Global Tropical Cyclone Winds from the QuikSCAT and OceanSAT-2 Scatterometers Bryan W. Stiles1, Rick Danielson2, W. Lee Poulsen1, Alexander Fore1, Michael J. Brennan3, Tsae-Pyng J. Shen1, and Svetla M. Hristova-Veleva1 (Email contact [email protected]) 1.Jet Propulsion Laboratory, California Institute of Technology, 4800 Oak Grove Dr. Pasadena CA, USA, 91109, 2. University Corporation for Atmospheric Research, PO Box 3000, Boulder CO, 80307 3. NOAA/NWS/NCEP/National Hurricane Center, 11691 SW 17th St, Miami FL, USA, 33165 © 2012. All rights reserved. Synopsis of Technique • • Using a simple neural network (e.g. ), we fit a nonlinear mapping from scatterometer data to wind speed. Inputs are: – Climatological Distributions of winds from 1999-2009 in the North Atlantic Basin Nominal QuikSCAT (version 2) Winds: Hurricane Ivan 23:37 UTC 11 Sept. 2004 8 sets of backscatter (Normalized Radar Cross-Section) values • • • 2 different azimuths, 2 different polarizations (and incidence angles), 2 different spatial scales (12.5 and 87.5 km) – a rain rate from the scatterometer noise channel . – cross track distance as a proxy for viewing geometry – Information from version 3 of QuikSCAT global wind retrieval product • • • • • Each sub-network is a weighted sum of tanh functions of 10-20 weighted sums of the inputs. Weights are constants that are set during a preliminary training procedure. Ground truth speeds are from H*WIND data from 2005 Atlantic hurricanes. - • Speed corrected for rain Maximum likelihood speed (no correction for rain) Rain Impact quantity Structure employs a set of sub-networks to simplify the mappings needed. - Used for fitting network weights. Attempt to correct wind direction in rain is left for future work. – Nominal direction retrievals from JPL QuikSCAT L2B products are maintained. Neural Network Structure Speed Net 2 Trained on 2005 QuikSCAT/H*WIND data QRAD rain rate Rain Impact Quantity MLE speed FINAL OUTPUT Neural Net Winds: Hurricane Ivan 23:37 UTC 11 Sept. 2004 CAT 1-3 by Quadrant SPEED Version 3 speed Rain Corrected Speed Network 1 Trained on AMSR/ SeaWinds Data CTD CTD QuikSCAT MLE speed Hurricane Speed Net 1 Trained on 2005 QuikSCAT/H*WIND data Backscatter values Hurricane Katrina (2005) as observed by QuikSCAT just before landfall; Same scale as Sandy above; 300 mile diameter TS winds CAT 4-5 by Quadrant QRAD rain rate QuikSCAT 1999-2009 Data Set QuikSCAT vs H*Wind (Atlantic) OceanSAT-2 Tropical Cyclone Winds • • • • • Examples Maximum Speed Tracks – Ivan 2004 Validation of QuikSCAT Winds. Vs. Best Track Max Speeds • collocations between 50 and 500 km • NNet bin averages are closer together, but suggest positive bias • JPL-V2,V3 weaker winds, bin averages farther apart, and negatively biased QuikSCAT vs SFMR and GPS drops Comparison between maximum best-track intensity (x-axis) and maximum QuikSCAT winds (y-axis) for all storm overpasses from 1999-2009. Different panels (a-f) correspond to different basins. Compared to neural net winds, version 2 winds are biased high in low intensity storms (due to rain effects) and saturate quickly above 30 m/s. Version 3 winds saturate quickly at higher winds but perform well below 30 m/s. OceanSAT-2 is a 13.4 GHz ocean wind scatterometer operated by the Indian Space Research Organization (ISRO). For the past two years ISRO has been collaborating with NASA/JPL and NOAA to refine the calibration of the OceanSAT-2 backscatter data. The goal of the collaboration is to extend the Ku-band scatterometer wind data record initiated by QuikSCAT. A crucial element of this effort has been the repointing of the QuikSCAT instrument to match the OceanSAT-2 viewing geometry (higher incidence angle). Although QuikSCAT ceased nominal operations in November 2009, its precisely calibrated backscatter measurements remain useful for cross-platform calibration. •OceanSAT-2 flyover of Cyclone Thane, Dec 28, 2011. •12.5 km MLE retrieval •Similar to QuikSCAT v2 KEY FOR QUIKSCAT VALIDATION PLOTS The different colors represent different QuikSCAT 12.5-km wind retrievals: Neural Network wind speed (red), nominal version 2 JPL L2B12 HDF product wind speed (blue) and version 3 JPL wind speed (green). Neural Net winds are optimized for high wind (>20 m/s) conditions with and without rain. Version 2 QuikSCAT data is optimized for speeds < 20 m/s and rain free conditions. Version 3 is optimized for < 20 m/s winds with and without rain. All grid cells with retrieved winds are presumed valid for purposes of comparison. No rain flagging is utilized. Dashed lines are binned by ground truth speeds (x-axis). Solids lines are binned by QuikSCAT speeds (y-axis). Number of samples in each bin are reported on the plots. Thin red solid and dashed lines indicate bin locations. For statistical reasons, random errors in ground truth winds lead to biases in the retrieved speed vs. truth curve. The sign of the bias differs depending upon which binning is used. A curve bisecting the area between the dashed and solid lines of the same color is a good proxy for QuikSCAT speed as a function of the true speed. Saturation results in truncation of the solid lines. References:  B. W. Stiles and R.S. Dunbar, “A Neural Network Technique For Improving The Accuracy Of Scatterometer Winds In Rainy Conditions.” IEEE TGARS, Vol 48 , No. 8, P 3114-3122, 2010.  B. W Stiles, S. Hristova-Veleva, et al, “Obtaining Accurate Ocean Surface Winds In Hurricane Conditions: A Dual Frequency Scatterometry Approach,” IEEE TGARS, Vol 48 , No. 8, P 3101-3113, 2010.  