historical drifter data and statistical prediction of...

Historical Drifter Data and Statistical Prediction of Particle Motion: A Case Study inthe Central Adriatic Sea

MILENA VENEZIANI*

RSMAS/MPO, University of Miami, Miami, Florida

ANNALISA GRIFFA

RSMAS/MPO, University of Miami, Miami, Florida, and ISMAR/CNR, La Spezia, Italy

PIERRE-MARIE POULAIN

OGS, Trieste, Italy

(Manuscript received 12 January 2006, in final form 30 May 2006)

ABSTRACT

In this paper, a method to analyze historical surface drifter data is presented that is aimed at investigatingparticle evolution as a function of initial conditions. Maps of drifter concentration at different times are builtand interpreted as maps of the probability of finding a particle at a given time in the neighborhood of agiven point in the domain. A case study is considered in a coastal area of the middle Adriatic Sea (a subbasinof the Mediterranean Sea) around the Gargano Cape, which is the focus of a newly planned experiment, theDynamics of the Adriatic in Real Time (DART). A specific application is considered that seeks to improvethe DART Lagrangian sampling planning.

The results indicate that the analysis of historical drifters can provide very valuable information onstatistical particle prediction to be used in experiment design. In the DART region, particle dynamicsappear mostly controlled by the upstream properties of the boundary current as well as by the presence ofa stagnation point located offshore of the tip of Gargano and separating two cross-basin recirculations. Asignificant seasonal dependence is observed, with drifters being more likely to leave the boundary currentin winter and fall, when the current is wider and more connected to the cross-basin recirculations. Futuredevelopments are discussed, including joint analyses with numerical model results.

1. Introduction

Predicting particle motion in the ocean is of greatimportance for a number of practical applications.These applications include prediction of pollutantspreading, oil spill containment, and search-and-rescueactivities, as well as assessment of biological quantitiessuch as larvae spreading, or warfare exercises like minetracking.

Particle prediction is a very challenging problem be-cause of the well-known conceptual reason that trajec-tories are extremely sensitive to the details of the ad-vective Eulerian velocity field and are highly chaotic(e.g., Aref 1984; Samelson 1996). It is therefore notsurprising that, despite the great advances in ocean cir-culation models in the last two decades, quantitativelypredicting the motion of single particles remains ahighly complex and demanding problem (e.g., Dunlapet al. 2004; Thompson et al. 2003). Even small errors inthe estimation of ocean currents can drastically changeparticle trajectories (Griffa et al. 2004). Furthermore,sources of errors in modeling ocean currents are ubiq-uitous due to uncertainties in the wind forcing, bathym-etry, coastline, and internal stratification, and themodel resolution and subgrid-scale parameterizations.Similar difficulties are found when particle prediction isperformed on the basis of extensive velocity datasets,

* Current affiliation: Ocean Sciences Department, Universityof California, Santa Cruz, Santa Cruz, California.

Corresponding author address: Dr. Milena Veneziani, OceanSciences Department, University of California, Santa Cruz, SantaCruz, CA 95064.E-mail: [email protected]

FEBRUARY 2007 V E N E Z I A N I E T A L . 235

DOI: 10.1175/JTECH1969.1

© 2007 American Meteorological Society

JTECH1969

such as current maps constructed with high-frequencycoastal radars (J. O’Donnell et al. 2006, personal com-munication; Spaulding et al. 2005). Uncertainties in theobserved velocity and the presence of unresolved scalesof motion can lead to significant errors in particle pre-diction. While it can be expected that future improve-ments in both models and measurements and in theircombined use through data assimilation will lead to sig-nificant enhancement of particle prediction, the presentstate of the art is still problematic.

Given the difficulties in deterministically predictingthe motion of single particles, in practical applications astatistical point of view is often adopted. Instead oftrying to reproduce the exact motion of single trajecto-ries, information is sought on the probability of findinga particle in a specific region and at a given time. In theframework of numerical modeling, this problem hasbeen treated using two main approaches. The “stochas-tic” approach (e.g., Thomson 1987; Dutkiewicz et al.1993; Paris et al. 2005) consists of perturbing the nu-merical trajectories predicted by a circulation modelusing a random noise, which represents the unresolvedmotion and is parameterized through a Lagrangian sto-chastic model (LSM). An ensemble of particle trajec-tories is generated for each initial condition, and theensemble evolution in space and time provides infor-mation on the evolution of the probability density func-tion (PDF). The “dynamical system” approach (e.g.,Wiggins 1992; Poje and Haller 1999; Kuznetsov et al.2002), on the other hand, focuses on the generalLagrangian structure of the flow generated by a circu-lation model, identifying its hyperbolic points andmanifolds, which tend to act as transport barriers. En-sembles of initial conditions (ICs) can be identified andare characterized by similar Lagrangian characteristicsor transport fates. Both the stochastic and the dynami-cal system approaches are promising, but they stillpresent a number of open questions. In the former,questions remain on the identification of the correctLSM and of its parameter values in order to effectivelyparameterize the unresolved motion. In the latter, de-velopment is still needed in the case of realistic flowswith high transients, where significant geometricalproperties might be difficult to identify and character-ize.

In this paper, we propose an approach to the prob-lem of statistical particle prediction based on the analy-sis of in situ historical drifter data. The methodologycan be considered complementary to the mentionedmodeling approaches, and is applied to a specific casestudy in the central Adriatic Sea, a subbasin of theMediterranean Sea, where a significant historical

drifter dataset has been collected over the years (e.g.,Poulain 2001). We focus on the region of a new experi-ment that is presently being planned, the Dynamics ofthe Adriatic in Real Time (DART; see information on-line at http://oceans.deas.harvard.edu/haley/DART05).The DART region is in the coastal area of the middleAdriatic, close to the Gargano Cape. It is an area withstrong topographic control and significant mesoscaleactivity related to instabilities of the boundary current.The DART experiment has an important Lagrangiancomponent and therefore provides an additional andmore specific motivation to the present study. We ex-pect to use our results as guidance for the planning ofthe experimental Lagrangian sampling strategy.

