than Petersburg?
• Developers speculate that super-cooled water droplets from the storm’s strong dynamics over the Pacific negated any scattering due to heavy snow.
• This illustrates why the algorithm is unreliable over water and coastal regions.
Super-cooled water droplets negated scattering from heavy snow rates in Edna & Thorne Bays
Less scattering due to heavy snow ratesPetersburg
SFR omitted heavy snow rates in Edna and Thorne Bays
SFR indicated heavy snow rates near Petersburg
Atmospheric Scattering.
Case Study 2Susitna and Copper River Valleys
SFR successfully captured a winter weather advisory in the Anchorage CWA, all interior locations.
• SFR rates similar to Talkeetna ASOS. (yellow)
• SFR showed lower snow rates north of Talkeetna through Broad Pass (orange) where spotters confirmed less.
• SFR showed higher values in NW Copper River Basin where no spotters exist. (red)
Source: Shaun Baines, WFO Anchorage
Talkeetna TalkeetnaCopper River Valley
TalkeetnaTalkeetnaCopper River Valley
Copper River Valley
Copper River Valley
SFR : 0.04 in/hrPATK: 0553Z 0.04 in.
0649Z 0.03 in.
SFR: 0.03 in/hrPATK: 0553Z 0.04 in.
0649Z 0.07 in.
SFR: 0.08 in/hrPATK: 0649Z 0.03 in.
0753Z 0.07 in.
SFR: 0.04 in/hrPATK: 0649Z 0.03 in.
0753Z 0.07 in.
Broad Pass
Broad Pass
Broad Pass
Broad Pass
Case Study 3Haines Highway
• Forecasters predicted a winter storm event in spite of model guidance.
• Thinking: a wedged cold air mass along the Haines Highway within the Chilkat Valley would remain entrenched with a rush of warm moist air overrunning aloft to produce heavy snowfall.
• A winter storm warning was posted for 3 to 9 inches of snowfall along the Highway.
A Winter Storm Warning was issued for the Haines Highway
with a level of confidence “a few hairs above shaky.”
Case Study 3Haines Highway
SFR: 03Z 2018 03 11
Haines
0.0 1.0 2.0 3.0
SFR image taken just before heavier snow rates were detected.
Note the heavier snow indicated by the darker blue areas south of the Haines Highway.
An example of how SFR can predict snowfall rates ahead of its actual occurrence.
Haines
Pleasant Camp
The Haines Highway is labeled in yellow. The international border is in red.
Alaska
Canada
Heavy Snow Rates
(darker blue areas)
Case Study 3Haines Highway
• Before interrogating SFR, forecasters made calls to spotters in the afternoon, most answering “Nothing happening.”
• Forecasters then discussed canceling the warning with no precipitation being reported as many models suggested.
• But SFR confirmed heavier amounts approaching the area (and original conceptual model of forecasters)
• Repeated calls to spotters as well as SFR data dissuaded forecasters from cancelling warning.
Winter Storm Warning Verification:
Co-op observers with the Border Customs Facility on the Haines Highway near Pleasant Camp recorded 20 inches of wet cemented snow before it changed to rain many hours later. (SFR Success)
Case Study 4King Salmon and New Stuyahok
SFR detected no snowfall from west of King Salmon north to New Stuyahok because of GFS detected warm layer aloft
SFR did not perform well in the lowlands of Alaska’s Naknek Bay region where heavy snowfall was almost completely omitted.
Source: Michael Lawson, WFO Anchorage
New Stuyahok
King SalmonWarm layer of temperatures > 0.5 C indicated at 1700 feet on GFS sounding.
Case Study 4King Salmon and New Stuyahok
The SFR algorithm detects snowfall first before calculating rates.
Because of the warm layer detected on the GFS sounding, developers reran the case using a higher atmospheric column temperature for detecting snow, raising it from 0.5 C to 1.2 C.
Notice the differences between the top and bottom SFR. On the bottom the SFR threshold for snow detection was increased from 0.5 C to 1.2 C and it performed much better.
As a result of this case and others, developers changed the snow identification threshold within Alaska only to 1 C, rather than the lower 48 version of 0.5 C.
SFR detects no snow where column T > 0.5 C
SFR detection improved drastically where column T > 1 C
Not Good
Very Good
Case Study 5: The Frozen AR / Juneau
• An Atmospheric River (AR) was forecast to move into Southeast Alaska with snow becoming heavy rain.
• NWS pushed message on social media hard
• Confidence was high.
• But things went terribly wrong and SFR was of no more help than most everything else.
IMERG1306Z
GOES PW 1306Z
Case Study 5: AR / Juneau
• SFR failed to identify snow in Juneau for whole event.
• Precipitation rates in adjacent mountain areas may have been comparable to what fell in Juneau
• Large variability in Juneau area precipitation type may have contributed to SFR miss.
• Juneau flagged by SFR as “no snow” because GFS model surface and column temperatures were too warm.