Ahmad, K. A., W. L. Jones, T. Kasparis, S. W. Vergara, I. S. Adams, and J. D. Park, “Oceanic rain rate estimates from the QuikSCAT Radiometer: A Global Precipitation Mission pathfinder”, J. Geophys. Res., 110, D11101, 26 pages, 2005. Hurricane Sandy (2012) as observed by OceanSAT-2 1 day before landfall as a tropical storm, 720 mile diameter of TS force winds Mean wind speed as a function of km from storm center (y-axis) and days (x-axis) from maximum intensity. Top = Cat 1-3 storms Bottom = Cat 4-5 Storm Movement Direction ABSTRACT We have produced a comprehensive set of tropical cyclone storm wind retrieval scenes for all ten years of the QuikSCAT mission and a sample set of storms observed by OceanSAT-2. The wind speeds were corrected for rain and optimized to avoid saturation at high winds using an artificial neural network method similar to that in  and . The QuikSCAT wind imagery and the quantitative speed, direction, and backscatter data can be obtained at http://tropicalcyclone.jpl.nasa.gov. The QuikSCAT wind speeds have been validated against best track intensity (i.e., maximum wind speeds), H*WIND tropical cyclone (TC) wind model analysis fields, and wind speeds from aircraft over flights (GPS drop wind sondes and step frequency microwave radiometer (SFMR) wind measurements). Storms from all basins are included for a total of 21600 scenes over the ten years of nominal QuikSCAT operations. Of these, 11435 scenes include the best track center of the cyclone in the retrieved wind field. Among these, 3295 were of tropical storms and 788, 367, 330, 289, and 55 were of category 1, 2, 3, 4 and 5 hurricanes, respectively, on the Saffir-Simpson Hurricane Wind Scale. In addition to the QuikSCAT TC winds, we have also processed wind fields from the Indian Space Research organization (ISRO) OceanSAT-2 satellite. OceanSAT-2 employs a scanning pencil beam Ku-band scatterometer with a design similar to QuikSCAT. JPL and NOAA have been working extensively with ISRO to aid in cross calibration between OceanSAT-2 and QuikSCAT. Toward this end the QuikSCAT instrument has been repointed in order to acquire data at the OceanSAT-2 incidence angles, and several meetings in India between the teams have taken place. The neural network that was trained on QuikSCAT data was used to retrieve OceanSAT-2 winds. The backscatter inputs to the network were transformed to match the histograms of the corresponding values in the QuikSCAT data set. • To date the ISRO/NASA/NOAA collaboration has resulted in: – More robust wind retrieval in low wind areas. – Improved wind accuracy as compared to numerical wind products and buoys. – Ongoing monitoring of calibration drift by comparison between QuikSCAT and OceanSAT-2 backscatter values. • The fruitfulness of the collaboration is further illustrated by the tropical cyclone winds on the following figures. – The OceanSAT-2 operational wind product is binned at 50 km with a conservative land mask employed to insure accurate winds. – JPL has retrieved winds at higher (12.5 km) resolutions and closer to the coast. – QuikSCAT neural network processing has been applied to OceanSAT-2 data. Acknowledgements • The work reported here was performed at the Jet Propulsion Laboratory, California Institute of Technology, and at the National Hurricane Center under contract with the National Aeronautics and Space Administration. • We would like to thank the Indian Space Research Organization for providing the excellent OceanSAT-2 scatterometer data set. • The work described here is funded by NASA’s Ocean Vector Winds program. • The website portal used to distribute the data set is funded by NASA’s Hurricane Science Research program. Acronyms: MAE = Mean Absolute Error, MBE = Mean Bias Error, FSP=Frequency of Superior Performance, TC= tropical cyclone, GPS=Global Positioning System, SFMR=Step Frequency Microwave Radiometer, H*WIND=Hurricane Research Division Real-Time Wind Analysis System Wind Fields, NHC = National Hurricane Center, JTWC=Joint Typhoon Warning Center, TS=Tropical Storm, ANN=Artificial Neural Network, CTD =Cross Track Distance, MLE = Maximum Likelihood Estimation Summary • Wind tropical cyclones speed fields have been Examples Maximum Speed Tracks – Isabel 2003 Note: Here blue is H*WIND Note: Here blue is H*WIND • NNet has slight positive bias at < 40m/s (H*Wind also positively biased) QuikSCAT–NHC Best Track Green denotes superior QuikSCAT retrieval Max. Wind (m/s) 34-kt Radii (km) •OceanSAT-2 flyover of Cyclone Thane, Dec 28, 2011. •Neural Network wind speeds using histogram matching of inputs. – Optimized for accuracy – Produced for all ten years of the QuikSCAT mission including over 5,000 observations of tropical storms and above. – Validated vs. H*WIND, SFMR, GPS drops, and best track wind speeds. – Made available online to the community at large in a browsable data base. • http://tropicalcyclone.jpl.nasa.gov • Ancillary data such as backscatter imagery and co-located rain information are also included. • We now begin the scientific exploitation of the QuikSCAT TC dataset. – Initial investigations will include: • Distribution of max winds as a function of storm movement and geographic location • Correlation of storm features with rapid intensification and de-intensification events • Storm size estimation and trend analysis 50-kt Radii (km) 64-kt Radii (km) • We also plan to produce similar datasets for OceanSAT-2 and ASCAT – Sample OceanSAT-2 fields have been produced and we are currently validating and optimizing them. – ASCAT fields are planned for the following year. • A paper describing the production and validation of the QuikSCAT data set is in preparation.