The goal of the paper is to describe a methodologyand its application to a coastal region. We aim at ad-dressing the following question: given a particlelaunched at point x0, what is the probability of finding itin the neighborhood of point x of the domain at a latertime t after the deployment? In practice, this is accom-plished by identifying some selected regions within theDART area and analyzing the evolution of ensemblesof historical drifters launched in these regions. Concen-tration maps of drifter positions are built as functions ofspace and time, and they are interpreted in conjunctionwith other more general statistics, such as the meanflow. The statistics are first performed considering thewhole ensemble of historical drifters, while conditionalstatistics are later introduced depending on the drifters’seasonal coverage. These concentration maps can beinterpreted as PDFs for particle positions, and they canbe utilized for several applications. For environmentalproblems, for instance, they can be used as “risk maps”in the case of pollutant release, indicating the mostlikely spread of pollutant concentration in space andtime. For experimental planning, as in the DART case,the maps can be used to identify the most likely cover-age obtained from a given release scheme, thereforeproviding information on the sampling strategy.

Because in situ datasets are always quite limited com-pared with numerical datasets, even in highly sampledareas, we expect that the information will be relativelycoarse, but it will be able to provide the range of ap-propriate space and time scales to be considered inmost practical applications. Also, we expect that theresults will set the ground for further investigations inconjunction with modeling studies, for instance, interms of validating model results and giving insightsinto the main factors contributing to uncertainties inthe modeled trajectories.

The paper is organized as follows. Section 2 gives abrief background of the Adriatic dataset and the maincharacteristics of the basin. In section 3, the specific

236 J O U R N A L O F A T M O S P H E R I C A N D O C E A N I C T E C H N O L O G Y VOLUME 24

DART area of interest is introduced and the method-ology used to analyze the data is described. The resultsof the analysis are presented in section 4. Conclusionsand a brief discussion of the implications for the DARTLagrangian sampling strategy are presented in sec-tion 5.

2. The Adriatic Sea and the historical dataset

a. Adriatic morphology and general circulation

The Adriatic Sea is the northernmost semienclosedbasin of the Mediterranean Sea, and is connected to theIonian Sea through the Strait of Otranto at its southernboundary (see Fig. 1 for geographic locations men-tioned herein). It can be separated into the followingthree different subbasins: 1) the southern Adriatic be-tween the Strait of Otranto (to the south) and thePalagruza Sill, with bottom topography as deep as 1200m in its center, called the South Adriatic Pit (SAP); 2)the middle Adriatic north of the Palagruza Sill, calledthe Pomo, Jabuka, or Mid-Adriatic Pit (MAP), whichcontains depressions as deep as 270 m; and 3) the north-ern Adriatic, a very shallow region (�100 m) to the

north characterized by a gentle bathymetry slope. TheAdriatic coastline is relatively smooth on the Italianside with the exception of two major capes—the Con-ero Promontory and the Gargano Cape. In contrast, theeastern coast is rugged and includes hundreds ofislands, and the bottom topography is often verysteep.

Hydrographic data (Artegiani et al. 1997) and driftertracks (Poulain 1999, 2001; Ursella et al. 2006, hereafterURS06) show that the mean large-scale, near-surfacecirculation is cyclonic, with water entering on the east-ern flank of the Strait of Otranto and continuing north-ward along the Albanian and Croatian coasts as a gen-erally weak and wide Eastern Adriatic Current (EAC).The EAC separates into three embedded cyclonic cir-culation structures by recirculating near the two north-ern slopes of the SAP and MAP and south of the IstrianPeninsula. All of these branches feed into the WesternAdriatic Current (WAC), which extends almost con-tinuously along the Italian coast from the Po Riverdelta in the northern subbasin to the Strait of Otranto.The WAC is swift and concentrated near the coast orminishelf slope and is mainly forced by the buoyancyinput from the Po River. Significant seasonal variability

FIG. 1. Description of geographic locations in the Adriatic Sea. The topography of the areais also shown (with 50-m contour levels).


of the WAC has been observed with enhanced flow tothe southeast in winter and reduced flow in summer.Wind forcing also affects the WAC: during events ofstrong northeasterly bora winds, the recirculation southof the Istrian Peninsula and the WAC are reinforced; inconditions of southeasterly sirocco winds, the EAC isstronger and the WAC weakens and even reverses(Poulain et al. 2004b).

b. Adriatic historical drifter data

As part of several research and military programs,satellite-tracked surface drifters have been released inthe Adriatic Sea since 1990 and have provided datauntil 1999. Maximum numbers of drifters occurred in1995 and 1997–98 in concomitance with programsfunded by the North Atlantic Treaty Organization(NATO) and the U.S. Office of Naval Research(ONR), respectively (Poulain 2001). The majority ofthese drifters were released in the southern Adriaticand the Strait of Otranto area. More drifters were de-ployed in the northern and middle Adriatic betweenSeptember 2002 and November 2003 as part of theONR Dynamics of Localized Currents and EddyVariability in the Adriatic (DOLCEVITA) and NATOAdriatic (ADRIA02/03) projects (URS06).

Most of the Adriatic drifters are Coastal Ocean Dy-namics Experiment (CODE) drifters that measure thesurface currents within the first meter of water (Davis1985). Their water-following characteristics have beenassessed (P.-M. Poulain and L. Ursella 2006, unpub-lished manuscript), and it was found that they followthe surface current within 2 cm s�1. All drifters weretracked with the Argos Data Location and CollectionSystem (DCLS) carried by the National Oceanic andAtmospheric Administration (NOAA) polar-orbitingsatellites. The drifter position data (latitude and longi-tude) were edited, interpolated, and low-pass filtered(with a cutoff period of 36 h). The low-pass time serieswere subsampled every 6 h and the velocity was com-puted by finite-centered differencing the 6-hourly inter-polated/filtered position data. Details on the drifterdata processing can be found in Poulain (2001), Poulainet al. (2004a, 2006), and URS06. The drifters used inthis study correspond mostly to the years 1992, 1995,1997–98, and 2002–03.

3. The DART area of interest and themethodology

The DART area of interest is situated in the middleAdriatic, around the Gargano Cape (Fig. 1). It is char-acterized by a complex topography (see also Fig. 4),with the Palagruza Sill extending from the cape toward

the eastern coast of the basin and dividing the two deepdepressions of the MAP (to the north) and the SAP (tothe south). A well-defined shelf can be seen along theItalian coast, approximately extending up to the 150-misobath and widening south of the Gargano Cape in theGulf of Manfredonia.

