Wind Shear Detection with Pencil-Beam Radars

M.W. Menitt, D. Klingle-Wilson, and S.D. Campbell

Wind Shear Detection withPencil-Beam Radars

Abrupt changes in the winds near the ground pose serious hazards to aircraft dUringapproach or departure operations. Doppler weather radars can measure regions ofwinds and precipitation around airports, and automatically provide air traffic controllers and pilots with important warnings of hazardous weather events. Lincoln Laboratory, as one ofseveral organizations under contract to the FederalAviation Administration, has been instrumental in the design and development of radar systems andautomated weather-hazard recognition teclmiques for this application. The TerminalDoppler Weather Radar (TDWR) system uses automatic computer algorithms toidentify hazardous weather signatures. TDWR detects and warns aviation users aboutlow-altitude wind shear hazards caused by microbursts and gust fronts. It alsoprovides advance warning of the arrival ofwind shifts at the airport complex. Extensiveweather radar observations, obtained from a Lincoln-built transportable testbed radarsystem operated at several sites, have validated the TDWR system. As a result, theFederalAviation Administration has issued a procurement contract for the installationof 47 TDWR radar systems around the country.

Low-altitude wind shear has been a majorcause ofU.S. air carrier fatalities in recentyears.A 1983 National Research Council study identified low-altitude wind shear as the cause of27aircraft accidents and incidents, with a total of488 fatalities, between the years of 1964 and1982 [1]. Since then, the National Transportation Safety Board investigated at least threeadditional wind shear incidents, including thecrash of Delta Flight 191 at Dallas-Fort Worthon 2 August 1985, which claimed another 137lives.

The Federal Aviation Administration (FAA)initiated the Terminal Doppler Weather Radar(TDWR) program in 1983. The purpose of thisprogram is to procure a ground-based Dopplerweather radar system for the automatic detection of wind shear hazards to aircraft, and toprovide real-time warnings to air traffic controllers for relay to pilots. Future enhancements tothe system will allow direct communication ofwarnings to pilots via radio data link.

Two prfmaryforms oflow-altitudewind shearpresent hazards to aircraft: the microburst andthe gust front. The microburst is a strong short-

The Lincoln Laboralory Journal, Volume 2, Number 3 (J 989)

lived outflow produced by strong thunderstorms. In certain regions of the country (forexample, the Denver, Colo.. area), even benignlooking clouds can produce microbursts. Gustfronts, also produced by thunderstorms, aredefined as the boundary between the outflowfrom the storm and the ambient winds. Whilemicrobursts are short in duration (10 to 20 min)and occur with little warning, gust fronts canlast for hours as they travel away from thegenerating storm system.

The detection ofwind shear hazards presentschallenges iIi three basic areas:(1) the reliable measurement of horizontal

winds, especially in low-signal, high~c1utter environments;

(2) the automatic identification and classification of the hazardous weather signaturesin thewindfieldmeasurements; and

(3) the assessment of the hazard level posedto aircraft, along with the effective communication ofthe hazard level to air trafficcontrollers and pilots, in a manner easilyunderstood by these non-meteorologistusers.

483

Merritt et aI. - Wind Shear Detection with Pencil-Beam Radars

The emphasis in the weather radar program atLincoln Laboratory has been on the first two ofthese areas. The effort has included the construction of a TDWR testbed radar system tocollect test data from several meteorologicalenvironments across the country. The algorithm development cycle has included bothoff-line testing and real-time operationaldemonstrations.

Phenomenology of Microburstsand Gust Fronts

Downdrafts within storms generate microbursts and gust fronts. Through a variety ofprocesses. including evaporative cooling andprecipitation loading. negatively buoyant airwithin a storm descends to the ground as a

downdraft. Upon reaching the surface. thedowndraft spreads out horizontally from a divergent center and produces a pool of cold airknown as the outflow (2).

Microbursts are formed when the divergencebeneath the downdraft reaches a specified intensity. namely a change in wind speed ofgreater than 10 m/s over a distance ofless than4 kIn. Figure 1 illustrates a potentially hazardous encounter of a microburst by an aircraft onfinal approach. Upon entering the microburst.an aircraft first experiences an increase in headwind. This increase causes the aircraft to flyabove the glide slope. The pilot. who is oftenunaware of the microburst. may attempt toreturn to the glide slope by reducing air speedand angle of attack. As the airplane continuesthrough the microburst. it encounters a strong

500

400

Qi 300Q)u..

200 ~ ~... ........... ....... ---100

a •

Fig. 1-Example of an aircraft encounter with a microburst. The spreading winds from a strong downdraft form themicroburst outflow. A penetrating aircraft first experiences an increase in headwind, followed rapidly bya downdraft, andfinally a tail wind. The loss in altitude across the event may result in ground impact.

484 The Lincoln Laboratory Journal. Volume 2. Number 3 (1989)

downdraft and then a tail wind, which results ina loss of lift. The airplane falls beneath the glideslope and the pilot must nowincrease power andangle of attack to bring the plane back to theglide slope. The aircraft, which requires a finiteamount of time to respond to the controls,crashes if it is too close to the ground torecover.

In most cases, the rain carried in the microburst provides a high-reflectivity signature. Ifthe environment beneath the cloud base is dry,however, the precipitation evaporates before itreaches the ground, which results in a lowreflectivity signature. The evaporation also intensifies the microburst by further cooling thedowndraft. The shape, size, and strength of themicroburst shear region often evolve rapidly,particularly during the early growth of the outflow. When a strong downdraft first impacts thesurface, it can change from a weak surfaceoutflow to a strong microburst in a time periodof 1 to 2 min. Most microbursts reach peakintensity and then decay in a time period of 10to 20 min. More complicated events may pulseso that the microburst strength peaks and decays several times before the event dissipatescompletely.

The gust front is formed at the leading edge ofthe outflow, where the cooler air from thedowndraft meets the warmer environmental air(Fig. 2). As the cool, dense outflow spreads out

...-~- . . .-

f/>'~'/,;" .

J'6-I

Downdraft

Fig. 2-An illustration of thunderstorm downdraft and outflow. The pool of cold air spread out beneath the storm iscalled the outflow. The gust front is the leading edge of theoutflow where environmental and outflow winds converge.

The Lincoln Laboratory JournaL. Volume 2. Number 3 (1989)


GustFront

----~Glide Slope

(a)

GustFront

-~;2l~Glide Slope

(b)

Fig. 3-Example ofan encounterbetween a gust front anda landing aircraft. (a) As the aircraft flies into an outflow, itexperiences an increase in head wind. (b) As the aircraftexits the outflow, it experiences a decrease in tail wind. Bothsituations result in a performance gain that causes theplane to rise above the glide slope. Strong turbulence mayoccur within the shear zone.

into the warm, less dense environmental air,horizontal convergence is produced at the outflow edge. During the passage of an outflow, anobserver on the ground would experience achange from warm light winds to cool gustywinds (thus the name gust front). Gust frontscan last several hours, can propagate hundredsof kilometers from the parent storm, and commonly exist long after the parent storm hasdissipated.

Gust fronts impact aviation in two ways: as ahazard to aircraft in flight and as a mechanismfor creating delays in the terminal area. The gustfront represents a transition zone of finite widthbetween the cold (outflow) and warm (environmental) air masses. The strong surface winds,updrafts, and turbulence that exist within thiszone may prove hazardous to an aircraft duringtakeoff and landing. Figure 3 illustrates the

485


conceptual model of an encounter between anaircraft and a gust front. In Fig. 3(a), the aircraftexperiences an increasing head wind as it entersthe outflow. In Fig. 3(b), the aircraft experiencesa decreasing tail wind as it exits the outflow. Theupdrafts and turbulence asociated with the gustfront often result in an uncomfortable ride. Bothpenetrations also result in a gain in aircraftperformance that can cause an arriving plane torise above the glide slope and overshoot thelanding zone on the runway.

The passage of a gust front over an airportoften produces a long-lasting change in theprevailing winds. The sudden wind shift canseriously impact the operation of the airport (forexample, a change in wind direction wouldrequire runway changes). With advance notification of a winJ shift, air traffic controllers canplan operations more effectively. The identification of the wind shift requires an estimation ofthe wind speed and direction behind the movinggust front boundary, typically in low-reflectivityregions.

