Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Analysis of Cobb-Predicted
Snow for Winters 2008-2010 at
Six Locations
Rachael A. WitterIowa State University
Mentors:
Dr. William A. Gallus, Jr., Christopher D. Karstens,
Logan R. Karsten
Iowa State University
Overview
• Background
• Motivation
• Data
• Methodology
• Results
• Conclusions
Background
• Microphysics that affect snowfall
accumulation (Cobb 2005)
• Snow crystal shape
• Snow density
• Snowfall algorithm to compute snow liquid
ratios (SLRs) that account for all of these
factors-Cobb method
Cobb Method
• Role of vertical motion in snow production
(Cobb 2005)
• Takes average SLR over snow events
• Accounts for not only temperature, but
vertical motion above snow production
zone (SPZ)
Other Snowfall Algorithms
• National Weather Service temperature
conversion table
• Scofield/Spayd diagrams (Scofield and
Spayd 1984)
• Base 10:1 SLR
• Do not take into account vertical motion
• Key for snow production and forecasting
Motivation
• Where does the Cobb method perform
well?
• Under what situations does the Cobb method
work well?
• Does one model perform better than the
other?
• Is there a correlation between forecast
statistics and seasonal snow total?
Data
• Raw BUFKIT data, post-processed using
Cobb method, for two winter seasons,
2008-2009 and 2009-2010
• Local COOP observations
• ASOS observations
BUFKIT Data
• Obtained from Pennsylvania State
University‟s BUFKIT archive
• Hourly (NAM) and three-hourly (GFS) data
• Forecasted quantities of SLR and snowfall
accumulation
• Summed into 24-hour periods to
correspond with COOP observations
BUFKIT Data
60 - 8454 - 78
48 - 7242 - 66
36 - 6030 - 54
24 - 4818 - 42
12 - 3606 - 30
0 - 24
Example
24-hour
Period
Runs of
NAM valid
for
Example
24-hour
Period
Runs of
NAM not
valid
Runs of
NAM not
valid
84 hours • Yielded 37 runs from
NAM and GFS for a given
24-hour period.
• Total of ~5300 24-hour
periods used for
verification per year, per
location.
COOP observations
• Obtained from NCDC
• Daily snowfall totals
• Water equivalent
• Average daily SLR calculated
ASOS Observations
• Used to remove any precipitation that fell
in a form other than snow from NCDC
totals
• Yielded corrected precipitation totals to
better represent a daily SLR
Six Locations
• Representative of three geographical regions:• Midwest
• Eastern coastal cities
• Great Lakes region
• 06Z-06Z forecasts used for KDSM
• 05Z-05Z used for others
• KMQT and KMSP originally included in study, missing ASOS data for 2 winter seasons
KALB – Albany, NY
KBOS – Boston, MA
KBUF – Buffalo, NY
KDSM – Des Moines, IA
KERI – Erie, PA
KJFK – New York City, NY
Method
• Contingency tables and forecast statistics
Yes No
Yes a b
No c d
Forecast
Observed
a = hit
b = miss
c = false alarm
d = correct null
Forecast statistics
•Probability of Detection (POD)
•a : number of correct warnings
•a + b : number of total observed
events
•False Alarm Ratio (FAR)
•c : number of falsely warned events
•a + c : total warnings issued
•Critical Success Index (CSI)
•a : number of correct warnings
•a + b + c : total warnings issued and
missed
Forecast statistics cont.
•Equitable Threat Score (ETS)
•a : number of correct warnings
•ch : number of correct forecasts
due to chance
Snowfall Thresholds
• Forecast statistics
computed for 3
thresholds of events
• Represent the models‟
ability to capture snowfall
events, and medium- to
high-end events
• 4 and 8 inch thresholds
not representative of
every event that occurred
Events greater than 0 inches
Events greater than 4 inches
Events greater than 8 inches
Results
• Forecast statistics for each threshold
• Correlation between forecast statistics and
seasonal snowfall total
KDSM KALB KBOS KBUF KERI KFJK KDSM KALB KBOS KBUF KERI KFJK
KDSM KALB KBOS KBUF KERI KFJK KDSM KALB KBOS KBUF KERI KFJK
Results for Zero Inch Threshold
• Only stations that receive lake-effect
snow as well as synoptic scale events
• Snow events correctly predicted 50% of
the time, falsely alarming only 25% of the
time.KBUF and KERI
POD 0.5-0.75
CSI 0.45-0.6
ETS 0.15-0.35
FAR 0.1-0.4
KDSM KALB KBOS KBUF KERI KFJK
KDSM KALB KBOS KBUF KERI KFJK KDSM KALB KBOS KBUF KERI KFJK
KDSM KALB KBOS KBUF KERI KFJK
Results for Zero Inch Threshold
• Locations situated on the northeastern
coast
• Snow events predicted correctly only 25%
of the time, falsely alarmed at least 50-
70% of the time.KBOS and KJFK
POD 0.2-0.3
CSI 0.1-0.2
ETS 0.1-0.2
FAR 0.4-0.7
KDSM KALB KBOS KBUF KERI KFJK
KDSM KALB KBOS KBUF KERI KFJK KDSM KALB KBOS KBUF KERI KFJK
KDSM KALB KBOS KBUF KERI KFJK
Results
• These locations could be considered
„inland‟, not influenced by lake-effect snow
or major body of water.
• Snow events correctly predicted and
falsely alarmed 35% of the time.
KALB and KDSM
POD 0.2-0.4
CSI 0.2-0.4
ETS 0.1-0.2
FAR 0.2-0.4
Results
• Majority of sites had higher values of POD,
CSI and ETS and lower FAR values for
winter 2009-2010
• Snow events better handled in 2009-2010
winter season
• GFS has higher FAR values, but also
higher POD, CSI and ETS values
• GFS captured snow events better than the
NAM but also “cried wolf” more often
Results
• Shows a positive correlation between POD, CSI and ETS and total amount of snow that fell at a location
• Negative correlation between FAR and seasonal snow total
• Correlations independent of model type and season
• Exception: Panel (d), more positive for 2009-2010 season compared to 2008-2009
• Improved performance during 2009-2010 season
Results
• POD, CSI and ETS values relatively low and FAR values relatively high• Models have a difficult time predicting medium-
and high-end events
• Some runs of the GFS captured 4-inch threshold events at all locations
• Several more of the NAM runs during 2009-2010 season at KALB, KERI and KJFK captured these events
• However, POD, CSI and ETS values still relatively low ( < 0.25)
Conclusions
• Locations affected by lake-effect snow had highest values of POD, CSI and ETS and lowest values of FAR
• GFS performed better than the NAM for both seasons
• Both models performed better for 2009-2010 season than 2008-2009 season
• Forecasters tasked with forecasting lake-effect snow are encouraged to use the Cobb method
Conclusions
• Forecasters should give slightly more
weight to GFS snowfall predictions
compared to NAM predictions
• The more snow a particular location
receives in a winter season, the better the
models do at predicting the event
Future Work
• Analyzing what differences come about by
using different methods of forecasting
snow
• Incorporating a larger sample size in a
well-confined geographic area, county
warning area (CWA)
Acknowledgements
• Dr. William Gallus, Jr.
• Chris Karstens
• Logan Karsten
• Daryl Herzmann
• Arthur Person
Questions