doi: 10.1098/rsif.2010.0116, 30-43 first published online 2 June 20108 2011 J. R. Soc. Interface

Adriaan M. Dokter, Felix Liechti, Herbert Stark, Laurent Delobbe, Pierre Tabary and Iwan Holleman operational weather radarsBird migration flight altitudes studied by a network of

References http://rsif.royalsocietypublishing.org/content/8/54/30.full.html#ref-list-1 This article cites 45 articles, 4 of which can be accessed free

This article is free to access

Rapid response http://rsif.royalsocietypublishing.org/letters/submit/royinterface;8/54/30 Respond to this article

Subject collections

(498 articles)environmental science (12 articles)biometeorology

Articles on similar topics can be found in the following collections

Email alerting service hereright-hand corner of the article or click Receive free email alerts when new articles cite this article - sign up in the box at the top

http://rsif.royalsocietypublishing.org/subscriptions go to: J. R. Soc. InterfaceTo subscribe to

This journal is © 2011 The Royal Society

on February 16, 2011rsif.royalsocietypublishing.orgDownloaded from

Bird migration flight altitudes studiedby a network of operational weather

radarsAdriaan M. Dokter1,*, Felix Liechti2, Herbert Stark2,Laurent Delobbe3, Pierre Tabary4 and Iwan Holleman1

1Royal Netherlands Meteorological Institute, De Bilt, The Netherlands2Swiss Ornithological Institute, Sempach, Switzerland

3Royal Meteorological Institute of Belgium, Brussels, Belgium4Meteo France, Toulouse, France

A fully automated method for the detection and quantification of bird migration was devel-oped for operational C-band weather radar, measuring bird density, speed and direction as afunction of altitude. These weather radar bird observations have been validated with datafrom a high-accuracy dedicated bird radar, which was stationed in the measurementvolume of weather radar sites in The Netherlands, Belgium and France for a full migrationseason during autumn 2007 and spring 2008. We show that weather radar can extract nearreal-time bird density altitude profiles that closely correspond to the density profilesmeasured by dedicated bird radar. Doppler weather radar can thus be used as a reliablesensor for quantifying bird densities aloft in an operational setting, which—when extendedto multiple radars—enables the mapping and continuous monitoring of bird migration fly-ways. By applying the automated method to a network of weather radars, we observedhow mesoscale variability in weather conditions structured the timing and altitude profileof bird migration within single nights. Bird density altitude profiles were observed that con-sisted of multiple layers, which could be explained from the distinct wind conditions atdifferent take-off sites. Consistently lower bird densities are recorded in The Netherlandscompared with sites in France and eastern Belgium, which reveals some of the spatialextent of the dominant Scandinavian flyway over continental Europe.

Keywords: bird migration; radar ornithology; weather radar; altitude profile;verification

1. INTRODUCTION

Spatio-temporal information on bird migration is ofinvaluable use to scientists and society alike, but sofar no sensor networks have been established that pro-vide continuous and automated quantification of birdmovements over large areas. Comprehensive monitoringof bird migration at continental scales can provide fun-damental insight into migration patterns, the impact onmigratory flight of synoptic scale factors like weatherand orography and the selection of stop-over areas bymigratory birds [1–3]. Strong potential exists for theapplied use of continuous bird migration observations,for example in aviation for purposes of improvingflight safety. In particular, military low-level flyinghas a high risk of en route bird strikes and spatialbird migration information is essential for generatingreliable flight warnings to pilots. Other warning systemscould be envisioned to reduce the collision risk of birdswith wind farms, off-shore platforms and other

man-made structures, by allowing people to predictand adapt to specific mass migration events.

Radar has had an immense impact on ornithologyand the study of bird migration, because of its uniqueability to monitor bird movements up to high altitudesand distances during both night and day [1,4–7].Weather conditions [7,8], topographical features likecoastlines [9,10] and orography [11] all influencemigratory flight. The exact interplay of these synopticscale factors and the migrational movements of birdsis hard to study in general, mainly because of a lackof observational data.

Operational weather radar networks exist in, forexample, Europe and the United States for meteorologi-cal applications. These networks have a large arealcoverage as illustrated in figure 1, showing part of theEuropean network OPERA (Operational Programmefor the Exchange of weather RAdar information; [12]).Although designed for precipitation monitoring,weather radars can also observe biological scatterers.Boundary layer clear-air weather radar echoes arecaused predominantly by arthropods (mostly insects)*Author for correspondence (a.m.dokter@uva.nl).

J. R. Soc. Interface (2011) 8, 30–43doi:10.1098/rsif.2010.0116

Published online 2 June 2010

Received 1 March 2010Accepted 10 May 2010 30 This journal is q 2010 The Royal Society

on February 16, 2011rsif.royalsocietypublishing.orgDownloaded from

and flying birds [13–15]. If weather radar could detectand quantify aerial bird densities automatically, theexisting weather radar network infrastructure could beused as a bird migration sensor network with anunprecedented coverage.

Although weather radars have been used in ornitho-logical research for several decades [16], to ourknowledge only two cases of continental bird migrationwere studied by a weather radar network [17,18], mainlybecause of the labour-intensive analytical workinvolved. To enable comprehensive monitoring of birdmigration by weather radar networks, it is essential to(i) develop an automated and portable method forboth the detection and the quantification of aerialbird densities and (ii) validate the reliability of thismethod with independent reference data. In thisresearch article, we address both these requirementsand report on the development and validation of abird migration quantification algorithm, extracting alti-tude profiles of bird density, speed and direction. Byapplying the algorithm to a network of weatherradars, we observed the progression of migratory flightsalong the migratory direction over a mesoscale range ofseveral hundred kilometres.

2. BIRD RADAR FIELD CAMPAIGNS

To obtain reference data for validating weather radarbird observations, we organized extensive field cam-paigns with a high precision dedicated bird radar. Anex-military pencil-beam X-band (l ! 3 cm) radar of

the type ‘Superfledermaus’ [19] was stationed in themeasurement volume of C-band weather radars inThe Netherlands (19 August–16 September 2007), Bel-gium (18 September–22 October 2007) and France (10March–9 May 2008; figure 1). The campaigns produceda large reference dataset consisting of bird densities andflight directions at 200 m altitude intervals and 1 h timeintervals for a four-month period in total. The birdradar system has been well calibrated and makes useof state-of-the-art echo identification and quantificationprocedures [20,21]. For each detected echo wing beatpatterns are automatically analysed, by which non-bird echoes like insects can be properly excluded. Batscannot be automatically distinguished from birds, butare relatively less abundant in the study area [22],and, therefore, unlikely to have influenced measure-ments. The bird radar thus provides a high-qualityreference for validating weather radar bird observationsover a full migratory season. Bird radar measurementprocedures are detailed in appendix A.

3. BIRD DETECTION ANDQUANTIFICATION BY WEATHERRADAR

Bird density quantification by weather radar relies onaccurate methods of target identification. Bird-scattered signals need to be automatically distinguishedfrom all other types of echoes, which includeprecipitation, ground echoes related to anomalouspropagation, clear-air returns by insects and to some

355˚ 0˚ 5˚ 10˚

45˚

50˚

0 100 200

km

7

15

23

31

39

47

55(dBZ)

EUMETNET/OPERA radar network

Wideumont

Trappes

Den Helder

De Bilt

Figure 1. Map of operational weather radars for part of western Europe. Radar sites are indicated by bullets, the weather radarsused in this study are labelled and coloured red. The OPERA reflectivity composite is overlaid for 19 April 2008 19.30 UTC.

Bird migration flight altitudes A. M. Dokter et al. 31

J. R. Soc. Interface (2011)

on February 16, 2011rsif.royalsocietypublishing.orgDownloaded from

degree Bragg scattering from refractive index turbu-lence. Bragg scattering can be present especially atthe top of the afternoon convective boundary layer[13,15], but is highly wavelength-dependent, appearingweaker on C-band by over 10 dBZe when comparedwith S-band [13,14].