One of the main goals of the DART experiment is tostudy the mesoscale instabilities that form in the WACboundary current near the Gargano Cape and their in-fluence on the downstream circulation. To this end, asuite of different measurements will be used duringDART and complemented by a near-real-time model-ing effort (see information online at http://oceans.deas.harvard.edu/haley/DART05). An array of currentmeters will be maintained for extended periods (orderof months), while two cruises of 2–3 weeks in winterand summer 2006, respectively, will provide high-resolution synoptic hydrographic data as well as micro-structure data. During the first cruise in March 2006, atotal of approximately 20 surface drifters will be de-ployed in the Gargano Cape area. The deploymentscheme should provide an approximately homogeneouscoverage in space and time during the cruise period.The historical data analysis presented in this paper willcontribute to identifying the suitable deploymentpoints, also providing estimates of the expected timescales for drifters to properly cover the area of interest.

As a first step, we subdivide the DART area intoappropriate dynamical subregions. The regions are cho-sen based on the mean flow and on the eddy kineticenergy (EKE) field computed by averaging the wholeensemble of historical drifter velocities over 0.1° squarebins (Fig. 2). The mean flow field (Fig. 2a) clearly showsthe WAC boundary current flowing southeast along thecoast. The current extends approximately up to the160-m isobath upstream of the Gargano Cape, and hasa well-defined core within the 100-m isobath from thecoast. Downstream of the cape, the current widens, fol-lowing the shelf. A small recirculation region can beseen in the lee of the cape, with reduced sampling sug-gestive of a “shadow zone” resulting from the classicaldetachment of a boundary current in the presence of anobstacle (Doglioli et al. 2004). The WAC appears con-nected with the interior flow through two branches ofthe cross-basin recirculations, along the slopes of theMAP and SAP, respectively. Also, a stagnation point inthe mean flow field can be seen on the Palagruza Sill,dividing the MAP and SAP recirculations. In the fol-lowing we will loosely refer to this point as the “hyper-bolic point,” because a hyperbolic (or saddle) point isindeed expected to be present in the specific flow real-izations (Poje and Haller 1999). The EKE field (Fig.2b) indicates that fluctuations are intense in the WAC,


especially near the tip of the Gargano Cape and at thesoutheastern edge of the DART area where the shelfnarrows and the current “reattaches” to the coast.

Based on this information, five regions have beenchosen and their boundaries are shown in Fig. 2 (thinblack lines). Two regions—north polygon (POLYN)and south polygon (POLYS)—are the polygons locatedinside the WAC upstream of the Gargano Cape. Al-though POLYN lies almost entirely outside the DARTarea, it has been included in the analysis because it isrepresentative of the structure of the boundary currentbefore being influenced by the cape, and it can there-fore provide an interesting upstream control point. Lo-gistically, ship transects during the cruises are expectedto extend up to POLYN, making deployments possiblein the region. The other three regions are locateddownstream of the Gargano Cape: the south GarganoCape (GARGS) covers the WAC pattern southward ofthe cape; the north Gargano Cape (GARGN) includesthe more complex region close to the hyperbolic pointand is influenced by the interior recirculation branch;and the Gargano recirculation (RECIRC) covers therecirculation region in the lee of the cape, in the Gulf ofManfredonia.

For each of these regions, we have selected all thedrifters that have either been deployed in the region orhave entered it at some time during their life. The de-ployment and entrance points are considered as initialconditions, that is, of independent trajectory realiza-tions. Because the number of actual deployments isvery small with respect to the entrance points, almostall of the ICs are located on the boundaries of the re-gions (see Figs. 4, 7). Notice that considering the en-trance points as independently given initial conditionsis an approximation because the entrance points de-pend on the flow field that advected the drifters. Also,in the practical implementation, a number of drifterscross more than one region in the DART area and theirtrajectories are therefore included in the analysis ofeach of the covered regions. Whenever possible, theICs in each region are further grouped to provide en-sembles of dynamically consistent initial conditions.For example, in POLYN and POLYS (see section 4a),the ICs in the core of the boundary current are sepa-rated from those in the more external part.

The evolution of particle motion for each IC group isstatistically characterized using a number of diagnos-tics. First, a qualitative assessment is performed by plot-ting the corresponding “forward” and “backward” tra-jectories for each region of the DART area. If thedrifter IC occurs at a time t0, the forward (backward)trajectories are those corresponding to the time periodt0 � t � t0 � tforw (t0 � tback � t � t0), where tforw (tback)

FIG. 2. (top) Mean flow obtained by averaging the historicaldrifter velocities over 0.1° square bins. Gray vectors are for binswith more than 5 but less than 10 independent measurements, n*(where n* is computed considering a Lagrangian decorrelationtime scale TL � 2 days); black vectors are for bins with more than10 independent measurements. The standard error ellipses arecomputed with respect to the major and minor axis of variability.Also shown are the 100- and 160-m isobaths. (bottom) EKE fieldobtained by averaging the Lagrangian eddy velocities over 0.1°square bins. Only results for bins with n* � 5 are shown. The thickblack lines identify the DART area of interest, while the thin linesidentify the following 5 subregions: POLYN (north polygon),POLYS (south polygon), GARGN (north Gargano Cape),GARGS (south Gargano Cape), and RECIRC (Gargano recircu-lation).


is a fixed interval. Clearly, if the IC is an actual deploy-ment, only the forward trajectory is computed. Visualinspection of the trajectory ensembles provides a directand general assessment of the transport from/to a cer-tain region.

A more quantitative analysis is then performed onthe forward trajectories, computing their concentrationmaps or PDFs over a 0.2° � 0.2° spatial grid at fixedtimes tforw. In each grid cell, the concentration is com-puted as the ratio between the number of drifterspresent in the cell at the time tforw, and the total numberof drifters in the considered IC group. The PDF mapsare complemented by other bulk descriptors, such asarrival times and ensemble properties. These analysesare first performed over the whole ensemble of driftertrajectories, and then are repeated considering seasonaldependence.

4. Results

A first characterization of the particle evolution isgiven in Fig. 3 by the forward and backward trajectories(tforw � tback � 10 days) for three significant regions:POLYN in the WAC boundary current upstream of theGargano Cape, and GARGN and RECIRC down-stream of the cape. POLYS and GARGS are qualita-tively similar to their corresponding northern counter-parts.

