The Use of GPS Radio Occultation Data for Tropical Cyclone Prediction

Bill Kuo and Hui LiuUCAR

Outline• Prediction of tropical cyclones:

– Current forecast skills– Comparison of operational model performance for 2011

• The NOAA Hurricane Forecast Improvement Project• Impact of assimilation approaches on NCEP GFS hurricane forecast

skills• GPS radio occultation (RO) and COSMIC• Assimilation of GPS RO data• Impact on the tropical cyclone analysis and prediction

– Hurricane Ernesto (2006)– Typhoon Morakot (2009)

• Outlook

Current capability of NHC

Good track forecast improvements. Errors cut in half in 15 years. Skill of 48h forecast as good as 24-h forecast 10 years ago.No improvement in intensity forecast. Off by 2 categories 5-10% of the time.

77 69 59 61 65 62 62 56

197219

163

210

166 177149 142 132 120

93 101 99 103 110 97

380396

297

417

311342

271243 237

215

170196

157191 200

179

353 356326

279316

254

324 310 309

0

50

100

150

200

250

300

350

400

450

500

Y1994 Y1995 Y1997 Y1998 Y1999 Y2000 Y2001 Y2002 Y2003 Y2004 Y2005 Y2006 Y2007 Y2008 Y2009 Y2010

時 間 年

誤差值

公里

12小時預報 24小時預報

48小時預報 72小時預報

48小時預報趨勢分析線 24小時預報趨勢分析線

Current capability of CWB in (km)

Good improvement from 1994 to 1005. 24 hour forecast errors reduced by half in 10 years. However, the skill becomes somewhat stagnant after 2005. Does not keep track of intensity forecast.

2011 Preliminary NHC Verifications

ECMWF and GFS did well (again).

ECMWF beat consensus at longer ranges.

TVCA beat FSSE.

AEMI not as good as GFSI, even at 5 days.

GFDL and HWRF middle of the pack.

NGPS, CMC, UKMET trailed.

Summary of 2011 Results and Implications

• Deterministic ECMWF (~16km) and GFS (~21 km) performed very well for 2011 (similar to that of 2010). Resolution matters.

• ECMWF beat multi-model consensus at longer ranges. High resolution deterministic global forecast is very useful.

• Multi-model consensus (TVCA) beat the FSU super ensemble (FSSE).

• Lower-resolution GSF global ensemble (AEMI, ~63km) did not perform not as well as high-resolution deterministic GFS (~21km), even at 5 days. Again, resolution matters.

• Performance of operational regional models (GFDL, 48/16/8 km, and HWRF, 27/9 km) was not as good as ECMWF and GFS global models. They are in the middle of the pack. Why didn’t regional model do better? What should they be used for?

• NGPS, CMC, UKMET global models trailed the other global and regional models. Hurricane forecast skill is different from general synoptic forecast skill (e.g., 500 mb height).

NOAA Hurricane Forecast Improvement Project (HFIP)

• 10-Year Goals:– Reduce average track error by 50% for day 1 through 5– Reduce average intensity error by 50% for day 1 through 5– Increase the probability of detection (POD) for rapid intensity

change to 90% at Day 1 decreasing linearly to 60% at day 5, and decrease the false alarm ratio (FAR) for rapid intensity change to 10% for day 1 increasing linearly to 30% at day 5. The focus on rapid intensity change is the highest forecast challenge indentified by the National Hurricane Center.

– Extend the lead time for hurricane forecasts out to Day 7 (with accuracy of Day 5 forecasts in 2003).