Individual bird echoes can in principle be resolved byhigh-resolution weather radar [23], but usually thespatial resolution of operational base data is toocoarse. Also, wing beat patterns cannot be recorded inoperational scanning modes. Therefore, alternativemethods of target identification need to be developed.At C-band bird-scattered signals are typically foundat low reflectivity factors of 210 to 10 dBZe [24], butinsects [13] and hydrometeors (i.e. water and ice par-ticles) [25] can give rise to signals of similar strength.Additional information is necessary to separate outbird-scattered signals.

Central to our method of identifying bird-scatteredechoes is an analysis of the radial velocity of the scat-terers, which is measured routinely by Dopplerweather radars. With the exception of strongly convec-tive systems, the wind field carrying hydrometeors isspatially smooth and slowly varying, resulting in alow spatial variability of the measured radial velocities[26]. This also holds for Bragg-scattered echoes andechoes caused by insects, whose active flight tends tobe slow and of which migration is predominantly wind-borne [27,28]. Bird migration gives rise to a muchhigher spatial variability in radial velocity [24,29–31],arising from individual variations in speed and directionof flying birds (see §4.1).

For most Doppler radars, an additional measure ofradial velocity variance is available, namely spectrumwidth, which derives from the echo pulse statisticsunderlying a single resolution volume (instead of resol-ution-volume to resolution-volume variances). It hasbeen suggested that spectrum width can be useful intarget identification of various biological scatterers[32], but in this study spectral width was highly variableover different cases and meteorological conditions tomake straightforward use of it in an automated algor-ithm. Nonetheless, there is potential for usingspectrum width in bird detection algorithms, especiallywhen using more sophisticated Doppler spectrumanalysis that extends beyond the assumption of a Gaus-sian Doppler spectrum by most signal processors, whichis often invalid for birds [24].

The developed bird detection algorithm produces avertical profile of bird density, speed and directionevery 5–15 min at 200 m altitude resolution, based ondata at the 5–25 km range. At these close distances,the radar beam is sufficiently narrow to probe distinctaltitudes, and range-dependent biases are largelyavoided, which otherwise need to be compensated for[33]. Each extracted altitude profile is based on a spatialaverage over the 5–25 km range measurement window.For broad-fronted movement like most passerinemigration [7], such an average will be representative,as migration will be spatially homogeneous. The algor-ithm is not designed to monitor very local migrationfeatures within the field of view of the radar. Detaileddefinitions are given in appendix A.

3.1. Bird detection

The algorithm is based on the existing wind-profilingalgorithms for Doppler weather radars, using thevolume velocity profiling (VVP) technique [26,34].This technique was successfully applied in the contextof bird migration profiling in a related study by vanGasteren et al. [30]. The VVP analysis delivers an alti-tude profile of the average speed and direction of thescatterers by fitting the data to a constant velocitymodel (see equation (A 1)). Additionally, at eachheight, a radial velocity standard deviation sr iscalculated (see equation (A 2)). Recent studies havesuggested that sr is a good discriminator betweenhigh-quality wind measurements and bird-contami-nated wind profiles [24,30,31]. van Gasteren [30]showed that air layers with a high radial velocity stan-dard deviation (sr . 2 m s21) occurred at altitudes,where simultaneously bird migration was detected byan independent bird radar, which demonstrated thatsr is also a good indicator for the presence of birds.Reliable bird density quantification could not bedemonstrated in the study of van Gasteren [30] owingto contamination from residual precipitation and alarge distance between the weather radar and birdradar sites (80 km). In §4.1, we validate sr as a discrimi-nator for bird presence, where we will also discusspossible events unrelated to bird migration that cancause high values of sr.

3.2. Removal of non-bird echoes

In addition to the sr criterion, we developed a targetidentification scheme to filter out non-bird echoes fromthe radar volume data. All reflectivity factors above20 dBZe are masked as non-bird echoes, since such highvalues occurred rarely during bird migration in ourstudy period and exceed reflectivity factors expectedfor broad-fronted passerine migration (20 dBZe corre-sponds to approx. 3500 passerine-sized birds per cubickilometre at C-band (see equation (3.1) and figure 5legend). For single-polarization Doppler weather radar,we developed a cell-finding algorithm that selectswithin each elevation scan contiguous areas above acertain reflectivity threshold (see appendix A).

When selecting contiguous reflectivity areas, spur-ious precipitation cells may be detected in areas ofintense bird migration as well. Examples of selectedcells are given in figure 2, showing plan position indi-cators for a case of bird migration (figure 2a (i)) andprecipitation (figure 2b (ii)). To identify selected cellsin bird migration areas and discard them from the pre-cipitation map, for each selected cell an average nearestneighbour variance scell is computed for the radial vel-ocities (see equation (A 3)). Analogous to sr discussedpreviously, the variance scell is lower for wind-bornescatterers (both hydrometeors and insects) than forbirds performing active flight. This difference is illus-trated in figure 2b, where scell is plotted as a functionof the average cell reflectivity factor for a number ofcells detected during intense bird migration events(green bullets) and cells detected during events withconvective showers (blue bullets). A combined criterion

32 Bird migration flight altitudes A. M. Dokter et al.

J. R. Soc. Interface (2011)

on February 16, 2011rsif.royalsocietypublishing.orgDownloaded from

using reflectivity and radial velocity information is usedto assign cells to the precipitation map: cells with eithera low variance or a high reflectivity factor (scell ,5 m s21 or Zcell . 15 dBZe) are removed from the dataas non-bird scattering. From the remaining areas, anaverage reflectivity is calculated for each 200 m altitudeinterval.

We find that bird migration gives rise to a relativelyhigh fraction of radial velocity dealiasing outliers (up to15%) compared with precipitation (1%) and daytimeclear air echoes [35] using the dual-PRF (pulse rep-etition frequency) technique. This is likely related tothe fact that bird scattering arises from a few large com-plex targets for which the backscatter phase varies intime with slight geometry changes of the bird body.In the VVP analysis, radial velocity outliers areremoved by an iterative fitting procedure. In the calcu-lation of scell, the identification of outliers is lessstraightforward and has not been implemented. As aresult, values of scell for bird migration are typicallyhigher than for sr, and therefore also the optimalthreshold in scell is higher (5 m s21) than for sr(2 m s21).

One polarimetric weather radar was available in ourstudy (i.e. in Trappes, France). By taking advantage ofthe additional information, especially the removal of

precipitation from the data is improved. We excludecontiguous areas with a high correlation coefficient,rHV . 0.9 (indicative of hydrometeors; [36]) or highdifferential reflectivity, ZDR . 3.0 dB (indicative ofinsect; [37–40]). For polarimetric radar, the VVPanalysis of radial velocity data remains necessary todetermine the presence or absence of birds, becausewe found that the ZDR criterion is often insufficient tofilter out insect echoes, especially during cases withstrong convective mixing.

3.3. Quantification of bird density

Bird density information can be obtained from reflectiv-ity measurements, analogous to quantitativeprecipitation estimation. For S-band weather radars,empirical relationships have been found between reflec-tivity and the volumetric density of migrating birds[17,41], but only for a few preselected cases of birdmigration in the absence of non-bird scatterers like pre-cipitation. At a radar wavelength l, the reflectivity h,equivalent reflectivity factor Ze and bird density rbirdare related according to [25,41,42]

h ! 103p5

l4jKmj2Ze ! rbirdsbird; "3:1#

reflectivity factor (dBZ) radial velocity (m s–1)

s cel

l (m

s–1 )

–20

(a) (b)

(i)

(ii)

–13 –7 0 7 13 20 –15 –10 –5 0 5 10 15

00

5

10

15

5 10 15 20 25

reflectivity factor, Zcell (dBZ)

Figure 2. (a) Plan position indicators (PPIs) for reflectivity factor and radial velocity for a bird migration event ((i) 5 October2007, 00.02 UTC) and an event with weak convective showers ((ii) 26 September 2007, 16.32 UTC). The PPIs show the 1.28elevation scan up to 25 km range. Borders of identified reflectivity cells are indicated in red. (b) Radial velocity standard devi-ation scell as a function of cell-averaged reflectivity factor Zcell for reflectivity cells detected during intense bird migration events(green bullets, scell ! 9+ 2 m s21) and during events with convective showers (blue bullets, scell ! 0.5+0.2 m s21) for the Wide-umont weather radar. Selected intense migration events were 4, 5, 6, 7 and 13 October 2007 from 17.00 to 09.00 UTC next day.Selected convective shower events were 24 September 2007, 9.00 to 18.00 UTC; 26 September 2007, 07.00 to 18.00 UTC; 17 Octo-ber 2007, 06.00 to 17.00 UTC; 18 October 2007, 09.00 to 17.00 UTC. Only cells consisting of over 800 resolution volumes areshown to limit the number of scatter points. For the largest reflectivity cell in each PPI, a connecting solid line is drawn toits corresponding scatter point. Cells inside the yellow shaded segments (scell , 5 m s21 or Zcell . 15 dBZe) are classified asnon-bird reflectivity cells and removed from the scan.

