Recent Improvements in Statistical

1
Recent Improvements in Statistical Tropical
Cyclone Intensity Forecasts Mark
DeMaria Regional and Mesoscale Meteorology
Team NESDIS/ORA, Fort Collins, CO John Knaff,
CIRA Fiona Horsfall, TPC John Kaplan, HRD MHHS
Seminar May 30, 2001
2
OUTLINE

• Overview of the Statistical Hurricane Intensity
  Prediction Scheme (SHIPS)
• Modifications for 2001
• Steering-level pressure predictor
• Inclusion of relative humidity
• 5-day forecasts
• Future Improvements
• Inclusion of GOES data
• Improved oceanic input
• Ensemble intensity forecasts
• Rapid Intensity Index (J. Kaplan)
• Neural network prediction

3
Model Evolution

• October 1988
• J. Kaplan forecast shift at NHC
• 1991-1995
• Statistical model developed for Atlantic, run in
  near-real time, Climo SST along track from VICBAR
  forecast
• 1996
• Operational transition, east Pacific version
  added, real-time SST
• 1997
• Forecast predictors added, forecast skill
  demonstrated
• 2000
• LBAR replaced by NHC Official track
• Inland decay added

AOML
4
SHIPS Intensity Model (2000 Version)

• Predicts maximum wind change at 12, 24 72 h
• Predictors chosen using multiple linear
  regression
• Atmospheric predictors to 48 h using 10-level dry
  mesoscale model, with vortex filtered
• Forecast modified over land using Kaplan/DeMaria
  decay model

5
Forecast Skill During 2000 Season
Errors Relative to SHIFOR (Sample Includes
Depression Stage)

• SHIPS skillful 0-48 h
• D-SHIPS skillful 0-72 h
• Bi-modal errors
• Alberto, Isaac High Skill
• Joyce, Debby Little Skill
• Tendency for high bias for sheared storms

6
(No Transcript)
7
Time Series of Forecast and Observed 48 hour
Intensity-Change Forecasts
40 N
Alberto
Joyce
10 N
Observed
SHIPS
SHIPS
Observed
8
GOES Visible Image of Hurricane Bertha 1996
9
Low-Shear High-Shear
Radial/Height Tangential Wind Cross- Sections
From AMSU Retrievals
10
Hurricane Debby August 23, 2000
11
Track Deviations After Shearing
Hurricane Debby Aug. 2000
Actual Track
12
Can Storm Motion Be Used to Infer Storm Vertical
Structure?

• Motion can be estimated from satellite/recon
  fixes
• Vertical structure of environmental flow
  estimated from NCEP global model analysis
• Find vertical weights on environmental flow that
  minimizes difference between storm motion and
  vertically averaged environment

13
Three-Layer Examplewith All East-West Flow
P1300 hPa W1 U1 P2500 hPa W2
U2 P3700 hPa W3 U3
Uk environmental wind at each level Wk
general vertical weights C storm motion Mk
weights for mass-average (1/4, ½, 1/4)
14
Minimization Problem
U Vertical Mean Wind ? Wk Uk E ½ C-U
2 Find Wk to minimize E, Set ?E/ ?Wk0 U1U1
U1U2 U1U3 W1 CU1 U2U1 U2U2 U2U3
W2 CU2 U3U1 U3U2 U3U3 W3
CU3
Problems 1. Wk0 if C0 2.
Linear System is singular if U1U2U3
15
Modified Minimization Problem

• Add constraints to system
• Require ? Wk 1 (similar to mass weights)
• Dont let Wk deviate too far from mass weights

L ½ C-U 2 ?1- ?Wk ½? ?(Wk -
Mk)2 ?Lagrange multiplier, ?penalty term
parameter Resulting linear system is symmetric,
non-singular
16
New SHIPS Predictor PSL
Pressure Levels 200, 250, 300, 400, 500, 700,
850 hPa Find Wk to Minimize L ½(Cx-U)2
(Cy-V)2 ?1- ?Wk ½? ?(Wk - Mk)2 Define
Pressure of Steering Level PSL PSL
?WkPk Sample Mean610 hPa Standard
Deviation100 hPa Significant correlation with
intensity change 24-84 hr Explains 1/4 as much
variance as shear
17
Steering Pressure for Debby 2000
18
Normalized Weights for Debby 2000
19
Steering Pressure for Joyce 2000
20
Normalized Weights for Joyce 2000
21
50
10
30
(From J. Franklin)
22
Test of Relative Humidity Predictors

• Average RH from r200 to 800 km from NCEP global
  analyses
• RHLO 850 to 700 hPa vertical mean
• RHHI 500 to 300 hPa vertical mean
• RHLO significantly correlated with intensity
  change
• Explains 1/3 as much variance as shear
• RHHI not correlated with intensity change