• 5-Year Goals:– Reduce track and intensity forecast errors by 20%

HFIP Hurricane Research Priorities

• Advancement in data assimilation (DA) technique– Technique to maximize the usefulness of observations– Identification and analyses of the sources of model error

• Physics advances– Air-sea and BL processes– Microphysical/aerosol/radiation processes

• Advanced model diagnostic techniques– Analyses and forecast of vortex low-order wavenumber evolution– Analyses and forecast of large-scale and hurricane environment evolution (e.g.

shear/track)• Development of high resolution ensembles

– Identify techniques to utilize ensemble members for improved intensity guidance

Pros Cons

3D-Var • Inexpensive• Practical approach for operational purposes

• Static (fixed) error covariance• Limited performance

4D-Var

• Use of observations at appropriate time• Flow-dependent error covariance• Suitably balanced analysis due to explicit use of model dynamics & physics• Treats observations simultaneously• Can handle a wide range of scales simultaneously• Can use all observations from frequently reporting stations (high temporal resolution)• Demonstrated capability (better than 3D-Var)

• Expensive• 4D-Var iterations are sequential – difficult to make full use of current computers & likely impossible in foreseeable future.• Require TL/AD models – difficult to maintain up-to-date code

EnKF

• Simple to design and code• No AD or minimization is needed• Provide estimates of error covariance (but noisy)• Allows for the full history of the evolution of forecast errors.• Can use the full model, including non-differentiable and stochastic terms.• Explicit estimate of errors.• Facilitates ensemble forecasting.

• Expensive• Treats one observation at a time, cost increases with data volume• Can handle limited range of scales (because of localization)• Can only estimate mean and covariances from a small sample, assuming Gaussian distributions• Is still experimental (comparable or slightly better than 3D-Var)

Hybrid• “Error of the day” for 3D-Var or 4D-Var can be obtained through ensemble• Easy to implement• Considered to be an attractive compromise

• In experimental phase• Still expensive

Data Assimilation Methods

Comparison of EnKF vs 3D-VarSame GFS model at T574L64 (~21km at 25N)

EnKF: T254L64 (~47km at 25N)

GSI (3D-Var): T574L64(~21km at 25N)

Use the same operational observational data stream

2010 North Hemisphere tropical cyclones

Hamill et al (2011) MWR

Comparison of EnKF, 3D-Var, and Hybrid

Deep layer (850-200 mb) vector wind, 26 July – 17 Sep, 2010, 20N-20S. Hamill et al. (2011)

Comparison between ECMWF (T639L62, ~28km at 25N) and GFS/EnKF (T254L64, ~47km at 25N) ensemble prediction of tropical cyclones.

ECMWF has significantly better performance over Western Pacific.

The GFS/EnKF has larger errors over the Western Pacific than Eastern Pacific or Atlantic.

Hamill et al. (2011)

ECMWF and GFS Ensemble Forecast Comparison

Differences between ECMWF and GFS/EnKF for 200 mb wind, averaged from 7/26 – 9/27, 2010. From Hamill et al (2011). Larger difference over Western Pacific is noted.

Radio Occultation Technique

COSMIC – Six-Satellite constellation mission, launched in April 2006

Characteristics of GPS RO Data• Limb sounding geometry complementary to ground and space nadir viewing

instruments• Global coverage• Profiles ionosphere, stratosphere and troposphere• High accuracy (equivalent to <1 K; average accuracy <0.1 K)• High precision (0.02-0.05 K)• High vertical resolution (0.1 km near surface – 1 km tropopause)• Only system from space to observe atmospheric boundary layer• All weather-minimally affected by aerosols, clouds or precipitation• Independent height and pressure• Requires no first guess sounding• No calibration required• Independent of processing center• Independent of mission• No instrument drift• No satellite-to-satellite bias• Compact sensor, low power, low cost

All of these characteristicshave been demonstrated inpeer-reviewed literature—many were proven by COSMIC

GPS RO observations advantages for weather analysis and prediction

• There are considerable uncertainties in global analyses over data void regions (e.g., where there are few or no radiosondes), despite the fact that most global analyses now make use of satellite observations.

• GPS RO missions (such as COSMIC) can be designed to have globally uniform distribution (not limited by oceans, or high topography).

• The accuracy of GPS RO is compatible or better than radiosonde, and can be used to calibrate other observing systems.

• GPS RO observations are of high vertical resolution and high accuracy. • GPS RO is an active sensor, and provides information that other satellite

observing systems could not provide• GSP RO provide valuable information on the 3D distribution of moisture

over the tropics, which is important for typhoon prediction.