Bird migration flight altitudes A. M. Dokter et al. 33

J. R. Soc. Interface (2011)

on February 16, 2011rsif.royalsocietypublishing.orgDownloaded from

with h in cm2 km23, Ze in mm6 m23, Km ! (m2 2 1)/(m2 $ 2) with m the complex refractive index of thescatterers, l (in cm) the radar wavelength, rbird inbirds per km23 and sbird the bird radar cross sectionat C-band in square centimetres. The bar over the ‘rbird-sbird’ product denotes that this quantity is strictly anaverage over the different bird types or species i accord-ing to

Pi rbird,isbird,i and over the scanned viewing

angles. With jKmj2 ! 0.93 for water [25], we find usingequation (3.1) that h is proportional to Ze by a factorof 28.0 at S-band (l ! 10 cm) (see also [42]), 361 atC-band (l ! 5.3 cm) and 3.51 % 103 at X-band (l !3.0 cm). Bird radar cross sections and thereforereflectivities h are of the same order of magnitudeat these radar bands. However, because of the l24

proportionality in equation (3.1), birds causemuch higher reflectivity factors Ze in S-band comparedwith C-band (and much lower reflectivity factorsin X-band).

Weather radar reflectivity can be converted toapproximate bird densities by dividing h in equation(3.1) by an average bird radar cross section sbird,which will be determined by the validation using birdradar measurements.

4. RESULTS AND DISCUSSION

4.1. Validation weather radar

In figure 3, we compare the bird density height profiledetermined by bird radar and weather radar for theperiod of 2–7 October 2007. A considerable nightlyvariance in flight altitudes and densities is recordedby the bird radar (figure 3a) during this period.While on 4 October, migration does not exceed1.5 km, on 6 and 7 October, the bird radar recordsmigration extending above 3 km height. These altitudeprofiles are closely reproduced by the weather radaralgorithm (figure 3b). Also, a remarkable correspon-dence is observed between both sensors for therecorded absolute numbers of birds, as illustrated bythe closely overlaying height-integrated bird densities(figure 3c).

For the quantitative verification, measurements wereused from all three campaigns in De Bilt, Wideumontand Trappes, comprising 118 days of continuous data.We limit the quantitative verification of the weatherradar algorithm to night-time hours only. During day-time bird flocks are common [7], which show no clearwing beat pattern and may therefore not be recognizedby the automatic target identification of the bird radar.For this reason, the dawn ascent of daytime migratingbirds is observed much more clearly by the weatherradar than by the bird radar (e.g. on 5–7 October;figure 3). During night-time, the weather radarrecorded above noise reflectivity factors in 61 per centof the surveyed time–height layers (i.e. a specificheight layer measured at a specific time) for airspaceup to 4 km above ground level (AGL), while the birdradar recorded a non-zero bird density in only 38 percent of the surveyed time–height layers. The datasetthus contains representative cases both with andwithout migrating birds.

4.1.1. Bird detection. Figure 4 shows for a large numberof time–height layers the radial velocity standard devi-ation, sr (see equation (A 2)) as a function of rawreflectivity (figure 4a, non-bird echoes included) andas a function of bird reflectivity (figure 4b, non-birdechoes excluded). Bird densities as determined by thebird radar are indicated by colours. The large majorityof non-zero bird densities are observed for time–heightlayers with sr . 2 m s21, which confirms that a radialvelocity standard deviation sr ! 2 m s21 successfullydiscards non-bird echoes. In figure 4a, a large numberof points cluster around sr ! 1 m s21 at high raw reflec-tivity values, which can be attributed to precipitationevents. A considerable number of time–height layerswith sr . 2 m s21 have obtained a high raw reflectivityfrom precipitation contaminations. The effect of remov-ing these non-bird echoes on the reflectivity is shown infigure 4b. Precipitation-contaminated scatter pointsshift horizontally to lower reflectivity values, and highreflectivity gets always associated to high bird densities.Some residual precipitation contaminations remain inthe region sr , 2 m s21, while time–height layers forwhich all resolution volumes get assigned to theprecipitation map are fully removed.

We define true bird presence when the bird densityrbird determined by bird radar exceeds 1 bird perkm23 and define a criterion for bird presence in weatherradar data by sr . 2 m s21. By collocating the birdradar and weather radar measurements, we can thusgroup all time–height layers into categories for thenumber of correct detections H, missed detections M,false detections F and correct non-bird detections Z.For the combined set of campaigns, we find that theweather radar system has a high probability of detec-tion (POD) to be H/(H $ M) (POD ! 97%). Theweather radar algorithm detects birds even at verylow densities and bird migration events are unlikely tobe missed. The false alarm ratio (FAR), F/(H $ F), ishowever rather high (FAR ! 42%), but the majorityof these false alarms occur in a regime of very lowbird reflectivity. Precipitation contaminations withreflectivity factors of 230 to 210 dBZe are frequent,but, as will be discussed below, such reflectivity factorscorrespond to very low bird densities below 1 bird perkm23 only. The rate of false detection steeply decreaseswith increasing bird reflectivity. Calculating the PODand FAR statistics for the subset of time–heightlayers with a reflectivity factor greater than 25 dBZe,we find POD ! 100% and FAR ! 3% only. Thisregime of bird reflectivity factors is of most interest,as it corresponds to events with moderate to high birddensities greater than 10 birds per km23.

Large radial velocity standard deviations, sr, canalso occur when the actual velocity field is not uniform(e.g. during strong wind shear; [43,44]) or when theterminal fall velocity is not constant (e.g. in a mixtureof snow and rain; [45]). Such nonlinear wind fieldsmay produce a high radial velocity standard deviationin the VVP analysis, which would falsely indicate anevent of bird migration. Fortunately, nonlinear windfields are usually associated with strong convection orfrontal passage, which is often accompanied by precipi-tation. Precipitating areas will be removed by the

34 Bird migration flight altitudes A. M. Dokter et al.

J. R. Soc. Interface (2011)

on February 16, 2011rsif.royalsocietypublishing.orgDownloaded from

precipitation masking algorithm, based on a high reflec-tivity factor (greater than 15 dBZe) or a low value ofscell. Since scell is a locally calculated quantity overthe nearest neighbours of each resolution volume, itsvalue increases only owing to velocity variations overscales of up to a few kilometres and not by large-scalenon-uniformities. Our bird radar measurements verifythat most daytime weather radar echoes in clear airwere caused by other scatterers than birds, whichwere correctly removed by the weather radar algorithmbased on a low value of sr. This confirms the suggestionby Caya & Zawadzki [43] that in clear air strong depar-tures of the wind field from linearity are rare, at leastduring our verification period.