As can be seen, particles entering POLYN originatefrom the northern WAC and from the recirculatingcross-basin branch along the MAP. Once they enter theboundary current, drifters tend to stay in it, movingsoutheastward following the bathymetry. Some escapepoints are visible, mostly upstream of the cape and at itstip. Particles entering GARGN also originate from twodifferent locations: the upstream WAC and the recir-culating SAP branch. After entering GARGN, they aremostly entrained in the boundary current movingsoutheastward, even though a clear escape point can beseen offshore of the tip of the Gargano Cape, withdrifters moving northward. The ensemble of ICs inRECIRC is reduced, because few drifters enter the leeof the cape. They mostly originate from the upstreamboundary current and skirt the tip of the cape as theymove downstream. They tend to recirculate once ortwice in the region and then move southeastward tojoin the boundary current.

A more quantitative description of the forward mo-tion is given in the following sections, which considerthe regions upstream and downstream of the cape sepa-rately.

a. Regions upstream of the Gargano Cape

The drifters’ ICs in the two upstream regions,POLYN and POLYS, are shown in the upper panels ofFig. 4. They are further divided in two groups for eachregion: BCint (filled circles) includes the ICs in theinternal part of the WAC, approximately defined as thearea within the 100-m isobath from the coast, whileBCext (open circles) includes the ICs in the externalpart of the WAC, between the 100- and 160-m isobaths.The temporal evolution of the PDF for the BCint drift-ers of POLYN is shown in Fig. 5. The results forPOLYS (not shown) are qualitatively similar. In addi-tion to the PDF map corresponding to a particular tforw

time value, each panel of Fig. 5 indicates the percentageof drifters that are inside the DART area and the per-centage of those that have stopped working at the timetforw. Drifters “die” for various reasons, from batteryfailure, to beaching, and being caught in fishermen nets.Their “mortality” has to be taken into account becauseit can significantly alter the number of realizations andeven bias the statistics (e.g., Falco et al. 2000). The totalnumber of drifters in BCint is initially equal to 58, while23 drifters are already dead at tforw � 15 days. Thismortality rate is rather high compared with the typicallifetime of a drifter in the Adriatic Sea [mean half-lifeof 35–40 days, see Poulain (2001) and URS06].

The PDF maps in Fig. 5 show the spreading of thedrifters as they move southeastward in the boundarycurrent. The spreading is mostly alongshore and it islikely to be due to the influence of the sheared current,as well as to the effect of having considered many dif-ferent realizations in computing the statistics. After 2days, a few drifters have already reached the tip of theGargano Cape, while the maximum of the concentra-tion is still located at the boundary between POLYNand POLYS. Most of the drifters (70%) are found inthe DART area after 8 days, while only 43% are stillpresent after 15 days. While most of them have exitedthe area following the current to the south, an escapepoint at the tip of the cape can also be seen.

Drifters with ICs in BCext (33 in total) have a quali-tatively similar behavior to those in BCint, but they alsoshow some significant differences (Fig. 6). The along-shore spreading is reduced with respect to BCint, prob-ably because of a weaker current shear. Also, particlesshow a higher probability of escaping the boundary cur-rent to enter the interior, so that a smaller percentage(54%) is found in the DART area after 8 days. Anescape point can be seen upstream of the cape, slightlynorth of the boundary between POLYN and POLYS.The presence of such an escape point is also suggestedby the mean flow structure in the corresponding loca-


tion (Fig. 2a), which shows a (weak) recirculation prob-ably connected to topographic features.

Results for POLYN and POLYS are summarized inTables 1 and 2. They are obtained by considering thewhole ensemble of initial conditions BCint and BCext,and focusing on a time period of 21 days, which ap-proximately corresponds to the duration of the plannedDART cruises. The statistics in Table 2 are aimed at

answering the following questions: how many driftersreach the tip of the Gargano Cape in 21 days; what istheir fate after they pass the tip; and, in particular, howmany drifters leave the boundary current to enter theAdriatic interior, and how many stay in the boundarycurrent or recirculate in the lee of the cape?

To answer these questions, two quantities ntot andndie_earlier are first computed for each region and are

FIG. 3. Forward and backward trajectories of the drifters from/to (top) POLYN, (middle) GARGN, and (bot-tom) RECIRC; the corresponding subregions are colored in gray for clarity. Both tforw and tback are equal to 10days.


listed in Table 1. The quantity ntot indicates the totalnumber of drifters with initial conditions in the region,while ndie_earlier indicates the number of drifters that dieduring the 21-day period without reaching the tip of thecape. The statistics in Table 2 are computed as percent-ages with respect to the difference between ntot andndie_earlier. The ndie_earlier drifters are removed because itis impossible to say whether or not they would havepassed the cape had they lived long enough. Specificanalyses to address the consequences of mortality andpossible associated biases (Falco et al. 2000) are out ofthe scope of this paper, and will be addressed in futureworks. The quantities in Table 2 indicate the percent-age of drifters that reach the tip of the cape (ngarg), thepercentage of those that migrate to the interior afterpassing the tip (ngarg2int), and the percentage of driftersthat enter the Gargano Cape recirculation (ngarg2rec).Also shown is the average arrival time Tm (in days)taken by the ngarg drifters to reach the tip of the cape.When computing these quantities, the tip of the Gar-gano Cape is identified with the eastern boundary ofPOLYS. Furthermore, ngarg2int and ngarg2rec are com-

puted by monitoring the trajectories inside the GARGS� GARGN area downstream of the Gargano Cape andby checking whether or not they permanently leave thearea during the 21-day period of interest.

The results in Table 2 are qualitatively similar forPOLYN and POLYS, indicating that the majority ofthe trajectories (83%–88%) reach the tip of the cape byremaining inside the boundary current (in agreementwith Figs. 5 and 6). After reaching the tip, approxi-mately 10% of the drifters leave the current and pen-etrate the interior, while the remaining �70% keepflowing southeastward in the WAC. Only �7% recir-culate in the lee of the cape. The arrival times Tm are9.3 and 5.8 days for POLYN and POLYS, respectively.While these are average times, the distribution of indi-vidual arrivals is shown by the histograms in the lowerpanels of Fig. 4. The distributions appear quite flat,especially for POLYN, in agreement with the along-shore spreading shown in Fig. 5, while for POLYS asmall peak can be seen at 2–4 days. Notice that POLYShas nonzero values in the class 0–2 days, mostly result-ing from a few ICs that lie very close to or even at the

FIG. 4. (top) Initial positions of the POLYN and POLYS drifters. Filled (open) circles identify the BCint (BCext)drifters. (bottom) Histograms of the arrival time taken by the ngarg drifters (see Table 2) to reach the Gargano Capefrom POLYN and POLYS.


eastern border of POLYS (Fig. 4). As a test, we haveremoved the ICs in the eastern half of POLYS and wehave found differences only in the first histogram class,while the statistics in Table 2 remain basically un-changed.