• Developers re-ran algorithm with higher threshold but SFR still failed to identify snow.
Juneau
Juneau AirportTemperatures
Time
03Z06Z09Z
GFS
37 F37 F38 F
Actual
32 F33 F32 F
Case Study 5: AR / Juneau
• The “changing to rain” paradigm was eventually abandoned.
1. Social media lit up with impressive snowfall amounts.
2. Aircraft soundings indicated isothermal atmosphere caused by snowmelt cooling and no sign of warming.
• It took 12 to 24 hours to fully change over to rain.
• SFR performed poorly but so did everything else including forecasters.
A late Winter Storm Warning was posted during the Event. Totals:
Douglas/ W Juneau 10-15 inches
Mendenhall Valley 6 to 8 inches
Auke Bay 4 to 6 inches
Downtown Juneau 2 to 3 inches
Juneau Airport 2.2 inches
Out the Road < 1 inch and rain
MDCRS: Juneau2018 02 13 0254Z
Case Study 6: Pelican
• Pelican is a tiny community of less than 100 residents.
• Located near the Gulf of Alaska, it is situated along a deep fjord known as Lisianski Inlet that shelters the town from direct marine influences and can make it a snow-hole in winter.
• Home to one Co-op observer who reports climate data each day in the morning.
• At the time the once-per-day co-op report was the only area surface observation.
Photo credit: Kimberly Vaughan
Case Study 6: Pelican
SFR: 2018 02 13 0447Z
Pelican, AK is situated in the center of the yellow circle. Only 4 inches of snow had fallen in Pelican as of 8 AM, which fell short of expectations. Due to lack of definite information, NWS Juneau let the warning expire at noon (21Z). This SFR image was timestamped at 2033Z. Latency, however, is continuing issue.
• Pelican Observer phoned Thursday morning to report a paltry 4 inches of snow when forecasters expected much more.
• Given earlier disappointing ground truth, no webcams, no available spotters, and a meteorological paradigm still in existence, forecasters allowed warning to expire without extension.
Case Study 6: Pelican
• A cold upper low in the gulf would be spreading snowfall along the eastern gulf coast. • A Winter Storm Warning was out for Pelican through noon on December 20.• GOES 17 imagery depicted the synoptic situation quite well.
GOES17 IR Legacy Band 2018 the morning of the storm.
Pelican
Case Study 6: Pelican
SFR: 2018 02 13 0447Z
• Because of SFR indicated heavy shower and “training” exhibited by GOES 17, the warning could have been extended.
• This SFR image was taken minutes before the decision was made to let the warning expire.
• But SFR latency is still an issue as forecasters did not receive data timely.• Based on higher resolution GPM
GMI imagery, SFR successfully depicted intense snow showers with rates of ~0.10 in/hr (the yellow and orange pixels).
• Given Pelican’s median snow ratios of 9:1 (Levin), this perhaps suggested snow rates close to 1 in/hr.
Pelican
Case Study 6: Pelican
• The observer phoned in the afternoon that it had snowed moderately all day with a total of around 1 foot from the storm. Very close to the original 13 inches forecasted.
• SFR Success for a coastal site (except it was late).
SFR: Best Practices for Forecasters
• Know SFR’s limitations before examining it:
1. Coastal Areas with super-cooled clouds are generally mischaracterized by the algorithm as “no snow.” During strong dynamic onshore flow, this weakness can spread inland.
2. SFR algorithm returns “No Retrievals” where surface temperatures are <= 7 F, which renders it useless for much of Interior Alaska, Arctic Alaska, and West Coasts at times.
• Monitor SFR during snow events! If not, all potential information gains are moot.
SFR: Best Practices for Forecasters
• Question SFR data before acting on it.
1. Is the SFR data plausible meteorologically? 2. Do other data sets and local knowledge
support SFR readings? • Radar, Satellite & Webcams• METARs, SNOTELs, & Mesonet stations• Reports from area Weather Spotters, Co-
op Observers, and Core Partners (Volunteer Weather Spotters need to be phoned much of the time).
• Crowdsourcing from mPing & Social Media
• Likewise, if other data sets imply heavy snow, then do not discount false negatives from SFR. Ponder how known issues may limit SFR for situation.
WFO Actions to Improve SFR Utility
• Continually assess SFR success (and failure) and report back to NASA SPoRT.
• Powerpoint mapping slides from forecasters are often worth a thousand words.
• Conduct regional snow ratio studies where data is available to aid developers in translating the microwave derived liquid data to best guess snowfall rate.
• Additional studies could be done in the snowy but windy regimes of the Bering and Arctic coasts. David Levin (2018)
SFR – Alaska Assessment
Thank you
[email protected]@noaa.gov
Contributors::
Shaun BainesEmily BerndtAaron JacobsMichael LawsonDavid LevinEdward LiskeKris White
NWS AnchorageNASA SPoRTNWS Juneau
NWS AnchorageNWS JuneauNWS JuneauNASA SPoRT