Development of the TDWR System

In the late 1970s the FAAjoined the NationalWeather Service and Air Weather Service in thedevelopment of a joint-use Next GenerationWeather Radar system (NEXRAD) to providecontiguous weather radar coverage over thecontinental United States. The principal role forthe NEXRAD system was to detect and monitorlarge severe storms, such as those that producetornadoes. Late in the development ofNEXRAD,scientific studies aimed at understanding themicroburst wind shear phenomenon began toidentifY the characteristics ofthe hazard, as wellas the radar system requirements for detectionand warning of the hazards.

After the crash ofPan American Airlines flight759 at New Orleans International airport (9 July1982), the National Research Council conducted an in-depth examination of the aviationwind shear situation [1]. The study concludedthat the FAA should develop a weather radarsystem that would be dedicated to identifying

486

wind shear hazards in the airport terminalenvironment. While the technical characteristics of the NEXRAD radar were comparable tothose deemed necessary for the terminal windshear protection task, the joint-use mission ofthe NEXRAD radar prevented it from being sitedand scanned in a manner appropriate to theterminal problem. The Terminal DopplerWeather Radar program was thus initiated,with the intent of using an enhanced NEXRADlike radar tailored to the task of wind sheardetection.

Lincoln Laboratory, already involved in thedevelopment ofNEXRAD products for the FAA,began construction of a testbed radar systemsimilar to NEXRAD to act as a data-collectionand algorithm-evaluation facility for TDWR [3]:The ongoing scientific studies of wind shearphenomena, particularly the JAWS project [4]and the investigation of the crash of Pan Amflight 759 (both in 1982), led to a reorientation ofthe program at Lincoln Laboratory to focus onterminal wind shear hazards.

The testbed radar and algorithm development centered on microburst and gust frontdetection, and the first attempts at microburstdata collection were made at MIT in 1983. Todate, the testbed radar has been transported toseveral different sites where radar observationswere performed on hundreds of microbursts.The radar was sited in Memphis, Tenn., dUring1985, and then moved to Huntsville, Ala., formeasurements dUring 1986. The radar observations made in Memphis and Huntsville providedthe first operationally oriented study of microbursts in the humid southeast environment.Following the operations in Huntsville, the radarwas relocated to Denver dUring 1987 and 1988.In 1988 the testbed was used to perform acomplete Operational Test and Evaluati~m

(OT&E) oftheTDWRsystem. This OT&E (whichCongress reqUired to allow the procurement ofthe TDWR system to proceed) demonstratedthat the TDWR, as specified, could provide auseful wind shearwarning function for air trafficcontrol.

The TDWR procurement was awarded to the

The Lincoln Laboratory Journal. Volume 2. Number 3 (1989)

Raytheon Corporation in November of 1988. Thefirst operational TDWR deliveries are scheduledto begin in volume in 1993. To obtain protectionat critical airports before that date, the FAAcontracted the NEXRAD supplier (Unisys) toadapt the initial 17 FAA-owned NEXRAD radarsto perfonn as interim TDWR systems. This terminal variation of the NEXRAD radar will include only software modifications to lUn theTDWR algorithms. Initial tenninal NEXRADdeliveries are scheduled to begin in 1990.

While the primary mission of the testbedradar system has been the development andevaluation of TDWR-related radar techniquesand detection algorithms, the measurementsfrom this radar have also been supplied toseveral national scientific studies on microburstphenomenology and stonn-scale meteorology.During 1986, the radar operated as part of theCOHMEX experiment [5], and the radar observations from Denver in 1987 were supplied tothe CINDE program (6). The ongoing data-collection program with the testbed radar hasgenerated the most extensive collection of microburst and gust front radar data available,and has been of considerable benefit to themeteorological community.

While the Lincoln Laboratory testbed provided the radar observations for the development and testing of wind shear detection algorithms, the Research Applications Program atthe National Center for Atmospheric Research(NCAR) actively addressed the user-interfaceissues and requirements for TDWR. NCARfonned a TDWR User Working Group composedof pilots, pilot group representatives, FAA airtraffic controllers, administration representatives, and scientists and engineers working onthe TDWR development. This group examinedthe various requirements ofthe TDWRusers (forwhich the system represented a totally newcapability), especially the types of infonnationthat pilots and controllers wanted and how thisinfonnation would be presented. The UserWorking Group, which met three times over thecourse of the system development, was responsible for the specification ofthe current concepts



ofTDWR display and warning (7).

Requirements for the TDWR

The primary goal of the TDWR is to detectmicroburst wind shear hazards and to providewarnings of these hazards to air traffic controlpersonnel for relay to pilots. The wjnd shearfrom a microburst presents a significant hazardto an aircraft only dUring low-altitude flight,such as approach and departure operations.The User Working Group concluded that microburst warnings were relevant to pilots while theywere operating at altitudes below 1,000 ft aboveground level (AGL). Given the nominal 3° glideslope used by most aircraft arrival operations,this altitude limit requires that warnings mustbe provided out to 6 kIn from the end of eachrunway.

The microburst detection must also betimely. The TDWR is required to provide pilotswith a warning ofhazardous wind shear at least1 min prior to their encounter with the hazard.To meet this time requirement, theTDWRmicroburst-recognition algorithm is designed to provide an initial warning when the surface outflowis still weak (10 mls or less). Since the TDWRupdates the microburst warning once per minute, the algorithm provides a timely warningwhen the outflow reaches a hazardous level.

A major economic benefit of TDWR is thedetection and prediction ofwind-shift lines. Thearrival of a wind-shift line at an airport has asignificant impact on operations, since it cancause a change in the active lUnway configuration. For example, the direction of takeoffs andlandings for a given runway might have to bereversed because of a wind shift. This reversalwould reroute approaching aircraft and forcedeparting aircraft to taxi to the other end of therunway.

To provide a useful product for lUnway configuration planning, the TDWR is required togenerate a 20-min prediction of the arrival of awind shift at the airport. To meet this prediction requirement, the TDWR must detect gustfronts to a range ofabout 60 kIn from the airport.

487

Merritt et aI. - Wind Shear Detection with Pencil-Beron Radars

The TDWR must also detect gust front windshears that impact the runways. Gust frontsin the terminal area, which are considered lesshazardous than microbursts, create turbulenceand substantial crosswinds that can affect

aircraft performance.

Data Acquisition in the TDWR

The TDWR antenna scan strategy represents

. ......:

o 2 mi

•

----- --

•

•

" .

•

•

••

•

••• •

j~i~ : i• 'I . ~--::.: ~ ..! ~';::-." '. -..-· .

..

•

•

..

•

•

.'..

. ~~.,

•. ',.

; ..l\~I-+.~ ....-:-o;:.---:::-,/ (' .

-'

-. --17---/-\

,.'"

"

" ./

Fig. 4-Windshear coverage region for the TDWR Operational Test and Evaluation atStapleton Airport in Denver. The figure shows the airport runways and the oval-shapedradar coverage region that extends 6 km beyond the end of each runway. The insetphotographs illustrate the Lincoln Laboratory testbed radar (15 km southeast of theairport), the University ofNorth Dakota weather radar (northeastofthe airport), andoneof the mesonet surface wind-measurement stations. The red dots represent mesonetstations, while the blue dots represent locations for the FAA LowLevel Wind ShearAlertSystem anemometer stations. The measurements recorded by these support instruments during the evaluation aided in the analysis ofthe meteorologicalevents observedby the testbed radar.

488 The Lincoln Laboratory Journal. Volume 2. Number 3 (1989)

Merritt et aI. - Wind Shear Detection with Pencil-Bearn Radars

300270240210180150

Time (s)

mb

120906030

40

35

30

C;a> 25:s-a>Clc« 20c0.~

15>a>W

10

5

00

Fig. 5-Antenna elevation-angle sequence used for TOWR operational testandevaluation in 1988. Thefirst scan in the sequence (shown in red) is a full-circle scan at an elevation of 0.5°. This scan is usedjointly by the gust front and microburst algorithms. A second full-circle scan at an elevaion angle of 1°is used by the gust front algorithm alone. The microburst algorithm further requires surface scans at1-min intervals. These scans (green) cover a 120° sector over the airport at an elevation angle of0.4°.The PRFselection algorithm requires a long-range, 10w-PRFfull-circle scan (blue). The remaining scansare used by the microburst precursor algorithm.

a complex trade-off of conflicting requirementsby the various components of the detection anddata-conditioning algorithms. The surface-divergence detection process, which is the primary component of the microburst-hazarddetection system, requires updates of surfaceinformation at least once per minute. The feature-recognition algorithms for structuresaloft require updates of the full storm volume,necessitating numerous scans at different elevation angles. Both of these detection algorithms require data only from the TDWR coverage region surrounding the airport and henceallow the radar to scan sectors on the orderof 1200 in width. Other algorithms, such as thegust front detection and automatic pulse repetition frequency (PRF) selection algorithm, require full-circle scans to identify gust frontsand storm cells in all directions (8).