In summary, the results indicate that drifters with ICsin the WAC upstream of the cape have a very highprobability of reaching the cape and staying in theboundary current even after they have passed the tip.This probability is especially high for drifters with ICsin the core of the current, within the 100-m isobathfrom the coast. A quantitative prediction of the trajec-tories and of their arrival times at the tip of the cape is,on the other hand, hard to obtain because of the along-shore spreading related to the shear of the mean cur-rent and to the high variability of the drifter realiza-

tions. We will return to this specific point in section 4cwhen considering the effects of seasonality.

b. Regions downstream of the Gargano Cape

The drifters’ ICs in the three downstream regionsGARGN, GARGS, and RECIRC are shown in Fig. 7.The ICs for GARGN correspond mostly to entrancepoints from the WAC boundary current, including bothBCint and BCext. They are grouped together becausethe sampling is reduced with respect to POLYN andPOLYS (46 drifters in total), and also because theyexhibit a similar behavior. The PDFs in Fig. 8 show thatthe drifter concentration follows two distinct patterns.The main pattern is entrained in the boundary currentand moves southeastward, while a secondary concen-tration branch moves northward toward the Adriatic

FIG. 5. Maps of drifter concentration computed over a 0.2° � 0.2° spatial grid at various fixed times tforw for theBCint drifters from the POLYN region. The initial number of these drifters is 58. See text for percentage defini-tions.


interior. This is not surprising considering the structureof the mean flow (Fig. 2a) in the northwestern part ofGARGN (where the ICs are located), which is an areavery close to the hyperbolic point between the twocross-gyre recirculations. The two out-flowing branchesof the hyperbolic point correspond to the two mainconcentration patterns seen in Fig. 8 (see, e.g., the panelfor tforw � 4 days). After 8 days, only 34% of the driftersare still within the DART area, with drifters exitingboth to the southeast and to the north of the region.

In the GARGS analysis, all the available ICs (104drifters in total) have been considered because theyexhibit a similar behavior. They include drifters enter-ing from the WAC, as well as drifters entering from thecross-basin recirculation around the SAP (Fig. 7). Thetime evolution of the PDF (Fig. 9) is strikingly differentfrom that for GARGN (Fig. 8), with virtually all of thedrifters entrained in the boundary current and moving

FIG. 6. As in Fig. 5, but relative to the BCext drifters from POLYN (their initial total number is 33).

TABLE 1. Total number of drifters with ICs in POLYN andPOLYS (ntot), and number of drifters that die during a 21-dayperiod without reaching the tip of the Gargano Cape (ndie_earlier).

ntot ndie_earlier

POLYN 91 24POLYS 96 16

TABLE 2. Statistics obtained from the total drifters (BCint �BCext) for POLYN and POLYS over a period of 21 days. Thepercentages are computed with respect to the difference betweenntot and ndie_earlier previously displayed in Table 1. See text fordefinition of the displayed quantities.

ngarg ngarg2int ngarg2rec Tm (days)

POLYN 56 (83.6%) 7 (10.4%) 5 (7.5%) 9.3POLYS 71 (88.7%) 9 (11.2%) 6 (7.5%) 5.8


southeastward. A certain degree of alongshore disper-sion can be seen at tforw � 2 days, but it is much weakerthan in the upstream boundary current (POLYN; seeFig. 5), probably because the downstream current is

wider and less sheared. At the southeastern corner ofthe DART area, the current narrows and intensifies(Fig. 2a), following the topography. After 8 days, only28% of the drifters are still in the DART area, and mostof the drifters have exited toward the south.

The available ICs in RECIRC (23 drifters in total)are located along the boundary with GARGS (Fig. 7).The PDFs in Fig. 10 show that, while almost all thedrifters are trapped in RECIRC during the first 2 days,they start exiting the region after 4 days. Most trajec-tories loop inside RECIRC once or twice and they thenleave the lee of the Gargano Cape. Overall, drifterstend to stay longer in the DART area with respect todrifters for GARGS or GARGN. After 8 days, in fact,47% of the particles are still in the area. Nevertheless,the RECIRC drifters also eventually tend to be en-trained in the boundary current moving southeastwardand exiting the area. Notice that the sampling in thisregion is scarce, suggesting that the lee of the capetends to be isolated from the main boundary currentsystem for most of the time. The mechanisms that leadto the formation of the recirculation connecting the leeto the main current are not obvious at this time andthey will be investigated in future works.

In summary, drifters with ICs in the DART areadownstream of the cape have an overall tendency to beentrained in the boundary current. Only in the regionoffshore of the tip of the cape on the Palagruza Sill is avery high variability observed, with drifters being en-trained in a northward branch and moving toward theAdriatic interior. This is related to the presence of ahyperbolic point, which induces significant spatial vari-ability. Also, the exact position of the hyperbolic pointis likely to vary in time, so that particle evolution mayvary according to the specific realization. In this region,the statistical analysis of historical data can provideonly partial answers, and it is expected that the joint useof modeled data will lead to improved results.

c. Seasonal dependence

The results discussed in the previous sections are ob-tained using the whole historical dataset, and they showa significant variability in the evolution of drifter tra-jectories. To resolve this variability at least partially,conditional statistics have been considered as a functionof seasons. The temporal definition of the seasons usedhere is similar to the calendar as appears in URS06:spring is from 21 March to 20 June, summer from 21June to 20 September, fall from 21 September to 20December, and winter from 21 December to 20 March.