Problems of using GPS RO data in weather models

• GPS RO data (e.g., phase, amplitude, bending angles, refractivity) are non-traditional meteorological measurements (e.g., wind, temperature, moisture, pressure), advanced assimilation systems are needed to effectively assimilate such observations.

• The limb-viewing geometry makes their measurement characteristics very different from in situ point measurements (e.g., radiosonde) or the nadir-viewing passive microwave/IR measurements. Advanced observation operators are needed.

• The GPS RO measurements are not perfect and subject to various sources of error (e.g., measurement noise, tracking errors, uncalibrated ionospheric effects, super-refraction,…etc).

• Model configuration and issues (e.g., domain size, model top, parallelization, regional/global, ionosphere, model errors) affect the choices of assimilation strategies and outcome.

Φ1, Φ2, a1, a2

α1, α2

α

N

GPS RO measurements & processing

T, e, P

Phase and amplitude of L1 and L2

Bending angles of L1 and L2

Bending angle (Neutral A.)

Refractivity

Ionospheric correction

Bending angle optimization & Abel inversion

A priori meteorological data & 1D-Var

Radio holographic methods, taking multipaths into accountGeometrical (single ray) w/o amplitudes

Satellites orbits & Spherical symmetry

Observation type and operators

Obs Type Obs Operator

N. A.

Retrieved T, e, P Interpolation in time & space

Refractivity Local refractivity

Refractivity Nonlocal refractivity

Refractivity Nonlocal excess phase

Bending angle Local bending angle

Bending angle 2D/3D Ray tracer

Phase + Ray shooting (orbits)

L1,L2Bending angle + Ionospheric &

Plasmaspheric modelPhase, amplitude

Com

plex

ity o

fO

pera

tor

Err

or C

hara

cter

istic

s

Neu

tral

atm

osph

eric

A Priori information is neededError structure is too complicated to properly characterize

Possible choices

Operator is too complicated for now

Pros Cons

3D-Var • Inexpensive• Practical approach for operational purposes

• Static (fixed) error covariance• Limited performance

4D-Var

• Use of observations at appropriate time• Flow-dependent error covariance• Suitably balanced analysis due to explicit use of model dynamics & physics• Treats observations simultaneously• Can handle a wide range of scales simultaneously• Can use all observations from frequently reporting stations (high temporal resolution)• Demonstrated capability (better than 3D-Var)

• Expensive• 4D-Var iterations are sequential – difficult to make full use of current computers & likely impossible in foreseeable future.• Require TL/AD models – difficult to maintain up-to-date code

EnKF

• Simple to design and code• No AD or minimization is needed• Provide estimates of error covariance (but noisy)• Allows for the full history of the evolution of forecast errors.• Can use the full model, including non-differentiable and stochastic terms.• Explicit estimate of errors.• Facilitates ensemble forecasting.

• Expensive• Treats one observation at a time, cost increases with data volume• Can handle limited range of scales (because of localization)• Can only estimate mean and covariances from a small sample, assuming Gaussian distributions• Is still experimental (comparable or slightly better than 3D-Var)

Hybrid• “Error of the day” for 3D-Var or 4D-Var can be obtained through ensemble• Easy to implement• Considered to be an attractive compromise

• In experimental phase• Still expensive

Data Assimilation Methods

Current DA use of GPS RO data3D-Var 4D-Var EnKF Hybrid

2D Bending angle SSI ECMWF

Local bending angle

GSI WRF-Var

ECMWF UKMO

Meteo France

Nonlocal excess phase

WFR-VarGSI* WRF-DART

Nonlocal refractivity JMA

Local refractivityNCEP

WFR-VarCWB

Env. CanadaJMA, CMA WRF-DART WRF-DA

Under test

*Ma et al. (2011) atmospheric river study, using regional GSI + ARW.