Figure 4 illustrates the coverage region cho-

sen for the TDWR OT&E in Denver dUring thesummer of 1988. The center of the figure showsthe airport runways, with the wind shear coverage region (which extends approximately 6 kInoff the end of each runway) shown as an ovaloutline. The figure also shows the location ofthetestbed radar and the supporting sensors. Figure 5 depicts the elevation angles chosen for thisscan strategy. The elevation-angle sequencewas chosen to meet the need for a I-min surfaceupdate rate for the microburst algorithm, tomeet the need for two full-circle surface scansevery 5 min for the gust front algorithm, and toprovide a worst-case vertical spacing from 1 kInup to 6 kIn above ground level for observation ofmicroburst features aloft.

This combination of scans provides frequentobservations of the surface windfield, includingone surface scan each minute, full-circle sUrface scans for gust front detection every 5 min,

The Lincoln Laboratory Journal. Volume 2. Number 3 (1989) 489


and a top-to-bottom volume coverage every 2.5min. The coverage of the volume above theairport area is important for the detection ofmicroburst features aloft, particularly for thedetection of storm reflectivity cores. The features aloft generally develop at altitudes of 5 to6 km AGL and take several minutes to descendto the surface [9, 10). The volume update rateachieved by the TDWR scan strategy allowsobservation ofthese descending core features atleast twice during their descent, to provide earlywarning of incipient microbursts. The TDWRsystem continually repeats the scan strategy described above, and supplies the resulting radarmeasurements to the wind shear detectionalgorithms.

Because several forms ofdata contaminationmay affect the radar measurements, specificdata-correction and data-editing proceduresare applied prior to algorithm operation. Themajor forms of data contamination are groundclutter residue, moving clutter, range aliasing,and velocity aliasing. Although the TDWR radarsystem must provide at least 50-dB groundclutter rejection [11], the residue from strongclutter targets such as mountains or nearbybuildings can result in data contamination. TheTDWR employs a static ground-clutter residu~map to edit the contaminated data regions [12).The maps are created on clear days so thatsignificant echoes can be associated with clutterresidue and compared with operational measurements. When the operationally measuredreflectivity levels compare to the clear-day echolevels, the data values are censored as contaminated. This approach edits data values coincident with strong clutter targets not removed bythe high-pass filtering performed in the signalprocessor, except when sufficient weatherpower overcomes the clutter contamination.Moving clutter targets such as birds, airplanes,or automobiles are not rejected by the groundclutter filtering, but are removed by a pointtarget rejection filter. The point-target rejectionfilter attempts to identify and delete spatiallyisolated discontinuities in the data. Careful filterdesign avoids editing the relatively small-scalemicrobursts that the system attempts to detect.

490

The effects of range aliasing and velocityaliasing are intrinsic to the pulsed Dopplerweather radar sensor, since no single PRF canprovide unambiguous measurements in bothrange and velocity (see the box titled "RangeAliasing and Velocity Aliasing with Pulsed Doppler Weather Radar"). The TDWR uses dynamicPRF selection logic to choose a PRF value thatminimizes the amount ofdistantweather echoesto be range aliased into the significant observation regions for both microburst and gust frontdetection. Periodic low-PRF scans identifY thelong-range weather echoes, and the PRF for theoperational scans is adjusted to handle theevolving out-of-trip weather pattern [13).

A complex combination of techniques addresses the velocity-aliasing problem. TheTDWR testbed (which operates at S-band) utilizes software techniques to detect and correctvelocity errors caused by aliasing (14). Thesetechniques rely on the spatial continuity of thevelocity field, and attribute extremely large gateto-gate velocity changes to aliasing. The production TDWR system will operate at C-band andhence be subject to more velocity aliasing. TheTDWR contractor has designed a dual-PRFwaveform to assist in the velocity de-aliasingprocess on those scans used for the wind-shiftestimation algorithm (which is particularlysensitive to velocity errors).

Microburst Detection

The TDWR microburst detection process provides reliable and timely warnings by employingradar observations both at the surface and aloft.The surface velocity measurements first identifyregions of divergence. If divergence regions arefound, strength and temporal-continuitythresholds determine if the regions are significant enough to warrant the generation ofmicroburst alarms. The detection of storm featuresaloft relaxes the strength and continuity requirements for alarm generation, and increasesthe likelihood that surface divergence regionswill be considered as microbursts.

Figure 6 illustrates the three stages of processing modules in the detection algori thm: fea-



Range Aliasing and Velocity Aliasingwith Pulsed Doppler Weather Radar

R

Pulsed Doppler weather radars use the pulse-to-pulse phasechange of the received signal toestimate the radial velocity of thescatterers in the target samplevolume. The phase-samplingapproach results in a maximumunambiguous velocity beyondwhich the measured velocity isambiguous (Le. the phase changeover a pulse period is greater thanone-half wavelength). This maximum velocity. or Nyquistvelocity.is given by

}.,Pv =-

4

where Ais the radar wavelengthand Pis the radar pulse repetitionfrequency (PRF). The pulsed nature of the radar operation alsointroduces a maximum unambiguous range.

c

2P

(where c is the speed of light)beyond which the reflections fromtargets are incorrectly located in

range. The radar wavelength.typically a fixed frequency forweather radars. couples thesetwo aliasing effects as follows:

clRV =

8

The PRF (a variable operatingparameter) adjusts the balancebetween the two effects. Figure Aillustrates the set of range/velocity aliasing limits available toweather radars operating at various wavelengths (the LincolnLaboratory testbed radar operates at S-band: the productionTDWR will operate at C-band).

To reduce the possibility ofrange-aliased echoes from distant storms. the unambiguousrange of a weather radar must beat least 450 krn. where the radarhorizon will extend above the tallest storm cells. To reduce the possibility of velocity aliasing. a velocity measurement range of atleast ±35 m/s must be availablefor severe storm applications.Figure A illustrates this area of

complete aliasing avoidance.where no range or velocity aliasing is likely to be experienced.The area is well outside of theachievable limits with either S- orC-band radars. As a result. practical weather radars must contend with both range- and velocity-aliasing effects.

Several techniques are available to reduce the data contamination effects from range andvelocity aliasing in the weatherradar context. The distributednature of the weather targets(compared to more common discrete-target radar systems) isboth a complication and an advantage for unfolding velocitymeasurements. Several softwarealgorithms were developed thatmake use of the spatial continuity of weather measurements forvelocity unfolding. Radar-waveform processing techniques canalso deal with the aliasing problem. but careful system designis required to maintain a highlevel ofground-clutter rejection.

Radar PRF (Hz)

1000 500 250

600

10 cm

450300

Operating Area forComplete ObscurationAvoidance •

150

60Ui'--.s~ 45'(30Qi>en 30::::l0::::lCl:0 15Ectlc

=>0

0

Unambiguous Range (km)

Fig. A-Trade-offbetween unambiguous range andvelocity intervals.Each curve indicates the feasible operatingpoints for the corresponding radar wavelength. The TDWR radar operates at C-band (5cm); theLincoln Laboratory testbedradar operates atS-Band (1 0 cm). Neithersystem is capable ofoperating free ofboth range andvelocityaliasing.



Fig. 6-The three microburst algorithm processing modules: feature extraction, vertical integration,and microburst recognition. The feature-extraction modules identify two-dimensional regions ofconvergence, divergence, rotation, andreflectivity. The vertical-integration modules assemble theseregions into three-dimensional features aloft, such as a reflectivity core or convergence aloft. Themicroburst-recognition modules use these features aloft to identify microburst precursors and to aidthe recognition of microburst surface outflows.

ture extraction, vertical integration. and microburst recognition. The feature-extraction modules identifY two-dimensional regions of shear(divergence. rotation. and convergence) frombase velocity data. Thresholding the base reflectivity field at several intensity levels (for ex-

492

ample. 15. 30. and 45 dBz) also identifies regions of precipitation. These feature-extractionalgorithm modules process the radar measurements made on each radar elevation scan.