As a first step, the mean flow has been computed forthe four seasons. This is displayed in Fig. 11. Although

FIG. 7. Initial positions of the BCint and BCext drifters in (top)GARGN, and initial positions of the total drifters in (middle)GARGS and (bottom) RECIRC.


the data coverage is obviously reduced with respect toFig. 2, and many bins have less than 10 independentmeasurements, the results are instructive and clearlyshow the flow evolution. During fall and winter, theWAC is well defined, wide, and energetic. The cross-basin recirculations around the SAP and the MAP areevident, connecting the WAC to the interior and to theEAC. During spring and summer, the WAC appearsweaker and more coastal (also resulting from the lackof observations offshore). Recirculation patterns (MAPand SAP) are less evident and the connection of thecoastal flow with the Adriatic open sea is reduced. Asdiscussed in a number of previous works (Zore-Armanda 1956, 1969; Artegiani et al. 1997; Hopkins etal. 1999; URS06), the Adriatic circulation and the

WAC in particular are strongly influenced by Po Riverdischarges and by wind conditions. The Po River dis-charge is high in fall and winter with a maximumaround December–January (URS06), while it is re-duced in summer with a minimum around July and Au-gust. This is one of the main reasons for the reducedWAC strength during the summer season. Regardingthe wind forcing, strong northeasterly bora winds oc-curring mostly in fall and winter are associated withreinforced WAC and cross-basin recirculations. Duringsummer, the wind regime is more variable, and it in-cludes episodes of strong southeasterly sirocco windsthat tend to weaken the WAC and the cross-basin re-circulations (Poulain et al. 2004b).

The seasonal dependence of drifter evolution is stud-

FIG. 8. As in Fig. 5, but relative to the sum of the BCint and BCext drifters from GARGN (their initial totalnumber is 46).


ied focusing on the regions upstream of the GarganoCape because they have an extended sampling and theyshow a well-defined seasonal signal. The qualitative be-havior of drifters in POLYN is illustrated by the back-ward and forward trajectories in Fig. 12. In fall andwinter, consistent with the mean flow pattern (Fig. 11),drifters enter POLYN from the WAC and from theMAP recirculation, while the recirculation branchweakens in spring and almost disappears in summer.Forward trajectories tend to stay in the WAC movingsoutheastward, although some escape points are evi-dent, especially in the fall, mostly resulting from thenorthward branch of the recirculation. Summer trajec-tories are significantly more confined to the coast withrespect to the fall and winter periods.

The time evolution of the PDF for the POLYN drift-

ers is shown in Fig. 13. To have sufficient coverage, allof the ICs (BCint � BCext) are considered to computethe drifter concentrations. Furthermore, the maps aredrawn for two extended seasons: fall–winter andspring–summer. The sampling is more reduced inspring–summer, with a total of 38 drifters versus 61 infall–winter. During the first few days, the spring–summer drifters stay close to the coast and show a highalongshore spreading, with some realizations reachingthe tip of the Gargano Cape after 2 days and approach-ing the DART southern border after 4 days. After 8days, the PDF covers the whole coastal region, with58% of the drifters in the DART area. Of the remain-ing drifters, 33% are dead and 9% have left the area tothe southeast, following the WAC. In contrast, the fall–winter drifters are less confined to the coast and they

FIG. 9. As in Fig. 5, but relative to the total drifters from GARGS (their initial number is 104).


have a higher probability of escaping the WAC northand at the tip of the Gargano Cape. After 8 days, thepercentage of drifters inside the DART area is approxi-mately the same as that in spring–summer (59%), butthe mortality is reduced (27%) and the remaining drift-ers (14%) have left the area mostly toward the interior.Also, during the first few days, the fall–winter PDFsshow a reduced spreading with respect to the spring–summer period, with drifters reaching the tip of thecape after 4 days instead of 2. This result is somewhatsurprising considering that the mean flow is generallyweaker in the spring–summer period than in fall–winter, but we attribute it to the increased flow vari-ability of the spring–summer seasons that is not re-solved by the drifter data coverage. During summer,the WAC is reduced because of the weak Po River

discharge, and it is therefore more directly influencedby the wind. Visual inspection of individual trajectoriesin late spring and summer shows the existence of twodistinct regimes. Some trajectories, probably under theeffect of strong bora winds, are extremely fast, movedown the boundary current, and cover the whole areain a few days. These fast trajectories are responsible forthe fast spreading of the spring–summer PDFs. Othertrajectories, instead, are trapped in mesoscale struc-tures and stay in the boundary current for much longer.Conditional statistics as a function of wind conditionsduring summer would help investigate this point fur-ther, but unfortunately the sampling is insufficient atthis time to conduct such investigation. Another aspectthat further complicates the statistical description insummer is that many “slow” drifters die, probably

FIG. 10. As in Fig. 5, but relative to the total drifters from RECIRC (their initial number is 23).


beached at the coast. In particular, 42% of the spring–summer drifters for POLYN die during the first 21 dayswithout reaching the Gargano Cape, versus the 28% forthe fall–winter period (Table 3). The spring–summerPDFs are then likely to be biased by the fast summerdrifters, and therefore only partially reliable. This as-pect will be considered when discussing the bulk statis-tics in the following.

As in section 4a, for the whole dataset (Table 2) wesummarize the seasonal drifter behavior in POLYNand POLYS by considering bulk quantities, which aredisplayed in Table 4. Because of the reduced sampling,some of the statistics cannot be considered significant.For instance, the percentage of drifters entering therecirculation in the lee of the cape is not reliable be-cause the actual number of drifters ngarg2rec is too small.Furthermore, during the summer, only very few driftersreach the tip of the cape in the first 21 days (ngarg � 4

for POLYN and 8 for POLYS), making it impossible tocompute significant statistics. For the other three sea-sons—fall, winter, and spring—the average arrival timeat the cape Tm (computed over more than 10 realiza-tions) is quite similar, ranging between 9–10 days forPOLYN and 5–6 days for POLYS. A significant sea-sonal dependence can be seen, instead, in the percent-age of drifters exiting the DART area toward the inte-rior after reaching the Gargano Cape (ngarg2int). Duringfall and winter, ngarg2int ranges between 10% and 20%,while during spring ngarg2int � 0. This is consistent withthe PDFs in Fig. 13.