The high spatial variability in radial velocity of birdscompared with wind-borne scatterers like insects andhydrometeors can be explained from the relativelysparse distribution of birds in airspace. When usingclose range data only (less than 25 km) up to bird den-sities of a few hundred birds per km23, a radarresolution volume contains only one up to a few individ-ual birds (assuming an even spatial distribution). Theradar resolution is therefore sufficient to resolve partof the speed and directional variation between

individual birds. In contrast, insect aerial densitiestend to be much higher than bird aerial densities [28].Each radar sample volume is thus more homogeneouslyfilled and individual speed and directional variations ofinsects are averaged out. To some degree this smoothingeffect is also observed for birds. As seen in figure 4, theradial velocity standard deviation sr slightly decreaseswhen the bird density increases. A second cause of thehigh radial velocity variance observed for birds is theirrelatively fast active flight (10–25 m s21; [46]), whichallows for more variation in speed and directionwithin individuals than for insects, of which the air-speed is low [47].

4.1.2. Quantification of bird density. In agreement withequation (3.1), we find a strong linear correlationbetween reflectivity h (measured by weather radar)and bird density rbird (measured by bird radar), as illus-trated in figure 5. An effective cross section for weatherradar at C-band can be calculated as sbird ! h/rbird(see equation (3.1)). We find a median sbird of 11+ 6cm2, with seasonal trends shown in figure 6. Eachdata-point refers to a nightly average of h/rbird,

02 Oct 2007 03 Oct 2007 04 Oct 2007 05 Oct 2007 06 Oct 2007time

0

50

100

150

bird

den

sity

(bir

ds k

m–2

)

1

2

3

4

heig

ht A

GL

(km

)

0

1

2

3

4(a)

(b)

(c)

0

1

2

3

4

1

2

3

4

1

2

3

4

heig

ht A

GL

(km

)he

ight

AG

L (k

m)

1

2

3

5

10

15

25

50

100

200

400

dens

ity (b

irds

km

–3)

–16

–12

–8

–4

0

4

8

12

refl

ectiv

ity fa

ctor

(dB

Z)

Figure 3. Comparison of the bird density altitude profiles determined by (a) bird radar and (b) weather radar. (c) Height-inte-grated bird densities are displayed for both weather radar (red) and bird radar (blue). Weather radar reflectivities were convertedto bird density by assuming a constant weather radar cross section at C-band of sbird ! 11 cm2 (see legends for figures 5 and 6).The period between sunset and sunrise is shaded in grey. Wind barbs in (c) show the wind profile from the HIRLAM numericalweather prediction model. Each half barb represents 10 km h21 and each full barb 20 km h21.

Bird migration flight altitudes A. M. Dokter et al. 35

J. R. Soc. Interface (2011)

on February 16, 2011rsif.royalsocietypublishing.orgDownloaded from

omitting time–height layers with a low bird reflectivity(less than 25 dBZe). From mid-September to mid-October (figure 6a), the average daily cross section issignificantly increasing (F22 ! 13.7, p , 0.002) from 7to 15 cm2 by 0.3+ 0.1 cm2 d21. A reverse temporaltrend is observed in spring when the cross section sig-nificantly decreases (F41 ! 4.3, p , 0.05) from 16 to10 cm2 by 0.1+ 0.04 cm2 d21 in the period mid-March to early May (figure 6b).

Wing beat data from the bird radar indicate that noc-turnal bird migration during all field campaigns wasstrongly dominated by passerine migration. Seasonalchanges in sbird must thus be explained from changesin the average body size of the migrating passerines.We can exclude insect scattering as a contribution tothe observed cross-section changes, as the number ofidentified insect echoes by the bird radar decreased inautumn and increased over the course of spring, whichare in fact the reverse trends that would explain theobserved cross-section variations (background scatteringby insects contributes to the measured reflectivity h byweather radar, but not to rbird determined by birdradar, leading to an effectively larger sbird). All largepasserines (greater than 50 g; [48]) in western Europeare Turdus thrushes, with the exception of some speciesthat are mostly daytime migrants like the European star-ling Sturnus vulgaris [5]. Both from visual observations,ringing data and radar observations [5,49,50], it is wellestablished that the migration of Turdus species in wes-tern Europe increases steeply during October, while theirreturn migration peaks relatively early in March. Theincreasing/decreasing abundance of large passerines inautumn/spring thus explains the seasonal changes inweather radar cross section.

For reasons of simplicity, we chose to use a constantbird radar cross section-sbird of 11 cm2 when converting

weather radar reflectivity into bird density byequation (3.1).

During some nights in early September, larger effec-tive sbird was observed up to 30 cm2 (data not shown).These cross sections are too large for the predominantlysmall passerines that migrate during this time of theseason [49]. We may speculate that these events arerelated to migratory noctuid moths engaging in

0 0.2 0.4 0.6 0.8 1.0 1.2 1.4% time–height layers

10

100

1000

bird

refl

ectiv

ity, !

(cm

2 km

–3)

–20

–15

–10

–5

0

5

refl

ectiv

ity fa

ctor

(dB

Z)

1 10 100bird radar, "bird (km–3)

# =17

cm2

# = 5 cm

2

Figure 5. Correlation weather radar reflectivity and bird radardensity in Wideumont, Belgium, corresponding to nightly(18.00–06.00 UTC) bird density estimates for 14 742 time–height layers recorded at 15 min time interval and 0.2 kmheight interval for the continuous period of 22 September–21 October 2007. We find sbird ! 11+ 6 cm2 and a correlationcoefficient R2 ! 0.73 (as calculated from a set of statisticallyindependent bootstrap samples). Solid black line; best fit:h ! 11rbird

0

1

2

3

4

5

6

# r (m

s–1

)

1 10 102 103 104

raw reflectivity (cm2 km–3)

–30

(a) (b)

–20 –10 0 10 20 30

raw reflectivity factor (dBZe)

#r = 2 m s–1

1 10 102 103 104

–20 –10 0 10 20 30

1

235

101525

50

100

200

400

bird

den

sity

, "bi

rd (k

m–3

)

bird reflectivity, ! (cm2 km–3)

105

bird reflectivity factor (dBZe)

1050

Figure 4. Radial velocity standard deviation sr as a function of raw reflectivity (a, reflectivity including non-bird echoes) and as afunction of bird reflectivity (b, reflectivity cleaned from non-bird echoes). Each scatter point refers to a time–height layer, i.e. aspecific height layer measured at a specific time. To limit the number of scatter points, only data for the Wideumont campaignare shown (22 September–21 October 2007, 14 742 detected non-empty time–height layers). In colour, the bird density is indi-cated as measured simultaneously by the bird radar. The large majority of non-zero bird densities are observed for time–heightlayers with sr . 2 m s21.

36 Bird migration flight altitudes A. M. Dokter et al.

J. R. Soc. Interface (2011)

on February 16, 2011rsif.royalsocietypublishing.orgDownloaded from

southward ‘return’ migrations [28,51]; however, dedi-cated measurements on the insect composition in theatmosphere would be necessary to confirm this. Becauseof the low migration intensities in early autumn, even asmall insect background can have a relatively strongimpact on the effective weather radar cross section.By mid-September, insect contamination no longerhad clear impact on sbird. During nocturnal springmigration, insect contamination is minor up to earlyJune. Weather radar-extracted nocturnal bird densitiesdrop below 5 birds per km23 by mid-May at the end ofthe migration season. In spring, insect contaminationsthus result in a bias of at most 5 birds per km23; how-ever, the contamination is likely less since by mid-Maybirds (e.g. Swifts) are still aloft and insect densities aremuch lower in March and April.

Although the aerial mass density of insects can belarger than for birds [28], scattering by birds is oftenenhanced relatively because of a larger radar cross sec-tion. For insects less than 1 cm, scattering enters theRayleigh regime, where the radar cross section steeplydecreases with the size of the object, while birds invari-ably give rise to a strong resonant (Mie) scattering. Itwill strongly depend on the typical insect size and den-sities whether insect contamination contributessignificantly to the total measured reflectivity [52].

Differences in the representativity of the bird radarreference data and the weather radar data will contrib-ute to the spread in the correlation between bird radarand weather radar in figure 5. An important differencebetween the two datasets is the size of the surveyedvolume. For each time–height layer, the weatherradar scans a 4 % 102 km3 volume, while for the sametime–height layer the fixed pencil beam of the birdradar surveys only a 0.01–0.1 km3 volume, dependingon the altitude (see appendix A for the exact scanningstrategies). Only in the case of ideal intense broadfront migration [7] will the correlation between thetwo radars be unaffected by the discrepancy in surveyedvolume.