The vertical-integration modules combinethe regions identified from scans aloft into three-


dimensional reflectivity and velocity structures.Velocity structures include convergence aloft,rotation aloft, divergence aloft (storm top divergence), and lower divergence (above the surfacebut below 1 km AGL). Reflectivity structuresinclude reflectivity cores, storm cells, and lowreflectivity cells.

The microburst-recognition modules consistof the surface-outflow, microburst-precursor,and surface-microburst algorithms. The surface-outflow algorithm attempts to identifymicroburst outflows by examining the temporaland spatial correlation of surface divergencefeatures. The microburst-precursor algorithmattempts to recognize precursor signatures thatindicate an imminent microburst, such as adescending reflectivity core coupled with a convergence aloft. The surface-microburst algorithm merges surface outflows with informationabout reflectivity structures, velocity structures, and precursor signatures to recognizemicrobursts.

The use ofmicroburst precursors to augmentsurface outflows allows the detection algorithmto detect many microbursts earlier in time thanwould be possible with the outflow informationalone. When Significant precursor structuresare identified, the algorithm declares a microburst alarm even when a very weak outflow isobserved. This conditioning ofthe alarm thresh01ds often detects microbursts in their earlieststages, minutes before the outflow itself is sufficiently strong to trigger an alarm.

Microburst-Outflow Signaturesand Their Detection

The divergence detection algorithm attemptsto locate two-dimensional regions of divergentshear, based on radial-velocity measurementsfrom the radar. The radial-velocity signatureobtained from a divergent flow is simply a rapidincrease with range of the radial velocity (wherepositive velocities indicate flow away from theradar). Figure 7 shows an example of such asignature. The color images in Fig. 7 illustratethe reflectivity field and the velocity field measured by the radar. Three microbursts are present in this data, outlined in red in the velocity



field. Figure 7 also shows the velocity profilealong a radial through the center ofthe strongestmicroburst. The divergence algorithm detectedthe section highlighted in red.

The algorithm scans the radial component ofeach velocity measurement to locate the shearsignature, forming shear segments whenever arun of generally increasing velocity values islocated. These shear segments are then associated across adjacent radials of the radar scan toform two-dimensional feature regions (Fig. 8).

The following description presents a briefoutline of the divergence detection process [15].The shear segments are identified by sliding awindow out in range, typically 0.5 km in extent,and starting a segment when the velocity valuesin the window are monotonically increasing.Once the start of a segment is found, the end isdetermined by moving the window out farther inrange as the values in the window generallycontinue to increase. This trend terminateswhen at least three-quarters of the samplepoints in the window have a velocity value lessthan the value at the first point in the window.

Each segment is then subjected to an iterative series of validation tests. On each iterationof the validation process, the segment endpoints are adjusted to insure that they arereasonable local extrema and have adequateslope or shear. Mter adjustment, the segment istested to insure that it has adequate length anda consistent positive trend. Finally, each endpoint is tested to insure that its velocity value isclose to the local median velocity value. If theshear segment passes all of these tests, it isaccepted; otherwise the test is repeated until thesegment is accepted or rejected.

The segments that survive these validationtests are then associated across radar azimuthsto form two-dimensional regions of shear. Anytwo segments that exhibit adequate overlap inrange (at least 0.5 km) and are within 2° inazimuth are joined together in the same region.The association process continues until all segments have been separated into regions. Theaggregates are now thresholded based on theirtotal area, number of segments, and maximumsegment strength. Regions with area less than 1square km, with fewer than 3 segments, or

493


Fig. 7-Radar signature of microburst outflow. (above) The radar reflectivity field for a strongmicroburst recorded on 21 June 1988 in Denver. (facing page) The radial-velocity field for the samemicroburst. The reflectivity image is in units of dBz, and velocity values are in m/s (negative valuesindicate velocities toward the radar). The region outlined in red is microburst outflow. Note the regionofaliased velocities in the velocity image at range 8 km andazimuth 310°. The Nyquist velocity for thisdata is ±18 m/s; the strong velocities in the microburst outflow have exceeded this limit. (below) Thevelocity profile through the center of the microburst shows a characteristic divergence signature ofoutflow along a radar radial at azimuth 313.5° for the case shown in the figure above. The portion ofthe velocity profile detected by the divergence algorithm is highlighted in red.

Divergence Shear Segment

with a maximum velocity differential (acrossthe strongest segment in the cluster) less than5 m/s are discarded. The clustering processresults in a set of significant regions of divergent shear, which are then passed to the threedimensional vertical-association and timecontinuity modules.

Microburst Features Aloftand Their Detection

Vi20

--.s 10~·u0 0a;>(ij -10i5CIla:

-207 8 9 10

Range (km)

11 12

In addition to surface outflow, microburstsare also associated with reflectivity and velocityfeatures aloft. These features aloft can aid the

494

recognition of surface outflows and can alsoserve as microburst precursors. A microburst



precursor is a set of features aloft that indicatesthe surface outflow will occur in 5 to 10 min.

Figure 9 shows a conceptual diagram for amicroburst typically found in the southeasternUnited States. In the first stage of development,a reflectivity core appears aloft (at about 5 kInAGL) in the parent storm of the microburst. Atthe same time, a convergence or inflow begins todevelop at or above the level of the reflectivitycore. In the second stage of development. thereflectivity core begins to sink to middle level(about 3 km). Convergence is evident now at thislevel and rotation often develops. Taken together. the descent of the reflectivity core andthe development of the middle-level velocitysignatures indicate that a strong downdraft hasdeveloped. The strong downdraft will producethe hazardous outflow at the surface. In thethird stage of development, the reflectivity corereaches the surface, and divergent flow at thesurface is now evident. Thus features aloft canallow the algorithm to declare the microburst

TIle Lincoln Laboratory Journal. Volume 2. Number 3 (1989)

outflowat an earlier stage than can be done withsurface data only. Additional algorithm development may use these features aloft to generateforecasts of surface outflows. prior to actualdivergence at the surface.

Figure 10 illustrates the detection procedurefor a microburst on 7 June 1986 at Huntsville,Ala. The bottom graph shows the time course ofthe surface outflow as determined by the algorithm and by an expert observer. The graphshows that the event is declared in a timelyfashion and is detected at all times when theoutflow velocity exceeds the microburst-alertthreshold of 15 m/s. The upper two graphs ofFig. 10 show the velocity and reflectivity featuresaloft detected by the algorithm. The upper andlower altitude limits for each structure areshown as a function of time. The initial microburst-precursor declaration is made at about1642 GMT. based on the detected rotation aloftand reflectivity core signatures. The reflectivitycore descends over the next 9 min, and reaches

495


the surface at 1651 GMfwhen the initial surfaceoutflow of 8 m/s is observed (convergence aloftis also detected at this time).

On the basis of the precursor signature detected from the reflectivity core and the rotationaloft, the algorithm declared the microburst at1651 GMf, rather than on the succeeding scan.For this case, the early declaration improved thetimeliness of the microburst alarm by over 1min, and the precursor signature was declared9 min in advance of the surface outflow.

Gust Front Detection andWind-Shift Estimation

Although gust fronts are often associatedwith thin lines of weak reflectivity and/or enhanced spectrum width, the most consistentand easily identifiable gust front signature isradial convergence in the Doppler velocity field.

Figure 11 provides an example of a thunderstorm outflow and gust front as seen by aDoppler radar. Figure 11 (left) shows the gustfront as a thin line of weak reflectivity in the reflectivityfield. Figure 11 (middle) shows the outflow as the region of positive velocit~es in theDoppler-velocity field. Figure 11 (right) showsthe linear signature of the gust front in thespectrum-width field. The gust front is the curvethat separates the outflow from the ambient air.

The red line in the figure represents the gustfront as detected by the algorithm.

The gust front algorithm, originally developedat National Severe Storms Laboratory (NSSL) in1984 [16] and continuously upgraded andimproved since that time, detects gust fronts byusing Doppler velocity alone [17]. The gust frontalgorithm consists of two separate algorithms:the gust front detection algorithm and the windshift estimation algorithm. For simplicity and

496

Fig. 8-Association ofdivergence shearsegments into two-dimensionaldivergence regions. Primitiveshear segments are shown in yellow, and resulting divergence regions in red.