In summary, particles with ICs in the WAC upstreamof the cape have a higher probability of leaving theboundary current toward the interior in fall–winterthan in spring–summer. During the summer, when theWAC is narrow and more confined to the coast, thetrajectories show a very high variability with two dif-

FIG. 11. Mean flow obtained by averaging the drifters velocities over 0.1° square bins for the four seasonalperiods. Gray vectors are for bins with more than 3 but less than 10 independent measurements; black vectors arefor bins with more than 10 independent measurements.


ferent regimes, probably linked to wind forcing. Epi-sodes of very fast trajectories alternate with episodes inwhich drifters are slow and trapped in mesoscalecoastal structures.

5. Conclusions

In this paper, an analysis of historical drifter data ispresented in the area of the DART experiment, in the

FIG. 12. (left) Forward and (right) backward trajectories relative to the POLYN subregion,during the four different seasons. The time intervals tforw and tback are equal to 10 days.


middle Adriatic around the Gargano Cape. The prop-erties of drifter evolution are investigated as a functionof the drifters’ initial conditions, and the specific goal ofproviding suggestions for the DART Lagrangian sam-

pling strategy is considered. The details of the samplingplan will obviously depend on several factors, such aslogistics, ship availability, and interaction with the otherexperimental and modeling activities. Nevertheless,

FIG. 13. As in Fig. 5, but relative to the total drifters from POLYN (Bcint � BCext), during twoextended seasonal periods: the (left) fall–winter and (right) spring–summer periods.


some general suggestions can be provided through ourresults. The DART cruise planned for March 2006 willinclude two successive finescale surveys of 2–3 days,separated by about 4 days. These surveys will mostlycover the POLYS and GARGS regions. One of thegoals is to maximize the overlap between the ship-based measurements and the drifter data, both in spaceand time. In other words, it would be ideal to releasedrifters so that most of them remain in the survey areawhile the ship is sampling. If a first portion of the drift-ers (�eight units) are deployed upstream along the firsttransect (in the western part of POLYS), they will typi-cally drift by the tip of the cape after 5–6 days andfurther stay in the survey area for 5–10 days (Fig. 9).We suggest spreading the deployments across the entirewidth of the first transect with some drifters inshore ofthe 100-m isobath and others more offshore. As shownby Figs. 5 and 6, BCint and BCext drifters spread down-stream similarly; therefore, we expect that the drifterswill cover the DART study area well. The probabilitythat a drifter escapes to the northwest or to the north isrelatively negligible and there is no high risk in releas-ing them along the first transect. We suggest seeding asecond group of drifters at the same location along the

first transect at the beginning of the second survey. Alag of about 4 days with respect to the first survey,combined with a typical residence time in the DARTarea of approximately 10 days, should produce a broadsampling of the study region in quasi synopticity withthe ship survey. Furthermore, some additional launchescan be planned to investigate specific regions, chosen incollaboration with the modeling component of DART.For example, a few drifters (two–three units) could beseeded in the neighborhood of the hyperbolic point inGARGN (Fig. 8) to investigate its synoptic locationduring the experiment, because it is expected to greatlyinfluence the pattern of transport and detrainmentfrom the boundary current. In addition, one–two drift-ers could be deployed in the Gulf of Manfredonia tosample its characteristic recirculation, because very fewdrifters are expected to move into that region (�7%).

Summarizing, the results indicate that the analysis ofhistorical data can provide very valuable informationon statistical particle prediction. A number of applica-tions can be foreseen aside from the discussed experi-ment planning, which include environmental planningand risk mitigation. The methodology should, however,be further improved in the future in order to make itapplicable to more general dynamical problems and inorder to address some of the open points identified inthis paper.

The method is expected to be particularly well suitedfor coastal flows with characteristics similar to the onesconsidered here, that is, with a well-defined current sys-tem and a significant topographic control. For flowswith higher variability and less constraints, such asopen-ocean flows or flows with a strong tidal or meso-scale component related to eddies and waves, it can beenvisioned that historical data will have to be aug-mented by synthetic model trajectories. In addition toproviding unlimited Lagrangian datasets for analysis,model studies will also help investigate the existence offlow regimes and the associated behavior of particlespreading. This is expected to be a valuable approachalso in our DART area to improve our understandingof periods of enhanced variability, such as the observedsummer variability. Because such variability is likelyrelated to wind forcing, which is especially effectiveduring summer when the buoyancy forcing from the PoRiver discharge is weak, conditional statistics with dif-ferent wind forcing could be carried out with synthetictrajectories from an appropriate hydrodynamicalmodel. A modeling study component would also helpinvestigate the characteristics of the complex hyper-bolic point region on the Palagruza Sill.

Another improvement to the proposed methodologythat is possible with broader data coverage is the use of

TABLE 3. As in Table 1, but relative to the four seasons. Noticethat a percentage of the ndie_earlier drifters are not truly dead, buthave just switched from one season to the other.

ntot ndie_earlier

POLYNFall 32 3Winter 29 14Spring 18 3Summer 20 13

POLYSFall 39 4Winter 36 9Spring 21 3Summer 17 8

TABLE 4. As in Table 2, but relative to the four seasons.

ngarg ngarg2int ngarg2rec Tm (days)

POLYNFall 25 (86.2%) 6 (20.7%) 2 (6.9%) 9.0Winter 13 (86.7%) 2 (13.3%) 0 (0.0%) 10.9Spring 14 (93.3%) 0 (0.0%) 1 (6.7%) 9.9Summer 4 (–) 0 (–) 0 (–) —

POLYSFall 32 (91.4%) 8 (22.8%) 2 (5.7%) 5.7Winter 24 (88.9%) 3 (11.1%) 2 (7.4%) 6.3Spring 16 (88.9%) 0 (0.0%) 2 (11.1%) 5.7Summer 8 (–) 1 (–) 0 (–) —


more uniformly distributed initial conditions fromwhich to compute the PDF maps. This could make thePDF statistics more general because they would repre-sent the probability of finding a particle in a cer-tain location given its initial condition in a wide spatialarea rather than in a concentrated location. We havechosen our ICs as entrance points to the consideredsubregions mainly for practical reasons. In fact, ourLagrangian data coverage was not wide enough to pro-vide uniformly spaced initial conditions and, at thesame time, ensure the statistical robustness of the PDFmaps.

Finally, an aspect to be considered in future investi-gations is the effect of the drifter mortality on the com-puted statistics. In our case study, for example, the highmortality observed during the summer is likely to berelated to the beaching of the more coastal and slowerdrifters, which tend to bias the statistics toward thefastest trajectories. We plan to address this issue in afuture work using synthetic Lagrangian stochasticmodel trajectories (R. Barbanti and P.-M. Poulain 2006,unpublished manuscript).

