titan - identifying and tracking convective storms as objects 1 heuristic probabilistic forecasting...
TRANSCRIPT
TITAN - identifying and tracking convective storms as objects
11
Heuristic Probabilistic Forecasting Workshop
Munich, Germany30-31 August 2014
Mike Dixon1 and Alan Seed2
1National Center for Atmospheric Research, Boulder, Colorado2Centre for Australian Weather and Climate Research, Melbourne, Australia
Identifying and tracking storms as 3-D objectsIdentifying and tracking storms as 3-D objects
Why?– Nowcasting of severe weather– Storm studies– Climatology studies
What? Storm identification and properties– 3D identification based on a reflectivity threshold– The dual threshold technique– Identifying convective regions– Scaling objects appropriately for forecast lead time
How? Storm tracking– Storm overlap– Optimized matching– Estimating the motion– Correcting motion using optical flow field tracking
22
STORM IDENTIFICATIONSTORM IDENTIFICATION
33
Primary identification – find contiguous regions with reflectivity Primary identification – find contiguous regions with reflectivity exceeding a specified threshold (in this case 35 dBZ)exceeding a specified threshold (in this case 35 dBZ)
44
Secondary dual-threshold identificationSecondary dual-threshold identificationThis allows us to split up storms that have just ‘touched’This allows us to split up storms that have just ‘touched’
instead of actually mergedinstead of actually merged
Find regions at lower thresholdFind regions at lower threshold(in this case 35 dBZ)(in this case 35 dBZ)
Within those regions, find sub-regions at Within those regions, find sub-regions at the higher threshold (in this case 40 dBZ)the higher threshold (in this case 40 dBZ)
55
Dual-threshold identificationDual-threshold identificationDeciding which sub-regions to use andDeciding which sub-regions to use and
growing the sub-regions to the original outlinegrowing the sub-regions to the original outline
Find valid regions – i.e. those with significantFind valid regions – i.e. those with significantsub-regions at the higher thresholdsub-regions at the higher threshold
Grow the valid regions out to theGrow the valid regions out to theoriginal threshold boundariesoriginal threshold boundaries
66
Sometimes, for clarity and simplicity,Sometimes, for clarity and simplicity,it is preferable to represent storms as ellipses instead of polygonsit is preferable to represent storms as ellipses instead of polygons
Polygon representation for storms can be Polygon representation for storms can be complicatedcomplicated
Ellipses are easier on the eyeEllipses are easier on the eye
77
Handling mixed convective / stratiform situationsHandling mixed convective / stratiform situations
(a) Identify the convective regions within the radar volume(a) Identify the convective regions within the radar volume
(b) Constrain the storm identification to the convective regions only(b) Constrain the storm identification to the convective regions only
88
Example of scan with large regions of stratiform / bright-band,Example of scan with large regions of stratiform / bright-band,along with embedded convectionalong with embedded convection
99
Vertical section along line 1-2
Column-max reflectivity Bright-band Convection
Convectivearea
Stratiformarea
Titan tends to merge both the convective and stratiform regionsTitan tends to merge both the convective and stratiform regionsinto a single storm identification.into a single storm identification.
Therefore we need to isolate the convective regions.Therefore we need to isolate the convective regions.
1010
Merged convective andstratiform regions
The Steiner et. al (1995) method for convective partitioning was tested.The Steiner et. al (1995) method for convective partitioning was tested.However, it seemed to over-identify convective areas.However, it seemed to over-identify convective areas.
The Steiner method computes the difference betweenthe reflectivity at a point and the ‘background’ reflectivity
defined as the mean within 11 km of that point.
The method estimates the convective regions based on the reflectivity difference, determining the radius of convection as a function of the difference value.
1111
Stratiformarea
A modified method was developed, based on the ‘texture’ of reflectivity A modified method was developed, based on the ‘texture’ of reflectivity surrounding a grid point. surrounding a grid point.
‘Mean texture’ of reflectivity – mean over the column oftexture = sqrt(sdev(dbz2))
computed over a circular kernel 5km in radius,for each CAPPI height.
Convective (cyan) vs Stratiform (blue)partition computed by thresholding
texture at 15 dBZ
1212
Storm identification on all regionsStorm identification on all regionscompared with using the convective regions onlycompared with using the convective regions only
Storms identified using a 35 dBZ threshold.The storms include the regions of bright-band,
leading to erroneously large storm areas
Storms using the same 35 dBZ thresholdbut including only the convective regions
1313
Spatial scaling appropriate for longer-term nowcasts - Spatial scaling appropriate for longer-term nowcasts - investigating approaches for a 2-hour lead time.investigating approaches for a 2-hour lead time.