TIle Lincoln Laboratory Journal. Volume 2, Number 3 (1989)


Storm Cell

-~ -

5 km -

10 km - -

Surface

Initial Stage

(a)

Descending Core

(b)

SurfaceMicroburst

(c)

Fig. 9-Stages in the evolution ofa microburst. (a) A reflectivity core initially forms aloft at 5 to 10 min prior to the onsetofsurface outflow. (b) As the downdraft develops, the core descends, andconvergence and rotation develop. (c) Finally,the core reaches the surface and the surface outflow begins.

convenience, these algorithms are togethercalled the gust front algorithm. This algorithmserves two functions: wind shear estimation foraircraft warning. and wind-shift estimation andforecasting for airport configuration planning.Wind shear warnings are issued when a gustfront impacts the runways or within 6 kIn of theends of the runways. The planning functionalerts anATC supervisor when a change in windspeed and/or direction caused by a gust frontwill occur at the airport within 20 min. Theoutputs ofthe algorithm are (1) a line identifyingthe location and shape of the gust front, (2) anestimate of the winds behind the gust front. (3)forecasts ofthe location of the gust front. and (4)an estimate of the wind shear an aircraft willexperience as it encounters the gust front.

The gust front algorithm processes two fullcircle scans (called tilts) during each volumeupdate of the antenna scan strategy. The elevation angles of the tilts are typically 0.5° and 1.0°.The gust front algorithm uses pattern-recogni-

tion techniques to identify gust fronts in theradial-velocity field. For each tilt. the algorithmsearches along radials for segments of decreasing radial velocity, which indicates radial convergence or shear (Fig. 12). If the value of theshear associated with each segment passesspecified thresholds, the segments are stored forfurther processing.

Shear segments are grouped on the basis ofproximity into features that are tested againstthresholds and either discarded, kept. or combined with other features. After separately processing both tilts. the algorithm tests for verticalcontinuity between the tilts. If the algorithmestablishes vertical continuity and the resultingfeature passes a length threshold, a gust frontdetection is declared. Gust front location isdetermined by fitting a line through the locations of the peak shears along all shear segments that constitute the gust front. Peakshear is the maximum decrease in radial velocity that occurs over a distance of 1 km along



Pre-Microburstcursor

Detected Alarm

6:::JCJ<l:: 4E~Q)"0 2.a+:::«

06

:::JCJ<l::

4E~Q)"0 2.a'';:::;

«0

~ 30en--.s.~ 20u0(j)> 10~ ..0 ..

:;:::(c)\5

0 01640 1650 1700 1710 1720

Time (GMT)

MicroburstWarning Alarm

Threshold Threshold

Fig. 10-Example of the use of microburst features aloft.The microburstalgorithm detected the surface outflowandfeatures aloft for an event observed by the Lincoln Laboratory testbed radar on 7 June 1986 at Huntsville, Ala. (a)Velocity features aloft detectedby the algorithm; (b) reflectivity features aloft detected by the algorithm; (c) surfaceoutflowas detected by the algorithm and determined by anexpertobserver. In this case, a microburst-precursordetection was made 9 min prior to the onset of surface outflow.The precursor is used to make an early declaration of thesurface outflow, which increases the timeliness of thealarm.

a given shear segment.The estimate of the wind shear that an air

craft might experience upon penetrating a gustfront is derived from the peak shear. The meanand standard deviation of the peak shears of all

498

shear segments in a gust front are computed.The wind shear estimate is simply the sum ofthese statistics.

To produce a forecast of the gust front location, the motion of the gust front with time mustbe established, which requires detections of thesame gust front in two consecutive volumescans (Fig. 13). The centroid ofeach detection oneach volume scan is computed. The algorithmcalculates the distance between the centroidson consecutive volume scans and, if this distance passes a threshold, the detections areidentified as the same gust front.

To generate a forecast, gust front propagation must be estimated. Gust fronts tend topropagate perpendicular to their orientations,so the component of the centroid-to-centroidvector perpendicular to the orientation is computed. The magnitude of the perpendicularcomponent is divided by the time differencebetween consecutive detections to derive propagation speed. Propagation direction is parallel to the perpendicular-component vector.The forecast is simply the current detectionmoved along the propagation direction by adistance defined by the propagation speedmultiplied by the desired forecast time (typically 10 and 20 min).

The gust front algorithm attempts to estimatethe wind speed and direction ahead of and behind the gust front. The algorithm uses datafrom the 0.50 tilt and assumes a uniform horizontal wind within specified spatial sectors.Estimates ofthe wind components are obtainedby regressing the smoothed Doppler velocitieswithin each sector onto sine and cosine functions, and minimizing the sum of the squarederrors between the measurements and the fittedvalues [18).

Alarm Generation

Mter the detection algorithms process thedata, detections are sent to the display systemfor alarm generation (NCAR developed the requirements for this portion ofthe system, as wellas the operational implementation used dUringthe TDWR OT&E). The display system generatestwo types of user displays: a geographic situ-



Fig. 11-An example of a thunderstorm gust front as seen by single Doppler radar. The gust front isidentifiable as (left) a thin line of weak reflectivity, (middle) a linear region of radial convergence in theDoppler wind field, and (right) a linear feature in the spectrum-width field.

ation display (GSD) and an alphanumeric display. Airport tower supervisors and TerminalRadar Approach Control (TRACON) controllersuse the GSD to understand the overall weatherconditions impacting airport terminal operations. Local controllers use the alphanumericdisplay for alarm messages to be read directly topilots. Figure 14 shows examples of both theGSD and alphanumeric displays [19).

To display microbursts, the display systemfirst computes a smooth shape to represent thealarm region. The shape is constrained toa rectangle with semicircular ends (like abandaid) that best fits the alarm region. Averaging is also performed with previous alarm regions to reduce the minute-by-minute jitter inthe alarm size and location. The bandaidshapes are used to generate the GSD display ofthe microburst alarms.


To display gust front detections, the displaysystem connects the series of points that describe the location ofthe gust front. The estimateof the wind speed behind the gust front isrounded to the nearest 5 knots and displayed asan alphanumeric character behind the front.Wind direction is shown by an arrow. The display software uses the most recent detectionand the gust front propagation speed to compute the forecasted locations. and then displaysthe forecasted locations as dashed lines. Thedetected and predicted gust front locations arepropagated each minute on the display. toupdate the gust front product along with themicroburst displays.

To provide the alphanumeric warning messages, the display system intersects the shapesand lines of a microburst or gust front with theairport complex. The airport complex consists of

499


(a)18

12v;-

6--g~ 0'00 -6Q5>

-12

-18

25.04 27.44 29.84 32.24 34.64

Range (km)

Shea' ~ Shea, ~segment_~ (b) segment_~

Feature Feature

;::~,-~ ;::~,-~Low Scan High Scan

Gust Front ....."..~4....)j"-

(c)

Fig. 12-Diagram illustrating gust front detection. (a) Radial-velocity data are searched for segments ofdecreasing Doppler velocity, which indicates radial convergence. The dots represent the locations of the peakshears along each segment. (b) Segments from the low (0.5) andhigh (1.0) tilts aregrouped respectively intofeatures, and (c) features are grouped on the basis of spatial proximity into gust fronts.

the area enclosed within a I-km buffer zone oneither side of the runway and 6 km past the endof the runway. If an intersection occurs, thesystem generates an alphanumeric alert for thatrunway, indicating the strength and location ofthe alann with respect to the runway. Since amicroburst is currently considered more hazardous to an aircraft than a gust front. a microburst alann has priority over a gust front alannon the same runway.

500

Evaluation of Wind Shear DetectionPerformance

Performance evaluation has been a majorcomponent ofthe wind shear algorithm development. Numerous performance evaluations wereconducted to establish the absolute level ofperformance of the detection algorithms and toidentifY relative benefits of alternate algorithmapproaches. A formal operational test and

The Lincoln Laboratory JournaL Volume 2. Number 3 (1989)


evaluation of the TDWR system was conductedat Stapleton Airport in Denver during the summer of 1988. The system output products werethen sent to on-duty air traffic controllers atStapleton for evaluation.