Impact of COSMIC on Hurricane Ernesto (2006)

Hurrican Ernesto:

Formed: 25 August 2006

Reached Hurricane strength: 27 August

Dissipated: 1 September 2006

15:50 UTC 27 August 2006

Picture taken by MODIS, 250 m resolution

Forecast experiments

• No GPS: initialized from AVN/GFS analysis at 2006-08-23-06Z

• GPS all: assimilate all 15 GPS profiles at 2006-08-23-06Z, followed by a 5-day forecast

• GPS 1 : only assimilate 1 GPS profile at 2006-08-23-06Z, followed by a 5-day forecast

WRF/DART EnKF system

36-km, 32 member ensemble

Use non-local excess phase observation operator

Low-level moisture change by assimilating GPS

Assimilation of 1 GPS RO sounding in the vicinity of Hurricane Ernesto caused a 1.8g/kg increase in moisture at 1.5 km. This was enough to get the hurricane genesis going.

NCAR 4-Day Ernesto Forecasts

Y.-H. Kuo (NCAR), 2007

Forecast with GPS Forecast without GPSThe Actual Storm

How does GPS RO data improve hurricane forecast?

WRF/DART ensemble assimilation of COSMIC GPSRO soundings

• WRF/DART ensemble Kalman filter data assimilation system• 36-km, 32-members, 5-day assimilation• Assimilation of 178 COSMIC GPSRO soundings (with nonlocal obs

operator, Sokolovskiey et al) plus satellite cloud-drift winds• Independent verification by ~100 dropsondes.

178 COSMIC GPSRO soundings during 21-26 August 2006

From Liu et al. (2011)

Distribution of GPS RO data

Experiment DesignName Conventional data GPS RONODA No No Continuous forecastONLY No Yes GPS onlyONLY6km No Yes above 6km GPS only above 6kmCTRL Yes No ConvectionalRO Yes Yes Conventional + GPSRO6km Yes Yes above 6 km Convectional + GPS6km

> The first three experiments try to identify the impact of GPS data in a clean setting (without the use of any other data)> The second three experiments assess the impact of GPS RO data in a more realistic operational setting.> Conventional data includes radiosondes, surface, aircraft reports, satellite cloud track winds, no radiance.

Analysis of 850 hPa Wind and Total Column Cloud Water(06Z August 23, 36 hours before Ernesto’s genesis, GPSonly run)

Convergence and convection in Ernesto’s genesis area.

Where Ernesto formed at 0000 UTC 25 August

Daily Analysis Increments of Q and T (GPSonly run)(850 hPa, August 23, 2 days before Ernesto’s genesis)

2-hour Forecast Difference of GPSonly-NODA (700 hPa)

00Z 24 August

Water vapor 06Z 23 August Wind

00Z 25 August Ernesto’s genesis time

Daily Analysis Increments of Wind(August 23, 2 days before Ernesto’s genesis, m/s)

GPSonly>6km run

700 hPa

GPSonly run 250 hPa

Strong correlation between GPS in the lower troposphere and winds at all levels.

2-hour Wind Forecast Differences (250 hPa) GPSonly-FCST 06Z 23 August GPSonly>6km-FCST

00Z 24 August

00Z 25 August Ernesto’s genesis time

2-hour Forecasts RMS fit to Radiosondes and dropsondes(Averaged over tropical Atlantic, 21-26 August, 2006)

Assimilation Experiments of RO and Conventional Data

• CTRL (NO GPS) run: Assimilate conventional data.• RO6km run: Add RO refractivity data only higher than 6km.• RO run: Add RO refractivity data of all levels.

• Non-local RO refractivity operator is used.

Analyses of SLP and 1000 hPa Vorticity (00UTC 25 August)

CTRL(NOGPS)

RO6km

RO

48h-forecast of Track errors and SLP IntensityCTRL

GPS>6km

GPS

Ensemble mean of 48-hour forecasts of Ernesto’s central sea level pressure (hPa, top) and track error (km, bottom) initialized from the analyses at 0000 UTC 25 August 2006

Total Integrated Water Vapor of 48-hour Forecasts(00UTC 25 August)

RO CTRLIR image

24h FCST

48h FCST

48-hour Forecasts of Ernesto (2006) with Assimilation of GPS and Conventional Observations

RO RO6kmIR Image

24h FCST

48h FCST

Assimilation of GPS > 6km shows less positive impact

Summary of Ernesto Study

• Assimilation of GPS RO data adds moisture in the lower troposphere, and produces noticeable wind increments, consistent with the convective environment.