For nowcasts of 30 to 60 minutes, the scale of storms as For nowcasts of 30 to 60 minutes, the scale of storms as measured by the radars is appropriate.measured by the radars is appropriate.
For longer lead time forecasts, say 1 hour to 2 hours, we For longer lead time forecasts, say 1 hour to 2 hours, we want to identify and track only larger scale features, so we want to identify and track only larger scale features, so we need a technique to isolate those features.need a technique to isolate those features.
1414
From Seed (2003) event lifetime vs. spatial scaleFrom Seed (2003) event lifetime vs. spatial scalebased on computed median correlation time for precipitation eventsbased on computed median correlation time for precipitation events
1515A. Seed, J Appl Meteor, Vol 42, No 3, March 2003.
~50km
2 hrs
30 mins
~12km
From Germann et. al (2006), for an expected lifetime of 2 hours,From Germann et. al (2006), for an expected lifetime of 2 hours,the spatial scale should be between 32 and 64 km.the spatial scale should be between 32 and 64 km.
We choose to test with a spatial scale of 50 km.We choose to test with a spatial scale of 50 km.
1616Germann et. al, J Atmos, Vol 63, No 8, August 2006.
2 hr lifetime~50 kmspatial scale
30 min lifetime~8 kmspatial scale
Computing the spectrum of the reflectivity field shows the Computing the spectrum of the reflectivity field shows the spatial frequency of the scenespatial frequency of the scene
Reflectivity over a 1200km x 1200 km gridReflectivity over a 1200km x 1200 km grid 2D FFT-based spectrum of reflectivity field2D FFT-based spectrum of reflectivity field
1717
Filtering the spectrum allows us to isolate the larger scale Filtering the spectrum allows us to isolate the larger scale features.features.
Reflectivity filtered for features 50 km and largerReflectivity filtered for features 50 km and largerSpectrum filtered to retain features of 50 km Spectrum filtered to retain features of 50 km
scale and largerscale and larger
1818This includes the stratiform regions. What if we use this procedure on the convective areas only?
Applying the 50km spatial filter to the convective regions Applying the 50km spatial filter to the convective regions highlights the larger scale convective featureshighlights the larger scale convective features
Convective reflectivity regionsConvective reflectivity regionsConvective reflectivity filtered for features 50 km Convective reflectivity filtered for features 50 km
and largerand larger
1919
Comparing convective storm identificationComparing convective storm identificationat different scalesat different scales
Identification of smaller-scale convective Identification of smaller-scale convective features, minimum size 30 kmfeatures, minimum size 30 km22
Identification of features at the 50km spatial Identification of features at the 50km spatial scale, minimum size 2500 kmscale, minimum size 2500 km22
2020
STORM TRACKINGSTORM TRACKING
2121
The intial tracking step looks for overlaps betweenThe intial tracking step looks for overlaps betweenstorm envelopes at consecutive times.storm envelopes at consecutive times.
2222
The current storm outline is in white, while the storm locations from the previous scan are in yellow.If multiple storms overlap a single storm, the largest overlap is used.
The secondary tracking step optimizes the pairing between stormsThe secondary tracking step optimizes the pairing between stormsnot matched not matched by the overlap step.by the overlap step.
2323
We minimize a cost function designed toidentify likely matches based on location and storm size.
Cost = (Distance between centroids) * weight1 + (Difference in Vol1/3) * weight2
Identifying mergers and splitsIdentifying mergers and splits
MergersMergers occur if the forecast location of two or more occur if the forecast location of two or more storms lie within the observed envelope of a single storms lie within the observed envelope of a single
storm at the next time stepstorm at the next time step
SplitsSplits occur if the observed location of two or more occur if the observed location of two or more storms at the next time step lie within the forecast storms at the next time step lie within the forecast
envelope of a single stormenvelope of a single storm
2424
Example of a storm splitExample of a storm split
Storm envelope at time 1Storm envelope at time 1 Storm splits into 4 parts at time 2Storm splits into 4 parts at time 2
2525
Forecasts are based on the linear extrapolation of previous locations, Forecasts are based on the linear extrapolation of previous locations, weighted by age (higher weight for more recent scans).weighted by age (higher weight for more recent scans).
Storm size is forecast to grow or decay based on size history.Storm size is forecast to grow or decay based on size history.
Forecast using ellipses for small-scale storms.Forecast using ellipses for small-scale storms.The past locations are shown in yellow.The past locations are shown in yellow.