During the operational evaluation. thetestbed TDWR system was operated each day

Fig. 13-Diagram illustratinggust frontprediction. For eachdetection, the orientation andcentroid location ofthe detection are computed. The location of the previous (to) detection, whose centroid is CO' is shown by the dash-dot curve.The location of the current (tl) detection, whose centroid isCI' is shown by the solid curve. The orientation of thecurrent detection is given by the thin solid line. L is thecentroid-to-centroid distance and L' is the component ofLthat is perpendicular to the orientation. D is the distance thegust front is expected to travel in 10 min; the dashed curveindicates the expected location of the gust front in 10 min.

The technical performance evaluation compares the wind shear alarms generated by the

Fig. 14-Displays for TDWR. The color display on the leftis the geographical situation display (GSD) used by towerand TRACON supervisors. It depicts location ofwindshearand wind-shift events as well as storm regions. The alphanumeric display on the right is for local controllers. Itprovides runway-oriented text messages to be readdirectlyto pilots.

from noon to 7 PM. All base radar observationswere routinely recorded and all TDWR systemalarms and products were archived for subsequent analysis. Human observers monitored theweather situation and visually examined theradar measurements for wind shear events. Forpost-mission analysis, the observers kept detailed logs that noted the presence ofwind shearevents. Figures 15(a) and 15(b) provide a summary of the number of microbursts and gustfronts that were noted by these observers overthe course of the evaluation period.

Two methods determined the performancelevel of the wind shear algorithms dUring theevaluation. The first method, described as atechnical evaluation. compared the wind shearalarms to human-generated ground-truth datathat indicated the actual location of microburstor gust front events. This method ofcomparisonis the primary metric for the ongoing development and reporting of algorithm performance.The second method, described as an operationalevaluation, compared system alarms with reports from pilots who were operating at Stapleton dUring the evaluation period.

Technical Scoring Approach

••••••••D ••••...

••••••••••••••

,l

I

#

/.,---'"

- - - Previous gust front detection (at time to)---- Current gust front detection (at time t1)

Co Centroid of gust front at time toC1 Centroid of gust front at time t1L Centroid-to-centroid distance

---- Gust front orientation (at time t1)

L' Component of L perpendicular to gustfront orientation

D Distance gust front is forecast to movein 10 min

••••••• Predicted 1O-min gust front location

(for time t1 + 10 min)


Merritt et aI. - Wind Shear Detection with PencU-Beam Radars

18

16

14

12

Number of10

Gust Fronts8

6

4

2

0

r- Weak

Moderatef- - -

Strong

r- Severe

~-

-

• -

• e I • •I ~ 11 n01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 7 18 19 20 2 22 23 24 25 26 27 28 29 30 3

July. 1988

18

6

14

12

Number of10

Gust Fronts8

6

4

2

0

Weak

Moderate- -

Strong

Severe ---

n~ I. • ~ rn

01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

August. 1988

Fig. 15 (a)-A summary of gust front occurrences from the TOWR OT&E during July and August 1988.

microburst and gust front algorithms with reference (or ground truth) alarms generated bycareful human analysis of the radar measurements. The ground-truth data base was createdby using the base radar measurements from thetestbed radar (the data input to the detectionalgorithms) along with data from other sensorssuch as surface wind stations and a secondDoppler radar.

The algorithm performance is quantified bytwo basic statistics: the probability ojdetection(POD) and the probablity ojJalse alarm (PFAJ,which are defined as follows:

502

PFA = t_al_s_e_al_arm--::-_s _correct alarms + false alarms

detected eventsPOD=------

total events

The definitions relate performance to threefundamental concepts: an event. a detection.and a Jalse alarm. These basic terms can bedefined in several ways to quantify differentaspects of the performance of a system. Toprovide an operationally significant measure ofthe behavior ofthe wind shear detection systemas it would appear to an end user (for example.


Merritt et al. - Wind Shear Detection with Pencil-Beam Radars

30

25

20

Number of 15Microbursts

10

5

o

r-

- -

.1111... • • _I01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

July, 1988

30

251-------------------

201-----------------------1._

Number of 15 1--- .----Microbursts

10 I------- ~I__-_l__----- ----j.

5 I------- ~I___I__.J__-------_l__~.

01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

August, 1988

Fig. 15 (b)-A summary of microburst occurrences from the TOWR OT&E during July and August 1988.

an aircraft pilot), the following definitions havebeen used.

Event-A wind shear event is defined as asingle observation (by the ground-truth analyst)ofa microburst or gust front in the radar data ona low-elevation scan. Each microburst or gustfront is typically observed on several sequentialscans and hence represents several events. Onlythose microbursts that fall within a 10-kInradius of the airport are considered in the scoring. Microburst events that are partially insidethe 10-kIn limit are scored if 10% or more of thetotal event area. or at least 1 square kIn. whichever is smaller. falls within 10 kIn. All gust frontevents located within 60 kIn of the radar arescored.

Detection-A microburst or gust front event isconsidered a detection if its area is overlappedby a corresponding ground-truth event.

False Alarm--A gust front declaration is con-

The Lincoln Laboratory Journal. Volume 2, Number 3 (1989)

sidered a false alarm if no ground-truth eventsoverlap the area of the alarm. A microburstdeclaration is considered a false alarm if theground-truth events overlap less than 10% ofthe area of the alarm. Otherwise. an alarm isconsidered correct.

The POD and PFA statistics describe theperformance of target detection systems. Thespatially distributed nature of the wind shearhazards. however. makes these simple detectionstatistics inadequate to describe completely thebehavior of the detection algorithms. The existing scoring techniques do not explicitly quantifYhow timely the detections are (i.e.. how soon thealarms start relative to the onset of the hazardover the airport region), how completely thehazard region is detected. or whether the hazardregions are detected one-for-one by the algorithms (i.e.. splitting and merging of alarms).The assessment of algorithm performance in

503


these areas is difficult. and a matter of ongoingstudy.

The following sections describe in more detailthe generation ofground truth and the technicalscoring results for the microburst and gustfront algorithms.

Microburst AlgorithmPerformance Assessment

The performance of the microburst detectionalgorithm dUring the TDWR OT&E was evaluated by using ground-truth information fromsingle-Doppler radar analyses performed at LincoIn Laboratory. Table 1 lists the days for whichground truth was generated for the 1988 cases.

The single-Doppler ground truth was createdby manually observing the radar-measuredradial windfield and drawing polygonal outlinesaround the microburst-outflow regions in thedata. This manual analysis task required detailed examination of each radar surface scan(once per minute) for the duration of each caseday scored. A total of475 scans were analyzed inthis way. Figure 16 shows an example of asingle-Doppler ground-truth case. The figureillustrates the truth outline for a strong micro-

burst, along with the alarm generated by thedetection algorithm for the same event.

To compute the POD and PFA statistics. theground-truth microburst outlines are comparedto the alarms generated by the microburst algorithm. Table 1 shows these statistics. Thesingle-Doppler ground-truth cases achieved anoverall POD of 90% with a PFA of 2%. Table 2shows the POD as a function of microburststrength, and indicates that almost all of thealgorithm misses are for weak microburstevents (those with a wind-speed difference ofless than 15 m/s). Very few misses occur forevents with a wind-speed difference of 15 mls orgreater. Fortunately. the algorithm is more reliable for the stronger events that are more hazardous to aviation. The events missed by thealgorithm occur primarily at the start or end ofa microburst's lifecyde. when the microburstoutflow is weakest.

Gust Front Algorithm PerformanceAssessment

As stated previously. the operational demonstration was conducted 7 hours per day for 60days. With a volume update rate of 5 min, the

Table 1. Single-Doppler Ground TruthCases Analyzed for 1988 Data

Number of Number ofDate Microbursts Events" POD(%) PFA(%)

10 June 17 158 85 421 June 17 213 93 125 June 12 169 95 37 July 10 102 87 017 July 6 58 88 4

Total 62 700 90 2

"Note: The number of events indicates the number of scan-by-scan observations of themicrobursts by the radar. Each microburst is typically observed on about ten consecutiveone-minute scans. The POD and PFA statistics are based on the ability of the algorithm todetect each individual event.

504 The Lincoln Laboratory Journal. Volume 2. Number 3 (J 989)


Fig. 16-Example of microburst ground truth from single-Doppler data. The green outline denotes aground-truth region as determined by manual examination of the velocity field (shown in units ofm/s).The red outline indicates the alarm generated by the microburst detection algorithm.