Acknowledgments. The authors are thankful to A.Molcard, T. Ozgokmen, A. Haza, and A. Poje for nu-merous stimulating discussions and to M. Rixen and E.Coelho for organizing the DART experiment and mo-tivating the work. This research was supported by theOffice of Naval Research Grants N00014-05-1-9994 andN00014-03-1-0291 and by the Consiglio Nazionale delleRicerche (Italy). Thanks to all the individuals who havecontributed to the drifter dataset used for this study,and to L. Ursella and R. Barbanti for their help with thedrifter data processing.

REFERENCES

Aref, H., 1984: Stirring by chaotic advection. J. Fluid Mech., 143,1–21.

Artegiani, A., D. Bregant, E. Paschini, N. Pinardi, F. Raicich, andA. Russo, 1997: The Adriatic Sea general circulation. Part I:Air–sea interactions and water mass structure. J. Phys.Oceanogr., 27, 1492–1514.

Davis, R. E., 1985: Drifter observation of coastal currents duringCODE. The method and descriptive view. J. Geophys. Res.,90, 4741–4755.

Doglioli, A. M., A. Griffa, and M. G. Magaldi, 2004: Numericalstudy of a coastal current on a steep slope in presence of acape: The case of the Promontorio di Portofino. J. Geophys.Res., 109, C12033, doi:10.1029/2004JC002422.

Dunlap, E., C. C. L. Tang, and C. K. Wang, 2004: Comparison ofmodel surface currents and drifter data from the GrandBanks. Canadian Hydrography and Ocean Science Tech.

Rep. 236, Department of Fisheries and Oceans Canada,Dartmouth, NS, Canada, 34 pp.

Dutkiewicz, S., A. Griffa, and D. B. Olson, 1993: Particle dif-fusion in a meandering jet. J. Geophys. Res., 98, 16 487–16 500.

Falco, P., A. Griffa, P.-M. Poulain, and E. Zambianchi, 2000:Transport properties in the Adriatic Sea as deduced fromdrifter data. J. Phys. Oceanogr., 30, 2055–2071.

Griffa, A., L. I. Piterbarg, and T. Ozgokmen, 2004: Predicta-bility of Lagrangian particle trajectories: Effects ofsmoothing of the underlying Eulerian flow. J. Mar. Res., 62,1–35.

Hopkins, T. S., A. Artegiani, C. Kinder, and R. Pariante, 1999: Adiscussion of the northern Adriatic circulation and flushing asdetermined from the ELNA hydrography. The Adriatic Sea,Ecosystem Rep. EUR 18834, Vol. 32, European Commission,85–106.

Kuznetsov, L., M. Toner, A. D. Kirwan Jr., C. K. R. T. Jones,L. H. Kantha, and J. Choi, 2002: The Loop Current and ad-jacent rings delineated by Lagrangian analysis of the near-surface flow. J. Mar. Res., 60, 405–429.

Paris, C. B., R. K. Cowen, R. Claro, and K. C. Lindeman, 2005:Larval transport pathways from Cuban snapper (Lutjanidae)spawning aggregations based on biophysical modeling. Mar.Ecol. Prog. Ser., 296, 93–106.

Poje, A. C., and G. Haller, 1999: Geometry of cross-stream mixingin a double-gyre ocean model. J. Phys. Oceanogr., 29, 1649–1665.

Poulain, P.-M., 1999: Drifter observations of surface circulation inthe Adriatic Sea between December 1994 and March 1996. J.Mar. Syst., 20, 231–253.

——, 2001: Adriatic Sea surface circulation as derived from drifterdata between 1990 and 1999. J. Mar. Syst., 29, 3–32.

——, R. Barbanti, R. Cecco, C. Fayos, E. Mauri, L. Ursella, andP. Zanasca, cited 2004a: Mediterranean surface drifter data-base: 2 June 1986 to 11 November 1999. OGS Tech. Rep.78/2004/OGA/31. [Available online at http://poseidon.ogs.trieste.it/drifter/database_med.]

——, E. Mauri, and L. Ursella, 2004b: Unusual upwelling eventand current reversal off the Italian Adriatic coast in summer2003. Geophys. Res. Lett., 31, L05303, doi:10.1029/2003GL019121.

——, R. Barbanti, L. Ursella, and R. Cecco, cited 2006: Adriaticsurface drifter database: 21 September 2002 to 29 February2004. OGS Tech. Rep. [Available online at http://poseidon.ogs.trieste.it/doga/sire/dolcevita/database_dolcevita/.]

Samelson, R. M., 1996: Chaotic transport by mesoscale mo-tions. Stochastic Modeling in Physical Oceanography,R. J. Aller, P. Miller, and B. Rozovskii, Eds., Birkhauser,423–438.

Spaulding, M. L., T. Isaji, A. Allen, P. Hall, and E. Howelett,2005: Application of a random flight model in search andrescue model systems. Third Lagrangian Analysis and Pre-diction of Coastal and Ocean Dynamics (LAPCOD) Meeting,Lerici, Italy, ONR and Cosponsors. [Available online athttp://www.rsmas.miami.edu/LAPCOD/2005-Lerici/.]

Thompson, K. R., J. Sheng, P. C. Smith, and I. Cong, 2003: Pre-


diction of surface currents and drifter trajectories on the in-ner Scotian Shelf. J. Geophys. Res., 108, 3287, doi:10.1029/2001JC001119.

Thomson, D. J., 1987: Criteria for the selection of stochastic mod-els of particle trajectories in turbulent flows. J. Fluid Mech.,210, 113–153.

Ursella, L., P.-M. Poulain, and R. Signell, 2006: Surface drifterderived circulation in the northern and middle Adriatic Sea:

Response to wind regime and season. J. Geophys. Res., 111,C03S04, doi: 10.1029/2005JC003177.

Wiggins, S., 1992: Chaotic Transport in Dynamical Systems.Springer, 301 pp.

Zore-Armanda, M., 1956: On gradient currents in the AdriaticSea. Acta Adriat., 8 (6), 1–38.

——, 1969: Water exchange between the Adriatic and the EasternMediterranean. Deep-Sea Res., 16, 171–178.


historical drifter data and statistical prediction of...

Documents