Forecast using polygons for larger-scale features.Forecast using polygons for larger-scale features.Each red outline is a 5-minute forecast, out to 30 Each red outline is a 5-minute forecast, out to 30
minutes.minutes.
2626
Similarly we can make forecasts of the storm area or volume, based on Similarly we can make forecasts of the storm area or volume, based on extrapolation of recent history.extrapolation of recent history.
And we can plot the time history of storm properties.And we can plot the time history of storm properties.
Forecasts of the storm area and volume, for each Forecasts of the storm area and volume, for each scan in the storm track lifetime.scan in the storm track lifetime. Time height profile of maximum reflectivity.Time height profile of maximum reflectivity.
2727
Sometimes we get tracking errors in challenging situationsSometimes we get tracking errors in challenging situations
2828
Example of radar scanning at 10 minute intervals, with fast moving storms.This can lead to problems with correct tracking.
Using a field tracked such as Optical Flow allows us to estimate the Using a field tracked such as Optical Flow allows us to estimate the ‘background’ movement of the echoes.‘background’ movement of the echoes.
2929
Example of tracking errors.Example of tracking errors.Neither storm in the NE quadrant is correctly tracked.Neither storm in the NE quadrant is correctly tracked.
3030
In this case no overlap occurs because of small storm sizes, long time between scans In this case no overlap occurs because of small storm sizes, long time between scans and fast movement.and fast movement.
By applying the Optical Flow vectors to storms with short histories,By applying the Optical Flow vectors to storms with short histories,we can improve both tracking the forecast accuracy.we can improve both tracking the forecast accuracy.
3131
How well did we do with forecasting the lineHow well did we do with forecasting the linefiltered using a 50 km spatial filter?filtered using a 50 km spatial filter?
Forecast at 23:05 UTC on 2014/06/08. Shown areForecast at 23:05 UTC on 2014/06/08. Shown are4 x 30 minute forecasts, to 2 hours.4 x 30 minute forecasts, to 2 hours.
2-hour verification at 01:05 UTC on 2014/06/09.2-hour verification at 01:05 UTC on 2014/06/09.This demonstrates that we can have some success This demonstrates that we can have some success
forecasting large-scale features at longer lead times.forecasting large-scale features at longer lead times.
3232
SEVERE STORM SEVERE STORM FORECASTINGFORECASTING
3333
Severe storm forecasting.Severe storm forecasting.
In the nowcasting role, a tool such as Titan appears to be In the nowcasting role, a tool such as Titan appears to be better suited to better suited to severe stormsevere storm forecasting than forecasting than precipitationprecipitation forecasting.forecasting.
One primary advantage of the ‘storm as an object’ One primary advantage of the ‘storm as an object’ approach is that severe weather attributes such as approach is that severe weather attributes such as mesocyclones, hail, heavy rain, turbulence and lightning mesocyclones, hail, heavy rain, turbulence and lightning can be ‘attached’ to the storm object and carried along with can be ‘attached’ to the storm object and carried along with the warning. the warning.
3434
Example 1: Thunderstorm Interactive Forecast Service (TIFS)Example 1: Thunderstorm Interactive Forecast Service (TIFS)Australian Bureau of MeterologyAustralian Bureau of Meterology
3535
TIFS takes TITAN results and presents them to the user ina suitable form for forecasting
Example 2: Automated Weather Alert System,Example 2: Automated Weather Alert System,State of Sao Paulo, BrazilState of Sao Paulo, Brazil
3636Warnings based on Titan forecasts are automatically generated for individual counties
Advanced dual-polarization radars can provide additional Advanced dual-polarization radars can provide additional information about the severe eventinformation about the severe event
RHI from SPOL S-band radar, showing RHI from SPOL S-band radar, showing severe storm vertical section (RHI)severe storm vertical section (RHI)
Same RHI showing the NCAR dual-polarization Same RHI showing the NCAR dual-polarization Particle ID product, with a hail core (yellow) and Particle ID product, with a hail core (yellow) and
heavy rain (red)heavy rain (red)
3737This additional information would enhance the utility of the forecast
Storm properties can include:Storm properties can include:
Geometric and reflectivity-Geometric and reflectivity-derived properties:derived properties:
– LocationLocation– AreaArea– Tops / vertical extentTops / vertical extent– Maximum reflectivityMaximum reflectivity– Rate of growthRate of growth– VILVIL
Properties derived from Properties derived from other applications:other applications:
– Presence of hailPresence of hail– Hail sizeHail size– Heavy rainHeavy rain– Precipitation fluxPrecipitation flux– Presence of lightningPresence of lightning– Lightning rateLightning rate– Presence of mesocyclone Presence of mesocyclone
(supercell flag?)(supercell flag?)– TornadoTornado– TurbulenceTurbulence
3838
Storm Studies and ClimatologyStorm Studies and Climatology
A tool such as Titan can help to condense vast quantities of A tool such as Titan can help to condense vast quantities of radar data into a more manageable size.radar data into a more manageable size.