•

gust front algorithm processed over 5.000 lowangle tilts. Each tilt potentially contained one ormore gust fronts. The goal of the gust frontalgorithm performance assessmentwas to truthand score. in near real time. every gust frontevent observed by the radar dUring the operational demonstration. To meet this ambitiousgoal in a timely manner. ground truth wasgenerated dUring real-time operations and thescoring procedures were automated. Thoughthe output ofthe algorithms was archived easilydUring operations. the real-time creation of aground-truth data base and the automation ofthe scoring process reqUired a significant software development effort.

Ground Truth

The gust front algorithm generated five products (detections. forecasts. wind-shift esti-


mates. and location and intensity ofwind shearalerts], each of which had to be scored. For thedetections. forecasts. and locations of the windshear alerts. the true locations of all gust frontswere needed. Ground truth for gust front location consisted of a series of points along theentire length of the gust front. The gust frontswere further categorized by strength. Thestrength (tlV) of the gust front was the averagepeak change in Doppler velocity perpendicular to and along the convergent portion of thegust front. Gust fronts were defined as weak(5 m/s::; tlV < 10 m/s], moderate (10 m/s::; tlV< 15 m/s], strong (15 mls ::; tlV < 25 m/s], andsevere (tlV ~ 25 m/s). Radar meteorologistsfrom NSSL and Lincoln Laboratory generatedreal-time ground truth. based on evidence ofgust front Signatures (radial convergence. azimuthal shear. and reflectivity thin line in singleDoppler radar data).

505

Merritt et al. - Wind Shear Detection with Pencil-Beam Radars

Table 2. Microburst Detection Performanceas a Function of Outflow Strength

Weak events POD(%) Strong events POD(%)Date (<15 m/s) (weak) (>15m/s) (strong)

10 June 78 71 80 9921 June 115 87 98 10025 June 42 81 127 997 July 51 77 51 9817 July 11 36 47 100

Ground truth for the wind-shift algorithmwas derived from surface wind measurementsmade by 38 portable weather stations locatedaround the airport [20]. This network ofweatherstations was called a mesonet. Winds ahead ofthe gust front were available to ATe via the LowLevel Wind Shear Alert System; thus only theestimate ofwind speed and direction behind thegust front was displayed and scored. Sincemesonet data constituted ground truth, onlythose wind shifts associated with gust frontsthat passed through the mesonet were scored.To determine which mesonet stations werebehind a gust front, the location of the gust frontfrom single-Doppler radar data was superimposed on the plotted mesonet data. Those stations which had experienced a change in windspeed and/or direction were used for analysis.The wind direction computed by the algorithmwas compared to the average wind diredionfrom the mesonet. The algorithm-computedwind speed was compared to the average of thepeak mesonet wind speeds that occurred dUring

a I-min interval. Pilot reports, recorded byobservers located in the tower, determinedground truth for the wind shear alert intensities.

Scoring Definitions, Rules, and Results

Representing the gust front ground-truthoutline by a straight line introduces errors intothe representation ofthe gust front location. Thegust front is a transition zone that is often notwell represented by a line. The difficulty inherent in identifying the precise location ofthe peakshear, and the use of straight line segmentsrather than curves to connect the points, alsointroduces errors. To compensate, truth is represented by a box 5 km wide and centered on thestraight line segments. Figure 17 illustratessuch a truth box.

The POD, defined as the number of truedetections divided by the number ofgust fronts,measures algorithm performance. A true detection is declared when any part of a gust front isdetected. Table 3 shows the POD as a function

Table 3. Probability of Detection of Gust Fronts

Moderate

73%

Strong

91%

Severe

100%

All

78%

506 The Lincoln Laboratory Joumal. Volume 2. Number 3 (1989)


60 km5 km

Fig. 17-Diagram illustrating truth, detections, and falsealarms. The 5-km-wide boxes represent ground truth, andthe solid lines represent detections. GF1 is an example ofa missed gust front (truth with no associated detection).EV21 is an example ofa false alarm (detection not associated with truth). The shaded area is the airport sector. Thevalues in square brackets are estimates ofwind speed anddirection behind the front.

plotted as a function of POL . . Table 4 gives themm

average POLas a function ofgust front strength.The location and intensity of the wind shear

constitute the wind shear warning. Location isscored by computing the number of wind shearalerts issued at the airport divided by thenumber of wind shear alerts that should havebeen issued. Table 5 shows the results of thisanalysis, which is termed the probability ojcorrectly locating gustJront wind shear.

The number offalse alarms issued divided bythe total number of alarms issued (3/206 or1.5%) defmes the probability ojJalse warning(PFW). In these three false alarms, gust frontswere in the airport vicinity, but the detectionsdid not agree well with the truth, and unnecessary warnings were generated. No detection inthe airport sector appeared where no gust frontexisted.

The measures of algorithm performanceshown in Tables 3 through 5 indicate that theability of the algorithm to detect gust frontsincreases with gust front strength. Thus, thealgorithm is better at detecting the more hazard0us gust fronts.

Wind shear intensity is scored by comparing

100

90 Severe -..

80

70

~60

~

0 5000... 40

30

20

10

00 15 30 45 60 75

POLmin (%)

Fig. 18-Probability of detection (POD) as a function ofpercent-of-Iength-detected threshold (PaLmi,,) for moderate, strong, severe, and all gust fronts.

of gust front strength.The PFA is defined as the number of false

detections divided by the total number of detections (true plus false). The PFA for the gust frontalgorithm was 27/1146 (2.4%). None of thefalse alarms occurred in the airport sector,defined as the area bounded by the azimuths2400 and 100 and the ranges 0 to 30 km from theTDWR testbed radar. Therefore the pilots received no false warnings.

The POD does not indicate how well thealgorithm detects a gust front. Figure 17 showsan example of a valid detection that may not beconsidered a good detection. A portion of thedetection labeled EV9 is located within the truthbox that represents a gust front identified asGF2. A valid detection is declared, but thealgorithm detection overlaps less than 10% ofthe total length of the gust front. If percent ojlength detected is signified by POL, a minimumPOL threshold (POL . ) can be applied so that

mmPOL must exceed the threshold before a validdetection is declared. Figure 18 shows POD

The Lincoln Laboratory JOtLmal. Volume 2. Number 3 (1989) 507


Table 4. Average Percent of Length Detected for Gust Fronts

Moderate

66%

Strong

68%

Severe

73%

All

67%

the intensity expressed in the wind shear alertmessage to pilot reports as logged by observersin the tower. The average absolute differencebetween pilot reports and alerts is about 10knots, with alerts overestimating wind shearrelative to pilot reports.

An aircraft that encounters a gust front isexpected to experience a gain in wind speed (Fig.3). In some instances, however, pilots report awind-speed loss. This fact results in a largewind-speed difference in the wind-speed erroranalysis. If the pilot reports of losses are removed from the analysis, the average absolutedifference between pilot reports and alerts isabout 5 knots. Cases in which pilots report awind-speed loss must be studied further todetermine why an inconsistency exists betweenthe reported and estimated shears.

Effective runway management is achieved byalerting an ATC supervisor, within 20 min ofarrival, when a wind shift is expected at theairport (forecasted location), and to what velocity the winds will change with its passage (windshift estimate). A valid forecast is declared if itfalls within the truth region for the time at whichthe forecast is valid. If the forecast falls outsidethe truth box, the forecast is considered a miss.A false forecast is declared if the gust front

dissipates before the validation time. The number of valid forecasts divided by the number ofevents for which forecasts were made gives theprobability ofcorrectforecast (PCF). The numberof false forecasts divided by the number ofevents for which forecasts were made plus falseforecasts gives the probability offalse forecast(PFF). The algorithm generates a forecast for 10and 20 min into the future. Table 6 shows thePCF as a function ofgust front strength. The PFFis 11% for the 10-minforecastsand 18% for the20-min forecasts. However, the algorithm wasable to produce forecasts for only 45% ofthe 270gust fronts and convergence boundaries thatoccurred dUring the 1988 test. When forecastsare generated, they are accurate; unfortunately,forecasts are not generated often.