4.2. Migration timing and altitude use

The effect of the environmental wind on flight altitudesbecomes evident by comparing the bird density profilewith the wind profiles calculated by the HIRLAM

numerical weather prediction model [53] (figure 3). Inline with previous studies [54,55], we observe that birdsadjust their flight altitude to make optimal use of tailwinds along the predominant migratory direction (inthis case towards southwest). For example, the HIRLAM

wind profile shows that on the night of 3–4 Octoberwind conditions at low altitude were more favourablethan at high altitude, where migrants would haveencountered strong south-westerly head and side winds.On this occasion, migration did not extend above 1 km.On the other hand, on the night of 5–6 October, tailwinds were present above 1 km and a large fraction ofmigration took place at high altitude.

Although decisions of birds to start migrating or notare primarily based on local weather conditions [8], wefind that migration patterns observed at single sitescan strongly depend on weather conditions elsewhere.This applies particularly to northern temperate climate,where frequent passage of high and low pressure systemscauses a large spatial variability in weather. At individ-ual sites, timing and altitude profile of bird migrationwas observed to be influenced by weather conditionsat locations further up the migratory flyway. We willrestrict our analysis to a spring migration event(19–20 April 2008), which illustrates the non-localcharacter of bird migration. The general mechanism offlight altitude selection and the adaptive response offlying birds to changing meteorological conditions arebeyond the scope of this manuscript.

We used four weather radars marked by red bulletson the map of figure 1. The OPERA reflectivity compo-site is overlaid to give an impression of the synopticweather situation. Active fronts related to a depressionabove the Bay of Biscay caused precipitation and east-erly winds above southern France. A weak occlusionfront at the Dutch–Belgian border slowly moved north-ward and diminished in activity in the course of thenight.

Figure 7a shows the retrieved bird density profiles forthe four sites ordered from north (top) to south(bottom). Arrows indicating flight directions are over-laid with the bird densities, while barbs in the pinkboxed insets show the HIRLAM wind profile at 00.00UTC. Large differences are visible between the sites,both in total density, timing and altitudinal profile ofthe migration.

01 Oct 15 Oct0

10

20

30(a) (b)

15 Mar 01 Apr 15 Apr 01 May

# bird

(cm

2 km

–3)

Figure 6. Seasonal trend in bird radar cross-section sbird at C-band. During autumn, the cross section increases in time (a, Wide-umont campaign), while in spring, the cross section decreases (b, Trappes campaign). These trends reflect the increasing/decreasing proportion of larger bird species (mainly Turdus thrushes) during the migratory season.

Bird migration flight altitudes A. M. Dokter et al. 37

J. R. Soc. Interface (2011)

on February 16, 2011rsif.royalsocietypublishing.orgDownloaded from

Migration in Trappes is characterized by strongdeparture and ascent to high altitudes above 2 km,where birds experience favourable tail winds alongthe predominant migratory direction (according toour bird radar-tracking measurements during noctur-nal migration towards 418 clockwise from north) asrevealed by the wind profile. Low-altitude migrationis avoided because of unfavourable easterly low-levelwinds. The migration intensity quickly drops after00.00 UTC and the arrival of birds from southern lati-tudes is limited. The drop in bird density is related tothe weather conditions south of Trappes, which, dueto rain and easterly winds, were unsuitable formigration.

Migration in Wideumont (280 km north east ofTrappes) starts out with strong departure at lower alti-tudes. Below 2 km, altitude tail winds are morefavourable than above 2 km, where initially strong wes-terly winds prevail (wind data not shown). After 22.00UTC, a remarkable double-layered bird density heightprofile is observed, with one band centred around800 m AGL and a second band centred around1900 m AGL. Since we started observing bird migrationby weather radar in autumn 2007, such double-layeredaltitude profiles have been observed regularly duringspring migration. The top altitude flight band appearsafter 4 h of flight time, which corresponds with ameasured bird ground speed of 70 km h21 to birdshaving travelled over a distance of about 280 km.These birds must therefore have departed in the vicinityof Trappes in northern France, where birds chose aflight altitude of 2500 m. Note that Wideumont islocated 585 m mean sea level (MSL), therefore the1900 m altitude band closely matches the cruising alti-tude after departure near Trappes. The double-layereddensity profile is not explained by the local wind profilein Wideumont. Rather, it is consistent with themigrants maintaining a constant flight altitude (abovesea level) once they had completed the ascent phaseof their migratory flight.

In De Bilt north of Wideumont, no birds depart inthe early night. This is explained by the weak occlusionfront slightly south of de Bilt, which blocks mostmigration. With this front weakening in activity,migration conditions become more favourable over thecourse of the night, and after 00.00 UTC, we do observethe passage and arrival of migrating birds. Nomigration is observed on the most northerly locatedweather radar in Den Helder.

Consistently lower bird densities are recorded inDe Bilt and even more so in Den Helder comparedwith the sites in Trappes and Wideumont. We calcu-lated migration traffic rates [20] defined as thenumber of migrating birds per hour crossing a 1 kmbroad vertical plane perpendicular to the migrationdirection at flight altitudes between 0.2 and 6 km.For nocturnal migration (between 18.00 and 05.00UTC), seasonally averaged traffic rates are listed intable 1 during an autumn period (22 September–21October 2007) and a spring period (15 March–15May 2008). Ordering the radars according to theirorthogonally projected distance from a line alongthe migration direction (towards 2218 in autumn

and 418 in spring as determined by the bird radar),we observe a northwest to southeast increase in thenumber of migrating birds. This increase revealssome of the spatial extent of the dominant Scandina-vian migration flyway. We may expect manyScandinavian breeding birds to fly across Denmarkand the southern tip of Sweden. Following theirendogenous migration direction, birds will concentrateeast of The Netherlands towards the centre of theflyway in Germany, eastern Belgium and France.Once data from the full OPERA weather radar networkbecome available, it will be possible to map for thefirst time the extent and dimensions of entiremigratory flyways over Europe.

5. CONCLUSIONS AND OUTLOOK

We have developed an automated method for birddetection and bird density quantification by weatherradar, extracting bird density, speed and direction asa function of altitude. The algorithm can run in nearreal-time with a time and altitude resolution of 5 minand 200 m, respectively, and was applied to four differ-ent Doppler weather radars in The Netherlands,Belgium and France. Generated altitude profiles arebased on data up to 25 km range and thus provide alocal characterization of the migration profile. Theextracted weather radar bird densities have been vali-dated with data from a high-accuracy dedicated birdradar, which was stationed in the measurementvolume of different weather radars during the peakmigration season of autumn 2007 and spring 2008. Wefind that Doppler weather radar is highly successful indetecting migrating birds and quantifying bird densitiesas a function of altitude. The probability of detection is(POD) is very high (99%), the false alarm ratio (FAR)is low (3%, provided that bird densities less than 10birds km23 are discarded) and weather radar reflectiv-ity can be quantitatively correlated to the birddensities determined by the bird radar.

By applying our bird quantification method to a net-work of weather radars, we observed how mesoscaleweather structured the intensity and timing of birdmigration. In particular, the altitude profiles at individ-ual sites were shown to be affected by the weatherconditions at other locations. While the local windfield is an important variable for understanding localbird densities aloft, weather conditions in the vicinityof take-off sites can also strongly determine theobserved flight altitude profiles.

The current study shows that automated birdmigration monitoring by operational weather radar net-works is feasible. The developed methods for birddetection and quantification can be easily extended tofull operational weather radar networks. With thedevelopment of the OPERA data centre for radar datawithin the coming 2 years [12], the establishment of acontinent-wide bird migration sensor network inEurope is within reach. Besides better scientific under-standing of continental bird migration, such anetwork can enable important applications both inaviation flight safety and environmental impact

38 Bird migration flight altitudes A. M. Dokter et al.

J. R. Soc. Interface (2011)

on February 16, 2011rsif.royalsocietypublishing.orgDownloaded from

assessments for land use planning and development.Large-scale continuous monitoring by radar networksmay improve our understanding of avian-borne diseasespread, by revealing in more detail the temporal andspatial dynamics of migratory flyways, though weknow of no examples of the use of radar in avianepidemiology so far [56].