23

• Extension of Forecast to Five Days
• 1954 First NHC Objective Forecasts to 24 hr
• 1961 Extended to 48 hr
• 1964 Extended to 72 hr
• 2001 Day 4 and 5 forecasts in test mode
• 2003 Operational Day 4 and 5 forecasts
• (if test is successful)
• Develop 5-day climatology and persistence model
  (SHIFOR) for evaluation of skill
• Extend SHIPS forecast to 5 days
• Test new SHIPS using 2000 season

24
5-Day SHIFOR Model

• Developed by A. Pike 1987
• Input Current max wind, 12 hr wind change, lat,
  lon, Julian Day, storm motion component
• Modifications
• Larger sample size 1967-1999
• Unnamed depressions eliminated
• Different predictors selected
• Forecasts to 5 days
• Test model using operational input 1995-2000

25
(No Transcript)
26
Test of New SHIFOR 1995-2000
27
Modifications for 5-Day SHIPS Forecasts

• Modify method for forecast of atmospheric
  predictors
• Current Run 10-level dry mesoscale model
  initialized with NCEP
  analyses with storm removed (valid only for 48
  h)
• New Predictors directly from AVN run out to
  5 days
• Modify averaging method for atmospheric
  predictors
• Current Area average (r 0 to 600 km)
• New Annular average (r200 to 800 km)
• Include unnamed depressions to increase sample
  size (1989-2000)
• Include PSL and RH predictors

28
How Accurate are AVN Forecasts of SHIPS
Predictors?

• 125 AVN forecasts of vertical shear predictor
  from 2000 Season
• Assume perfect track forecast
• Evaluate shear forecasts against corresponding
  AVN analyses
• Normalize errors by standard deviation of shear
  parameter
• Compare errors with persistence forecasts of
  shear parameter

29
Comparison of Old and New SHIPS/SHIFOR For 2000
Hurricane Season Sample
30
5-Day Track Forecasts for SHIPS Evaluation
Operational SHIPS uses 6 hr old NHC Official
Track forecast, shifted to current storm
position For 2000 test, extend from 72 to 120 hr
using CLIPER
31
2000 Intensity Forecast Errors
136 107 85 70
54
32
2000 Intensity Forecast Skill
33
Future Improvements

• Include predictors from GOES data
• Test ocean heat content as predictor
• Direct inclusion
• Emanuel coupled model
• Ensemble forecasts
• Neural Network prediction

34
CIRA Tropical Cyclone IR Imagery Archive

• Most Atlantic and East Pacific storms since 1995
• GOES-E, GOES-W, Meteosat storm-centered images
• Storm-centered, re-mapped to Mercator Projection
• gt 38,000 Atlantic, East Pacific images
• Include IR brightness temperature predictors to
  refine SHIPS forecasts
• Input Large-scale SHIPS intensity forecast,
  BT parameters
• Output Revised SHIPS forecast

15 Oct 00 UTC 15 Oct 12 UTC
16 Oct 00 UTC 16
Oct 12 UTC
Sample Images from Irene 1999
35
Extract Parameters from IR Imagery
Emily (40 kt) 25 Aug 99 1515 UTC Floyd (40
kt) 8 Sep 99 1515 UTC
36
0-100 km Brightness T vs. Max Wind
SHIPS Predictor Is T Deviation From Trend Line
37
SHIPS Forecast Improvement withIR Temp
Adjustment for 1995-99 Cases(Dependent Data
Evaluation)
38
Systematic Evaluation of GOES Imagery (J. Kossin,
CIRA Post Doc)
39
Altimetry Data from TOPEX/Poseidon
Source www.aoml.noaa.gov/phod/cyclone/data
40
TPC Real-Time Heat Content Analysis (Replace SST
with HC)
41
Use Emanuel Coupled Model to Estimate Upper Bound
on Intensity
Gloria 1985
(Emanuel 1999)
42
Additional Improvements

• Ensemble forecasting
• Run SHIPS for tracks from all other guidance
  models
• Run SHIPS for ensemble tracks (e.g., VICBAR
  system at HRD)
• Rapid Intensity Index
• Use SHIPS input to estimate probability of rapid
  intensification (J. Kaplan, HRD)
• Neural Network prediction
• J.-J. Baik demonstrated improvement for W.
  Pacific storms
• West Pacific SHIPS by end of 2001
• ONR project

43
SUMMARY

• New SHIPS predictors (PSL and RHLO)
• Directly using AVN forecast fields improves
  prediction
• 5-day version of SHIPS/SHIFOR available for 2001
• SHIPS skill to 84 h in 2000
• Continued work on forecast improvements