Andrea Schumacher 1, Mark DeMaria 2, John Knaff 3, Liqun Ma 4 and Hazari Syed 5 1 CIRA, Colorado State University, Fort Collins, Colorado 2 NOAA/NWS/NHC, Miami, FL 3 NOAA/NESDIS/STAR, Fort Collins, Colorado 4 NOAA/NESDIS/OSPO, College Park, Maryland 5 NOAA/NESDIS/OSPO, SSAI, College Park, Maryland UPDATES TO THE NESDIS TROPICAL CYCLONE FORMATION PROBABILITY PRODUCT AMS 31 st Conference on Hurricanes and Tropical Meteorology, San Diego, CA, 31 Mar 2014
Motivation & Past Work (24-hr TCFP) Extending to 48 hrs Algorithm Overview Examples from 2013 2013 Verification Future Work OUTLINE AMS 31st Conference on Hurricanes and Tropical Meteorology 2
Need for objective, real-time TC formation guidance Development of NESDIS TC Formation Probability (TCFP) product Estimates probability of TC formation in next 24 hrs 5 x 5 degree latitude/longitude grid boxes Originally just in Atlantic and E. Pacific (2006), extended to N. W. Pacific (2009), then made global (2013) Uses synoptic (GFS), convective (geostationary water vapor), and ocean (Reynold’s SST) predictors MOTIVATION & PAST WORK AMS 31st Conference on Hurricanes and Tropical Meteorology 3
Official forecast agencies provide 48-hr TC formation products NHC Tropical Weather Outlook JTWC TC Formation Alerts Mismatch between TCFP guidance and user needs Basic idea Use GFS forecast fields Satellite predictors most important at early times Challenges Pre-TC disturbances can move far in 48 hours Fixed-domain scheme can’t account for motion Disturbance-following schemes have advantage Balancing increase in size spatial averaging vs. loss of signal EXTENDING TCFP TO 48 HOURS AMS 31st Conference on Hurricanes and Tropical Meteorology 4
Discrete 5 x 5 degree grid boxes Some predictors would be divided between 2 grid boxes when disturbance at the edge of a grid box E.g., PCCD (% cold cloud coverage) Would lead to “probability pulsing” Introduced new spatial averaging 1-degree resolution domain 0-360E, 45S to 45N Value at each point is average over r = 500km (larger than 5 x 5 degree boxes) Still some “probability pulsing” most likely due to diurnal variations in convection SPATIAL AVERAGING AMS 31st Conference on Hurricanes and Tropical Meteorology 5 R = 500 km New product estimates the probability of TC formation within 500 km in the next 48 hours
Development data ATL: NEPA: NWPA: NIO: SIO: SHM: Limiting dataset = water vapor imagery Development done by basin AMS 31st Conference on Hurricanes and Tropical Meteorology 6 DATA PredictorSource VSHD mb Vertical Shear (kt) GFS RVOR mb Relative Vorticity (10-6s-1) GFS MLRH mb Relative Humidity (%) GFS THDV - Vertical Instability Parameter (°C) GFS HDIV mb Horizontal Divergence (10-6s-1) GFS MSLP - Mean Sea Level Pressure (mb) GFS TADV mb Temperature Advection (10-6°Cs-1) GFS BTWM - Cloud-cleared Brightness Temperature (°C) Water Vapor PCCD - WV pixels Colder than -40 °C (%) Water Vapor RSST - Reynold’s Weekly SST (°C) Reynolds DNST - Distance to Nearest Active TC (km) ATCF PLND - Land Coverage (%) CPRB - Climatological TC Formation Probability (%) ATCF
# non-genesis cases >> # genesis cases Screened out cases where one or more predictors were very hostile to genesis Screened out cases within 300 km of existing TC, over land For each predictor, found threshold for which less than 1% of formation cases occurred Examples for VSHD, RVOR, and MLRH: SCREENING AMS 31st Conference on Hurricanes and Tropical Meteorology 7 Predictor name and unitsN. Atl N.E. Pac N.W. Pac N.I.O.S. Pac*S.I.O. VSHD mb Vertical Shear (kt)>28.2>21.4>25.5>28.1>28.2>22.1 RVOR mb Relative Vorticity (10 -6 s -1 )<-1.0<-0.8<-0.4<0.1>0.1>0.0 MLRH mb Relative Humidity (%)<22.3<38.6<32.9<38.7<29.8<32.6
Linear discriminant analysis (LDA) is a method used to find a linear combination of predictors that separates two or more classes of events Here, classes are “genesis” and “non-genesis” LDA finds coefficients provide max separation in standard deviation space Top 4 contributing predictors by basin: LINEAR DISCRIMINANT ANALYSIS AMS 31st Conference on Hurricanes and Tropical Meteorology 8 N. AtlN.E. PacN.W. PacN.I.O.S. PacS.I.O. RVORCPRBRVOR CPRBRVORPCCDDNSTPCCD MSLPPCCDCPRB PCCDMSLP TADVDNSTMSLP
LDA yields a linear function (F) that provides binary classification Genesis or No Genesis However, F itself is not binary Separated genesis cases into 10 equal bins by F and computed genesis frequency ESTIMATING 48-HR PROBABILITY AMS 31st Conference on Hurricanes and Tropical Meteorology 9 F1F2# Gen Cases# TotalGen Freq (%) P(24h U 48h) = P(24h) + P(48h) + P(24h ∩ 48h)
AMS 31st Conference on Hurricanes and Tropical Meteorology 10 PRODUCT EXAMPLES 36 hours prior to 2 genesis events: Manuel (E Pac) Man-Yi (NW Pac)
AMS 31st Conference on Hurricanes and Tropical Meteorology VERIFICATION ATLNEPANWPANIOSIOSHEM Brier Skill Score (wrt Clim) Multiplicative Bias Max Probability Genesis (%)
EXCEL HAS FAILED ME THIS MORNING!!!! Basic takeaway from plots Under-prediction in all basins 2013 VERIFICATION - RELIABILITY AMS 31st Conference on Hurricanes and Tropical Meteorology 12
Tropical Weather Outlook includes 120-hr formation probability New in 2013 From a practical standpoint, extending TCFP to 120 hours is straightforward Challenge: Disturbance motion may be even bigger problem Increase size of spatial averages? Preliminary work underway Hybrid TCFP / disturbance-following genesis model E.g., TCGI, regional dynamical model TCFP can identify regions of highest formation likelihood Trigger disturbance-centric model FUTURE WORK AMS 31st Conference on Hurricanes and Tropical Meteorology 13
TCFP extended to 48 hours New spatial averaging scheme works well Increased overall probabilities 24hr product: 10-20% 48hr product: 20-50% Still systematically underpredicting, especially at higher probabilities Brier skill scores suggest higher skill than climatology SUMMARY AMS 31st Conference on Hurricanes and Tropical Meteorology 14
References QUESTIONS? AMS 31st Conference on Hurricanes and Tropical Meteorology 15