Currently, polarimetric weather radar is rapidlybecoming the new operational standard. Dual-polariz-ation techniques will likely contribute to futureimprovements of the current bird migration quantifi-cation algorithm. We expect that automated birdmigration quantification can also be implemented forweather radars operating at different radar wavelengths

19-Apr-08 18:00 20-Apr-08 00:00 20-Apr-08 06:000

50

100

150

19-Apr-08 18:00 20-Apr-08 00:00 20-Apr-08 06:000

50

100

150

0

1

2

3

4

0

1

2

3

4Den Helder

19 Apr 2008 18.00 20 Apr 2008 00.00 20 Apr 2008 06.000

50

100

150

0

1

2

3

4

0

1

2

3

4De Bilt

0

1

2

3

4

0

1

2

3

4Wideumont

18.00 20.00 22.00 00.00 02.00 04.00 06.00 08.00 10.000

1

2

3

4

0

1

2

3

4Trappes

1 2 3 5 10 15 25 50 100 200 400density (birds km–3)

–16

(a)

(b)

–12 –8 –4 0 4 8

reflectivity factor (dBZ)

Den Helder

De Bilt

Wideumont

Trappes

heig

ht A

GL

(km

)

bird

den

sity

(b

irds

km

–2)

time (UTC)

Figure 7. Bird densities as a function of time and altitude at different weather radar sites (figure 1) for the night of 19 April 2008; (b)shows the height-integrated bird densities over 0.2–4 km AGL. The period between sunset and sunrise is shaded in grey and civiltwilight is shaded in light grey. The pink boxed inset on the right-hand side of each panel shows the HIRLAM wind profile at 00.00UTC. Bird flight speed and directions are indicated as barbs overlaid on the altitude profile. Each half barb represents 10 km h21

and each full barb 20 km h21. Around 00.00 UTC, a double-layered bird density profile is observed in Wideumont. We attribute thetop band to birds that departed in the vicinity of Trappes, while the lower band results from birds departing more locally.

Bird migration flight altitudes A. M. Dokter et al. 39

J. R. Soc. Interface (2011)

on February 16, 2011rsif.royalsocietypublishing.orgDownloaded from

like S-band and X-band; however additional researchand validation is required, because of the very differentrelative cross section of hydrometeors, insects and birdsat these wavelengths. Weather radars are capable of pro-viding spatial information on bird migration by usingdata from their full surveillance area [1,3,33]. A challen-ging task will be the development of automated methodsto identify and quantify these spatial migration patternsat regional scales. Dedicated methods to identify (flocksof) large birds will be of interest to aviation, as thesebirds pose the highest collision risk.

For early-warning systems of bird migration, oper-ational bird density forecasts are essential, which couldbe made in combination with spatially explicit birdmigration models [57,58]. Such models may account forthe inherently non-local character of bird migrationand deal with the sparseness of radar observations, butwill depend on data assimilation of areal bird densityinformation. In meteorology, the integration of modelsand observations has become common practice and hasgreatly improved weather prediction. We expect that asimilar synergy between bird migration observations byweather radar and migration models will greatly improvethe description and predictability of bird movement.

We would like to thank Robert Diehl and four anonymousreviewers for their constructive comments on the manuscript.We would like to thank our field campaign collaborators:Claudio Koller, Thomas Luiten, Valere Martin, PirminNietlisbach, Marco Thoma and Andreas Wagner. TonVermaas of the Royal Dutch Airforce at Soesterberg Airfield,Serge Sorbi of the Belgian Royal Airforce at Saint-Hubert AirBase and Mr Bizeul of Lyonnaise des Eaux providedextensive support stationing the bird radar. We thank HansBeekhuis, Geert De Sadelaer, Christophe Ferauge and Jean-Claude Heinrich for technical assistance. We also thank Hansvan Gasteren and Judy Shamoun-Baranes for usefuldiscussions. This work was conducted within the frameworkof the European Space Agency’s (ESA) Flysafe project.

APPENDIX A. MATERIALS ANDMETHODS

A.1. Weather radar methods

The weather radars used in this study were located inDen Helder, The Netherlands (52.954 N/4.791 E,

51 m MSL), De Bilt, The Netherlands (52.103 N/5.179 E, 44 m MSL), in Wideumont, Belgium(49.915 N/5.505 E, 592 m MSL), and in Trappes,France (48.775 N/2.009 E, 191 m MSL). Typical C-band weather radar technical parameters are listedin [31]. Range/azimuth resolution equalled 1 km/18,1 km/18, 250 m/18 and 240 m/0.58 for the fourradars, respectively. All available Doppler scans wereused in the analysis, i.e. depending on the radarfrom 8 to 13 scans distributed between 0.4 and 258elevation.

For 200 m height intervals up to 6 km AGL, weapproximate the local velocity field by a spatially con-stant model. The components of the local velocityfield in the x-, y- and z- directions are approximatedby constants u0, v0 and w0. The radial velocity Vr isthen given by

Vr"u;f# ! sinf cos u u0 $ cosf cos u v0

$ sin uw0; "A1#

with u, f the elevation and azimuthal angle, u0 and v0the Cartesian ground speed components and w0 the ver-tical speed [34,45]. After a first least-squares model fit toequation (A 1), points with a deviation from the modelof over 10 m s21 are removed as outliers resulting fromdealiasing errors [26]. Final velocity parameters aredetermined by a second model fit. From the fit residualsfor each resolution volume i of the radial velocity datato equation (A 1), we determine the VVP retrievedradial velocity standard deviation sr in a least-squaressense:

s2r !

1N &M

XN

i

'Vr;i & Vr"fi; ui#(2; "A2#

where Vr,i are the observed radial velocities, N is thenumber of data points and M is the number of esti-mated parameters in the radial velocity model (M ! 3in equation (A 1)). Height-intervals with data gapsalong the azimuth larger than 458 are rejected [26],which requires migration within the measurementwindow to be broad-fronted up to a certain degree.Data points N are all resolution volumes with anabove-noise reflectivity (including those assigned tothe precipitation map), excluding gates identified asclutter (see below).

For single-polarization Doppler radar data, contigu-ous cells of reflectivity are detected by a cell-searchingalgorithm [59], which defines cells as groupings of resol-ution volumes within an elevation scan for which eachresolution volume has a reflectivity factor greater than0 dBZe and at least five directly neighbouring resolutionvolumes that also meet this requirement. We find that aminimum threshold of 0 dBZe cuts out nearly all pre-cipitation if an additional fringe of 3 km width isadded around the selected cells to remove the low-reflec-tivity borders of precipitation areas. Cells withinelevation scans are analysed by a cell-averaged nearestneighbour radial velocity standard deviation scell,according to

s2cell ! kkV 2

r lneigh & kVr l2neighlcell; "A3#

Table 1. Average nocturnal migration traffic rates (in birdskm21 h21) during an autumn period (22 September–21October 2007) and a spring period (15 March–15 May 2008)for the four radar sites shown in figure 1. Autumn data forTrappes were unavailable. Average bird ground speeds usedin the calculation were obtained from the VVP radialvelocity analysis. Both in autumn and spring, a northwest tosoutheast increase in the number of migrating birds isobserved.

migration traffic rate (birds km21 h21)

season Den Helder De Bilt Trappes Wideumont

autumn 2007 243 504 — 857spring 2008 85 170 515 562

40 Bird migration flight altitudes A. M. Dokter et al.

J. R. Soc. Interface (2011)

on February 16, 2011rsif.royalsocietypublishing.orgDownloaded from

where the brackets k. . .lneigh denote an average over aresolution volume and its eight direct neighbours andk. . .lcell denotes an average over all resolution volumesassigned to a cell.