• GPS RO data in the lower troposphere is crucial for creating a favorable environment for hurricane genesis

• Assimilation of GPS RO data produced improved analysis and subsequent forecasts both in terms of track and intensity.

Impact of GPS RO data on Typhoon Morakot (2009)

Assimilation experiments for Morakot

• Assimilation from 00Z August 3 to 18Z August 8, 2009 with 2-hourly cycling.

• 36km analysis grid with 35 levels• 48h Forecast with nest grids 36km/12km/4km starting

at August 7 00Z.• CWB observations are used• CTL run: Assimilate CWB conventional observations• GPS run: CTL run + COSMIC data.

Track analyses (August 3 06Z - 8 18Z)

NOGPS

GPS

Ensemble mean

Observed

SLP intensity analyses (August 3 06Z - 8 18Z)

NOGPS

GPS

Ensemble mean

Observed

24h Rain forecast (August 7 - 8 00Z)

NOGPS

GPS

Ensemble mean

Observed

48h Rain forecast (August 8 - 9 00Z)

NOGPS

GPS

Ensemble mean

Observed

Rain Probability Forecast (August 7-8 00Z)

Ensemble mean

Observed

Rain Probability Forecast (August 8 - 9 00Z)

Ensemble mean

Observed

WRF/DART Analyses and Forecasts for Morakot (2009)

• Control (with GPS RO): CWB operational observations including TC BOGUS data, 6-hourly analysis cycle.

• NOGPS: remove the GPS RO data from the control run.• Assimilations started 12UTC Aug. 3, 2009. 45km grid only

for both assimilation and forecasts.• CWB TWRF is also full cycled for the same period.• Cost of the 16 ensemble forecasts is equivalent to one

nested (45/15/3km) deterministic forecast.

Forecast initialized at 12 UTC 5 AugustNo GPS With GPS

Forecasts initialized at 12 UTC 6 AugustNo GPS With GPS

NO GPS

GPS

Difference

Two week assimilation, 1-14 June 2007, 850 mb wind field.

Summary of Morakot Study

• Assimilation of the GPS data:– Improve analysis and prediction of typhoon track– Improve rainfall prediction– Enhanced upper-level divergence, increased low-

level moisture flux, and intensity 500 mb height• Performance of WRF/DART is superior to

WRF-Var• Positive impact of GPSRO is also seen in

WRF-Var

48-h of track forecast errors for Typhoon Jangmi (2008) with and without COSMIC.

Typhoon Prediction During T-PARC

Assimilation of GPSRO data from COSMIC improved the quality of regional analysis and typhoon track errors. For four storms in T-PARC 2008, the average improvement is 13%.

From H. Liu and Jeff Anderson

COSMIC-II SoundingsGeographic Coverage

(6@72°, 6@24°)

1 hour

6 hour

3 hour

24 hour

Summary• GPSRO observations are not affected by clouds and precipitation• GPSRO observations provide three-dimensional water vapor

information over the tropics, which are crucial for:– Tropical cyclone genesis– Tropical cyclone intensity forecast– Tropical cyclone track forecast

• Current FORMOSAT-3/COSMIC sounding distribution is NOT optimal for tropical prediction. The data density is the lowest over the tropics (0.4 sounding over 500 km x 500 km in one day).

• FORMOSAT-7/COSMIC-2, with 10 times more soundings will significantly improve the skills of tropical cyclone prediction:– Improve track, intensity and rainfall forecast for typhoons affecting Taiwan– Directly contribute to all the goals of NOAA’s Hurricane Forecast

Improvement Project (HFIP)