The accuracy of the wind-shift estimate isdetermined by comparing the wind-shift estimate to the mesonet data. The average absolutedifference in wind speed and direction betweenthe wind-shift estimate and the mesonet data is3 mls and 30°, respectively. The wind-shiftspeed is 1.5 mls larger, on the average, thanthat determined from the mesonet data (whichmay be explained by the difference in height ofthe measurements [21]). The average wind-shiftdirection is about 5° counterclockwise of themesonet wind direction.

Table 5. Probability of Correctly LocatingGust Front Wind Shear

508

Moderate

64%

Strong

86%

All

70%

Summary

The Lincoln Laboratory program on windshear detection has resulted in the developmentof a significant capability for warning pilots ofimminent hazards along their approach anddeparture flight paths. This effort has coordinated research and development programs withother government-sponsored organizations

TIle Lincoln Laboratory Journal. Volume 2. Number 3 (1989)


Table 6. Probability of Correctly Forecasting Wind Shift

Moderate Strong Severe All

10

20

94% 96% 100% 95%

82% 84% 0% 83%

•

(most notably NSSL and NCAR), resulting in anintegrated system for the identification anddisplay ofhazardous wind shear and operationally significant wind shifts. Lincoln staffparticipated in the development of the TDWR systemspecification, and assisted the FAA in evaluatingcontract proposals for the system procurement.Several field measurement programs, particularly the TDWR OT&E at Denver in 1988, carefully assessed the performance of this system.

The results of the OT&E test and evaluationprogram indicate that the current wind shearand wind-shift algorithms perform well andprovide a significant benefit to the safety andefficiency ofterminal operations. Ongoing development and testing efforts aim to enhance theaspects of the system operation in whichalgorithm performance could be substantiallyimproved.

Acknowledgments

The work described in this article is the resultof several years ofdevelopment and field testingperformed by dozens ofindividuals from LincolnLaboratory, the National Severe Storms Laboratory, and the National Center for AtmosphericResearch. The authors would like to acknowledge the contributions of their Lincoln Laboratory colleagues Stephen Olson, Hayden Schultz,and Valerie Coel to the development of the realtime microburst algorithm, and Mark Isaminger, Darelyn Neilley, Richard DeLaura, RobertHallowell, and Joseph Cullen for their efforts onsingle-Doppler truthing and scoring. WilliamMahoney and F. Wesley Wilson (NCAR) providedconsiderable insight and assistance on algorithm development and evaluation. Michael


Eilts and Laurie Hermes (NSSL) were responsible for the development and evolution of thegust front and wind-shift algorithms, and muchof the ground-truth analyses for these algorithms. This work was sponsored by the FederalAviation Administration.

References

1. National Research Council, Low-Altitude Wind Shearand Its Hazard to Aviation (National Academy Press,Washington. DC, 1983).

2. M.M. Wolfson, "Characteristics of Microbursts in theContinental United States," LincolnLaboratoryJoumal1,49 (1988).

3. J.E. Evans, and D. Johnson, ''The FAA TransportableDoppler Weather Radar," Preprints. 22nd Con] onRadar Meteorology, Zurich, 10--13 Sept. 1984 (Amer.Meteor. Soc.), p. 246 (1984).

4. J. McCarthy, J.W. Wilson, and T.T. Fujita. ''The JointAirport Weather Studies Project," Bull. Amer. Meteor.Soc. 63, 15 (1982)

5. J. Dodge, J. Arnold, J. Wilson, J. Evans, and T.T. Fujita,''The Cooperative Huntsville Meteorological Experiment(COHMEX)," Bull. Amer. Meteor. Soc. 67, 417 (1986).

6. J. Wilson,J. Moore, G. Brant Foote, B. Martner,A. Rodi,T. Uttal, and J. Wilczak, "Convective Initiation andDownburst Experiment (CINDE)," Bull. Amer. Meteor.Soc. 69, 1328 (1988).

7. C. Biter, TDWR/LLWSAS User Working Group, Summary oj the 15-17 February 1989 Meeting, NationalCenter for Atmospheric Research, Boulder, CO (1989).

8. S.D. CampbellandM.W. Merritt, "TDWRScanStrategyRequirements," Project ReportATC-144, Lincoln Laboratory (1988).

9. R.D. Roberts and J. Wilson, "Precipitation and Kinematic Structure of Microburst Producing Storms,"Preprints. 22nd Con] on Radar Meteorology, Zurich,10--13 Sept. 1984 (Amer. Meteor. Soc.), p. 71.

10. S.D. Campbell, "Microburst Precursor RecognitionUsing an Expert System Approach," Proc. 4thlntl. Con]on Interactive InJolTTlation and Processing SystemsJorMeteorology, Oceanography and Hydrology, Anaheim,CA, 30 Jan.-3 Feb. 1989 (Amer. Meteor. Soc.), p. 300.

11. J.E. Evans, "Ground Clutter Cancellation for theNEXRAD System," Project Report ATC-122. LincolnLaboratory (1983).

12. D.R. Mann, ''TDWR Clutter Residue Map Generationand Usage," PrqjectReportATC-148, Lincoln Laboratory(1988).

509


13. S. Crocker. "TDWR PRF Selection Criteria," ProjectReportATC-14 7, Lincoln Laboratory (1988).

14. M.D. Eilts and S.D. Smith. "Efficient Dealiasing ofDoppler Velocities Using Local Environmental Constraints." Environmental ResearchLaboratories Report,National Severe Storms Laboratory (1989). DOT/ FAA/SA-89/1.

15. M.W. Merritt. "Automated Detection of MicroburstWind Shear for Terminal DopplerWeather Radar," Proc.Digital Image Processing and Visual CommunicationsTechnologies in Meteorology. SPIE 846, 61 (1987).

16. H. Uyeda and D.S. Zmic. "Automatic Detection ofGustFronts." J. Atmos. and Oceanic Tech. 3, 36 (1986).

17. S.D. Smith. A. Witt, M. Eilts, D. Klingle-Wilson. S.Olson. J. Sanford. "Gust Front Detection Algorithm forthe Terminal Doppler Weather Radar: Part I. CurrentStatus." 3rd Inti. Con] on Aviation Weather Systems,

MARKW. MERR11Tisastaffmember in the Air TrafficSurveillance group. His research interest is in the detection of aviation weather

hazards by Dopplerradar. He received an S.B.E.E. degree in1981 and anS.M.E.E. degree in 1983. both from MIT. Markhas been at Lincoln Laboratory since 1981.

STEVEN D. CAMPBELL hasworked for Lincoln Laboratory since 1978 in areas related to air traffic safety. Hedeveloped ground-support

hardware for the TCAS aircraft collision avoidance systemproject. and was project engineer for the Lincoln-developedsoftware and hardware for the GPS satellite navigationproject. Since 1984. he has been working on an artificial intelligence approach to microburst recognition for theTDWRprogram. His current focus of interest is in microburstprediction. He received his B.S.E.E.. M.S.E.E., and Ph.D.degrees in electrical engineering from the University ofWashington.

510

Anaheim, 30 Jan.-3 Feb. 1989 (Amer. Meteor. Soc.).p. 31 (1989). .

18. S.D. Smith. "Sectorized Uniform Wind Algorithms."NEXRADJointSystems Program Office Report. NationalSevere Storms Laboratory (1986).

19. D. Tumbull.J. McCarthy, J. Evans. and D. Zmic. "TheFAA Terminal Doppler Weather Radar (TDWR)Program." 3rd Inti. Conference on Aviation WeatherSystems. Anaheim. 30Jan.-3Feb. 1989(Amer. Meteor.Soc.). p. 414 (1989).

20. M.M. Wolfson. J.T. DiStefano. and B.E. Forman. "TheFLOWS Automatic Weather Station Network inOperation." ProjectReportATC-134. Lincoln Laboratory(1986).

21. M.D. Eilts. "Low Altitude Wind Shear Detection withDoppler Radar." J. Climate and Appl. Meteor. 26(1), 96(1987).

DlANAKLINGLE-W1LSON isa staff member in the AirTraffic Surveillance group.She received a B.A. degree ingeography from Indiana

University. an M.S. degree in atmospheric science from theUniversity of Michigan. and a Ph.D. degree in atmosphericscience from Purdue University. Diane also worked for theTravelers Insurance Corporation in Hartford. Conn.. in theNatural Hazards Risk Assessment Program. She joinedLincoln Laboratory in 1987.


,

Documents

Wind Shear Detection with Pencil-Beam Radars