For dual-polarized Doppler radar, we make com-bined use of dual-polarization and Doppler quantities.Areas consisting of non-bird scatterers are removedbased on a high cross-polar correlation coefficientrHV . 0.9 or a high differential reflectivity, ZDR . 3.0.Height layers with low radial velocity standard devi-ations sr , 2 m s21 are rejected like in singlepolarization Doppler radar, which discards insectechoes for which dual-polarization moments often over-lap with those of birds.

To exclude spurious echoes owing to ground clutter,static clutter maps were generated for each weatherradar by performing a reflectivity average over severalclear-air days outside the bird migration season foreach resolution volume. All resolution volumes with areflectivity average above 210 dBZe during these daysare rejected permanently. A dynamic clutter map wasimplemented that excludes resolution volumes with aDoppler velocity in the interval [21,1] m s21, which fil-ters out most echoes related to anomalous beampropagation [25].

A.2. Bird radar methods

The bird radar has been stationed within themeasurement volume weather radars in De Bilt,Wideumont and in Trappes. Hourly migration trafficrates were determined at the airfield of Soesterberg(52.13 N/5.28 E, 10 m MSL), the airfield of StHubert (50.03 N/5.44 E, 577 m MSL) and at Flinssur Seine (48.58 N/1.52 E, 59 m MSL) at 6, 12 and24 km distance from the respective weather radarsites.

Fixed beam measurements [20] were carried out atthree elevation angles (5.68, 22.58, 798), which alloweda good coverage of all flight altitudes up to 6 km. Thebeam was directed towards WNW (2938), thus perpen-dicular to the main migratory direction. During night-time, the three elevations were monitored every halfhour between 17.00 and 05.00 UTC, and for the restof the day every hour. Detection range was restrictedto 7.5 km and recording time for a single elevationwas 4 min. Echoes showing wing beat patterns corre-sponding to birds were automatically selected by a‘vector support machine’ [21], which was trainedbased on a large sample of visually classified targetsby an expert.

Between fixed beam measurements, the radar wasoperated in tracking mode and tracks of single tar-gets were recorded for at least 20 s to determineflight paths. During the night, this was performedby an automatic search algorithm randomly selectingtargets from all relevant heights. During daytime,tracking was performed manually with an operatorselecting targets on the radar screen and anobserver watching through a telescope mountedparallel to the radar antenna. Tracked targets wereidentified by the observer in the best case tospecies level.

Bird densities are calculated from the number ofrecorded echoes, the average bird ground speed vectorand the calibrated bird- and aspect-specific surveyedvolume [20].

REFERENCES

1 Gauthreaux, S. A. & Belser, C. G. 2003 Radar ornithologyand biological conservation. Auk 120, 266–277.

2 Gauthreaux, S. A. & Belser, C. G. 2005 Radar ornithologyand the conservation of migratory birds. Technical report,USDA Forest Service.

3 Diehl, R. H. & Larkin, R. P. 2005 Introduction to theWSR-88D (NEXRAD) for ornithological research. InBird conservation implementation and integration in theAmericas: Proc. 3rd Int. Partners in Flight Conf. 2002,Asilomar, CA, 20–24 March 2002, pp. 876–888. U.S.Department of Agriculture Forest Service Pacific South-west Research Station.

4 Eastwood, E. 1967 Radar ornithology. London, UK:Methuen.

5 Alerstam, T. 1990 Bird migration. Cambridge, UK: Cam-bridge University Press.

6 Bruderer, B. 1997 The study of bird migration by radar.Part I: the technical basis. Naturwissenschaften 84, 1–8.(doi:10.1007/s001140050338)

7 Bruderer, B. 1997 The study of bird migration by radar.Part II: major achievements. Naturwissenschaften 84,45–54. (doi:10.1007/s001140050348)

8 Erni, B., Liechti, F., Underhill, L. G. & Bruderer, B. 2002Wind and rain govern the intensity of nocturnal birdmigration in central Europe—a log-linear regression analy-sis. Ardea 90, 155–166.

9 Alerstam, T. 1978 Reoriented bird migration in coastalareas: dispersal to suitable resting grounds? Oikos 30,405–408.

10 Buurma, L. S. 1995 Long-range surveillance radars as indi-cators of bird numbers aloft. Isr. J. Zool. 41, 221–236.

11 Bruderer, B. 1996 Vogelzugforschung im bereich derAlpen 1980–1995. Ornithol. Beob. 93, 119–130.

12 Holleman, I., Delobbe, L. & Zgonc, A. 2008 Update on theEuropean weather radar network (OPERA). In Proc. 5thEur. Conf. on radar in meteorology and hydrology,Helsinki, Finland, 30 June–4 July 2008. FinnishMeteorological Institute.

13 Wilson, J. W., Weckwerth, T. M., Vivekanandan, J.,Wakimoto, R. M. & Russell, R. W. 1994 Boundary layerclear-air radar echoes: origin of echoes and accuracyof derived winds. J. Atmos. Oceanic Technol. 11,1184–1206. (doi:10.1175/1520-0426(1994)011,1184:BLCARE.2.0.CO;2)

14 Martin, W. J. & Shapiro, A. 2007 Discrimination of birdand insect radar echoes in clear air using high-resolutionradars. J. Atmos. Oceanic Technol. 24, 1215–1230.(doi:10.1175/JTECH2038.1)

15 Contreras, R. F. & Frasier, S. J. 2008 High-resolutionobservations of insects in the atmospheric boundarylayer. J. Atmos. Oceanic Technol. 25, 2176–2187.(doi:10.1175/2008JTECHA1059.1)

16 Gauthreaux, S. A. & Belser, C. G. 1998 Displays of birdmovements on the WSR-88D: patterns and quantification.Weather Forecast. 13, 453–464. (doi:10.1175/1520-0434(1998)013,0453:DOBMOT.2.0.CO;2)

17 Diehl, R. H., Larkin, R. P. & Black, J. E. 2003 Radarobservations of bird migration over the great lakes. Auk120, 278–290.

18 Gauthreaux, S. A., Belser, C. G. & Blaricom, D. V. 2003Using a network of WSR-88D weather surveillance

Bird migration flight altitudes A. M. Dokter et al. 41

J. R. Soc. Interface (2011)

on February 16, 2011rsif.royalsocietypublishing.orgDownloaded from

radars to define patterns of bird migration at large spatialscales. In Avian Migration, pp. 335–346. Germany:Springer.

19 Bruderer, B., Steuri, T. & Baumgartner, M. 1995Short-range high-precision surveillance of nocturnalmigration and tracking of single targets. Isr. J. Zool. 41,207–220.

20 Schmaljohann, H., Liechti, F., Bachler, E., Steuri, T. &Bruderer, B. 2008 Quantification of bird migration byradar–a detection probability problem. Ibis 150, 342–355. (doi:10.1111/j.1474-919X.2007.00797.x)

21 Zaugg, S., Saporta, G., van Loon, E., Schmaljohann, H. &Liechti, F. 2008 Automatic identification of bird targetswith radar via patterns produced by wing flapping.J. R. Soc. Interface 5, 1041–1053. (doi:10.1098/rsif.2007.1349)

22 Limpens, H., Mostert, K. & Bongers, W. 1997 Atlas van deNederlandse vleermuizen. Utrecht, The Netherlands:KNNV Uitgeverij.

23 Nebuloni, R., Capsoni, C. & Vigorita, V. 2008 Quantifyingbird migration by a high-resolution weather radar. IEEETrans. Geosci. Remote Sens. 46, 1867–1875. (doi:10.1109/TGRS.2008.916467)

24 Koistinen, J. 2000 Bird migration patterns on weatherradars. Phys. Chem Earth B 25, 1185–1194. (doi:10.1016/S1464-1909(00)00176-3)

25 Doviak, R. J. & Zrnic, D. S. 1993 Doppler radar andweather observations, 2nd edn. New York, NY: AcademicPress.

26 Holleman, I. 2005 Quality control and verification ofweather radar wind profiles. J. Atmos. Oceanic Technol.22, 1541–1550. (doi:10.1175/JTECH1781.1)

27 Drake, V. A. & Farrow, R. A. 1988 The influence of atmos-pheric structure and motions on insect migration. Annu.Rev. Entomol. 33, 183–210.