Having done so, it is possible to study both individual storm Having done so, it is possible to study both individual storm cases, and to analyze long periods to determine the cases, and to analyze long periods to determine the climatology of convection and severe weatherclimatology of convection and severe weather..
3939
You can summarize an event by displaying all of the tracks You can summarize an event by displaying all of the tracks making up that event.making up that event.
4040Magenta lines show the movement vectors of the storm centroids.
Yellow ellipses depict the storm extent.
You can have some fun and animate the event as it unfoldsYou can have some fun and animate the event as it unfolds
4141
Example of a Climatology StudyExample of a Climatology Study
The following 2 slides show the results of a climatology The following 2 slides show the results of a climatology study of the Madden-Julian Oscillation (MJO), using data study of the Madden-Julian Oscillation (MJO), using data taken by the NCAR S-Pol radar in the Maldives during the taken by the NCAR S-Pol radar in the Maldives during the DYNAMO field experiment, October 2011 to January 2012. DYNAMO field experiment, October 2011 to January 2012.
These results are thanks to Sachin Deshpande, of the Indian These results are thanks to Sachin Deshpande, of the Indian Institute of Tropical Meteorology (IITM), Pune, INDIAInstitute of Tropical Meteorology (IITM), Pune, INDIA
4242
Storms were identified & tracked by TITAN throughout its lifetime. The majority of storms were of short durations, i.e. predominantly less than 2 hours.
Active MJO periods showed long lived storms compared to suppressed MJO periods.
Short lived storms are due to shallow and isolated convection associated with SubMCS.Long lived storms are due to organized convection associated with MCSs.
DYNAMO field projectDuration of storms during Active and Suppressed phases of MJO
1 3416 6831 102460
4
8
12
16
20
Sto
rm D
ura
tio
n (
hrs
)
Storm Number
15 Oct To 27 Oct
01 October To
14 October
1 3646 7291
15 Nov To 30 Nov
Storm Number
01 November To
14 November
1 4196 8391
15 Dec To 30 Dec
Storm Number
01 DecemberTo
14 December
0 2 4 6 8 10 12 14 16 18 201
10
100
1000
10000
Fre
qu
en
cy
Co
un
t
Storm Duration (hrs)
October
0 2 4 6 8 10 12 14 16 18 20
Storm Duration (hrs)
November
0 2 4 6 8 10 12 14 16 18 20
Storm Duration (hrs)
December
Source: Personal communication, Sachin Deshpande,Indian Institute of Tropical Meteorology (IITM), Pune,
INDIA
Frequency of occurrence of storms of different scalesduring DYNAMO
Larger storm areas and top heights for MCS (during 15-30 Oct) as compared to SubMCS (during 1-14 Oct).
Contribution of small sized storms to total population is more compared to large size storms but the contribution of MCSs is maximum to the total storms when area is considered.
Shallow convective echoes (SCEs)Etops lower than 1 km below 0o level
Deep convective cores (DCCs)ETops : at least 8 kmIncludes young and vigorous cellswith strong updrafts
Wide convective cores (WCCs)Horizontal area: at least 800 km2
Contains region where intense individual cells merge together
01 Oct 14 Oct 2011 15 Oct – 27 Oct 2011
0
5
10
15
20
25
1 10 100 1000 10000
Count
Echo
Top
Hei
ght (
km)
( c ) 01 Oct - 14 Oct
0
5
10
15
20
25
1 10 100 1000 10000
Count
Echo
Top
Hei
ght (
km)
(d) 15 Oct - 30 Oct
0
100
200
300
400
500
600
700
800
1 10 100 1000 10000
Count
Stor
m A
rea
(km
2)
(a) 1-14 Oct
0
800
1600
2400
3200
4000
4800
5600
6400
7200
8000
1 10 100 1000 10000
Count
Stor
m A
rea
(km
2)
(b) 15-30 October
Source: Personal communication, Sachin Deshpande,Indian Institute of Tropical Meteorology (IITM), Pune,
INDIA
Suppressed phase Active phase
4545
THANK YOUTHANK YOU