28 Chapman, J. W., Reynolds, D. R. & Smith, A. D. 2003Vertical-looking radar: a new tool for monitoringhigh-altitude insect migration. Bioscience 53, 503–511. (doi:10.1641/0006-3568(2003)053[0503:VRANTF]2.0.CO;2)

29 Zhang, P., Liu, S. & Xu, Q. 2005 Identifying Dopplervelocity contamination caused by migrating birds,part I: feature extraction and quantification. J. Atmos.Oceanic Technol. 22, 1105–1113. (doi:10.1175/JTECH1757.1)

30 van Gasteren, H., Holleman, I., Bouten, W., van Loon, E. &Shamoun-Baranes, J. 2008 Extracting bird migration infor-mation from C-band weather radars. Ibis 150, 674–686.(doi:10.1111/j.1474-919x.2008.00832.x)

31 Holleman, I., van Gasteren, H. & Bouten, W. 2008 Qualityassessment of weather radar wind profiles during birdmigration. J. Atmos. Oceanic Technol. 25, 2188–2198.(doi:10.1175/2008JTECHA1067.1)

32 Larkin, R. P. & Diehl, R. H. 2002 Spectrum width of flyinganimals on Doppler radar. In Proc. 15th Conf. on Biome-teorology/Aerobiology and 16th Congress ofBiometeorology, Kansas City, MO. Boston: AmericanMeteorological Society.

33 Buler, J. J. & Diehl, R. H. 2009 Quantifying bird densityduring migratory stopover using weather surveillanceradar. IEEE Trans. Geosci. Remote Sens. 47, 2741–2751. (doi:10.1109/TGRS.2009.2014463)

34 Waldteufel, P. & Corbin, H. 1979 On the analysis ofsingle-Doppler radar data. J. Appl. Meteor. 18, 532–542.(doi:10.1175/1520-0450(1979)018,0532:OTAOSD.2.0.CO;2)

35 Holleman, I. & Beekhuis, H. 2003 Analysis and correctionof dual-PRF velocity data. J. Atmos. Oceanic Technol. 20,

443–453. (doi:10.1175/1520-0426(2003)20,443:AACODP.2.0.CO;2)

36 Bringi, V. N. & Chandrasekar, V. 2001 Polarimetric Dop-pler weather radar. Cambridge, UK: CambridgeUniversity Press.

37 Mueller, E. A. & Larkin, R. P. 1985 Insects observed usingdual-polarization radar. J. Atmos. Oceanic Technol.2, 49–54. (doi:10.1175/1520-0426(1985)002,0049:IOUDPR.2.0.CO;2)

38 Achtemeier, G. L. 1991 The use of insects as tracers for‘clear-air’ boundary-layer studies by Doppler radar.J. Atmos. Oceanic Technol. 8, 746–765. (doi:10.1175/1520-0426(1991)008,0746:TUOIAT.2.0.CO;2)

39 Zrnic, D. S. & Ryzhkov, A. V. 1998 Observations ofinsects and birds with a polarimetric radar. IEEE Trans.Geosci. Remote Sens. 36, 661–668. (doi:10.1109/36.662746)

40 Bachmann, S. & Zrnic, D. 2007 Spectral density of polari-metric variables separating biological scatterers in theVAD display. J. Atmos. Oceanic Technol. 24, 1186–1198. (doi:10.1175/JTECH2043.1)

41 Gauthreaux, S. A. & Belser, C. G. 1999 Reply to Blackand Donaldson’s comment on ‘Display of bird movementson WSR-88D: patterns and quantification’. WeatherForecast. 14, 1041–1042. (doi:10.1175/1520-0434(1999)014,1041:R.2.0.CO;2)

42 Black, J. E. & Donaldson, N. R. 1999 Comments on ‘Dis-play of bird movements on the WSR-88D: patterns andquantification’. Weather Forecast. 14, 1039–1040.(doi:10.1175/1520-0434(1999)014,1039:CODOBM.2.0.CO;2)

43 Caya, D. & Zawadzki, I. 1992 VAD analysis of nonlinearwind fields. J. Atmos. Oceanic Technol. 9, 575–587.(doi:10.1175/1520-0426(1992)009,0575:VAONWF.2.0.CO;2)

44 Boccippio, D. J. 1995 A diagnostic analysis of the VVPsingle-Doppler retrieval technique. J. Atmos. OceanicTechnol. 12, 230–248. (doi:10.1175/1520-0426(1995)012,0230:ADAOTV.2.0.CO;2)

45 Browning, K. A. & Wexler, R. 1968 The determination ofkinematic properties of a wind field using Doppler radar.J. Appl. Meteor. 7, 105–113.

46 Alerstam, T., Rosen, M., Backman, J., Ericson, P. G. P. &Hellgren, O. 2007 Flight speeds among bird species: allo-metric and phylogenetic effects. PLoS Biol. 5, e197.(doi:10.1371/journal.pbio.0050197)

47 Reynolds, A. M., Reynolds, D. R., Smith, A. D. &Chapman, J. W. 2010 A single wind-mediatedmechanism explains high-altitude ‘non-goal oriented’headings and layering of nocturnally migrating insects.Proc. R. Soc. B 277, 765–772. (doi:10.1098/rspb.2009.1221)

48 Dunning, J. B. 1993 CRC handbook of avian body masses.Boca Raton, FL: CRC Press, Inc.

49 Alerstam, T. 1976 Nocturnal migration ofthrushes (Turdus spp.) in southern Sweden. Oikos 27,457–475.

50 Lensink, R., van Gasteren, H., Hustings, F., Buurma, L.,van Duin, G., Linnartz, L., Vogelzang, F. & Witkamp,C. 2002 Vogeltrek over Nederland 1976–1993. Haarlem:Schuyt & Co.

51 Chapman, J. W., Nesbit, R. L., Burgin, L. E., Reynolds,D. R., Smith, A. D., Middleton, D. R. & Hill, J. K. 2010Flight orientation behaviours promote optimal migrationtrajectories in high-flying insects. Science 327, 682–685.(doi:10.1126/science.1182990)

52 Vaughn, C. R. 1985 Birds and insects as radar targets.Proc. IEEE 73, 205–227.

42 Bird migration flight altitudes A. M. Dokter et al.

J. R. Soc. Interface (2011)

on February 16, 2011rsif.royalsocietypublishing.orgDownloaded from

53 Unden, P. et al. 2002 HIRLAM-5 scientific documentation.Technical report. Swedish Meteorological and HydrologicalInstitute (SMHI).

54 Bruderer, B. & Liechti, F. 1995 Variation in density andheight distribution of nocturnal migration in the southof Israel. Isr. J. Zool. 76, 1106–1118.

55 Liechti, F. 2006 Birds: blowin’ by the wind? J. Ornithol.147, 202–211.

56 Peterson, A. T. & Williams, R. A. J. 2008 Risk mapping ofhighly pathogenic avian influenza distribution and spread.Ecol. Soc. 13, 15.

57 Erni, B., Liechti, F. & Bruderer, B. 2005 The role of windin passerine autumn migration between Europe andAfrica. Behav. Ecol. 16, 732–740. (doi:10.1093/beheco/ari046)

58 Vrugt, J. A., van Belle, J. & Bouten, W. 2007Pareto front analysis of flight time and energyuse in long-distance bird migration. J. AvianBiol. 38, 432–442. (doi:10.1111/j.0908-8857.2007.03909.x)

59 Gonzales, R. C. & Woods, R. E. 1992 Digital imageprocessing. Reading, MA: Addison-Wesley.

Bird migration flight altitudes A. M. Dokter et al. 43

J. R. Soc. Interface (2011)

on February 16, 2011rsif.royalsocietypublishing.orgDownloaded from