Estimation of yellowfin tuna (Thunnus albacares) habitat in waters adjacent to Australia’s East Coast: making the most of commercial catch data

Estimation of yellowfin tuna (Thunnus albacares) habitat inwaters adjacent to Australia’s East Coast: making the most ofcommercial catch data

JAMES DELL,1,2,3,* CHRIS WILCOX1,2 ANDALISTAIR J. HOBDAY2

1UTas ⁄ CSIRO Quantitative Marine Science Program,Institute of Marine and Antarctic Studies, University of

Tasmania, Hobart, Tas. 7005, Australia2Wealth from Oceans National Research Flagship, CSIRO,Marine and Atmospheric Research, Hobart, Tas. 7000,

Australia3School of Zoology, University of Tasmania, Hobart,Tas. 7005, Australia

ABSTRACT

The physical environment directly influences the dis-tribution, abundance, physiology and phenology ofmarine species. Relating species presence to physicalocean characteristics to determine habitat associationsis fundamental to the management of marine species.However, direct observation of highly mobile animalsin the open ocean, such as tunas and billfish, is chal-lenging and expensive. As a result, detailed data onhabitat preferences using electronic tags have onlybeen collected for the large iconic, valuable orendangered species. An alternative is to use commer-cial fishery catch data matched with historical oceandata to infer habitat associations. Using catch infor-mation from an Australian longline fishery andBayesian hierarchical models, we investigate theinfluence of environmental variables on the catchdistribution of yellowfin tuna (Thunnus albacares). Thefocus was to understand the relative importance ofspace, time and ocean conditions on the catch of thispelagic predator. We found that pelagic regions withelevated eddy kinetic energy, a shallow surface mixedlayer and relatively high concentrations of chlorophyll aare all associated with high yellowfin tuna catch in theTasman Sea. The time and space information incor-porated in the analysis, while important, were lessinformative than oceanic variables in explaining

catch. An inspection of model prediction errorsidentified clumping of errors at margins of ocean fea-tures, such as eddies and frontal features, which indi-cate that these models could be improved by includingrepresentations of dynamic ocean processes whichaffect the catch of yellowfin tuna.

Key words: East Australia Current, fishery catch data,habitat prediction, Thunnus albacares, yellowfin tuna

INTRODUCTION

An understanding of a species distribution and habitatutilization helps inform natural resource management,through identifying the most probable spatio-temporalenvironmental envelope where the species can befound. This information is important for planningharvest strategies (Garcia et al., 2007), ecosystem-based management (King and McFarlane, 2006), spa-tial management (Hobday and Hartmann, 2006;Hobday et al., 2010), by-catch abatement (Brotherset al., 1999; Howell et al., 2008) and catch rate stan-dardization for stock assessments (Bigelow et al., 2002;Bigelow and Maunder, 2007). Predictions from speciesdistribution models are most effective when there aresufficient location data, free from sampling bias,available to parameterize the environmental charac-teristics that influence the biology, range and habit ofthe species in question (Pearson et al., 2006; Pearmanet al., 2008). However, there are few examples in thepelagic environment where this is the case. Mostspecies location data are fisheries-dependent but, whilebiased by vessel, economic, gear and behaviouraleffects related to fishing operations, these data are themost abundant and widely sampled available. Here weuse commercial fishing data to investigate whichenvironmental variables best characterize the habitatdistributions of yellowfin tuna (Thunnus albacares) inareas where fisheries-independent habitat informationon this species is limited.

Much of the information on habitat utilization forhighly mobile pelagic species, such as tunas and bill-

*Correspondence. e-mail: [email protected]

Received 29 October 2010

Revised version accepted 18 May 2011

FISHERIES OCEANOGRAPHY Fish. Oceanogr. 20:5, 383–396, 2011

� 2011 Blackwell Publishing Ltd. doi:10.1111/j.1365-2419.2011.00591.x 383

fish, was first gathered using fishery catch records orresearch voyages (Alverson, 1959; Sund et al., 1981).More recently, direct observation in the open oceanusing real-time acoustic tag technology and vessel-mounted acoustic receivers or networks of acousticlistening stations attached to fisheries aggregationdevices (FAD) have been used to study habitat asso-ciations (Holland et al., 1985; Josse et al., 1998; Brillet al., 1999). Development of archival tags andincreased focus on data recording in fisheries nowprovide detailed information that can be used toanalyse habitat associations of pelagic predators in themarine environment (Gunn and Block, 2001). Anincreasing number of marine species have been fittedwith electronic tags that measure aspects of theenvironment directly experienced by the animal overlong time periods and potentially large spatial domains(Nielsen et al., 2009). The animals most often taggedin the pelagic zone are those that support the mostvaluable fisheries or are ‘charismatic’ or endangeredspecies, meaning these detailed data on habitat pref-erences are not available for all species in all regions.

In contrast, commercial longline fishing operationsoccur widely across the open ocean and provide analternative source of data for those less-studied species.Capture information from commercial fishing opera-tions has become increasingly detailed and consistentlycollected following increasing pressure from nationaland international agreements calling for greateraccountability and sustainable management of marineresources (e.g. SOFIA – http://www.fao.org/fishery/so-fia/en and the Kobe process tabled at the RegionalFisheries Management Organisations 2007 meeting –http://www.tuna-org.org/meetingspast.htm). Althoughcaution is required when using fishing data to modelspecies distributions (Bishop, 2006), fishing times,location and catch composition of longline operationscan be linked to remotely sensed and modelled oceandata, providing the opportunity to better understandthe habitat association of pelagic species.

Commercial fisheries data are often consideredbiased with respect to mapping species distributionsbecause fishers choose to target species at the centre oftheir distributions to maximize catches and minimizesearch costs. Australia’s Eastern Tuna and BillfishFishery (ETBF) maintains a data set of catch recordscontaining data fields that can be used to minimizethese biases. Multiple species, each with a preferredhabitat, are caught in the ETBF. The fishery is definedas a mixed-species fishery (Caton and McLoughlin,2005; Hobday et al., 2009). The boundaries of theETBF represent an area that combines the preferredspatial and temporal envelopes of many species and, as

a result, extends beyond the preferences of any singlespecies. For example, southern bluefin tuna (Thunnusmaccoyii) (SBT), broadbill swordfish (Xiphias gladius),big-eye tuna (Thunnus obesus) and albacore (Thunnusalalunga) can all function for extended periods attemperatures and depths well below the physiologicallimits of species such as yellowfin tuna (YFT) andstriped marlin (Tetrapturus audax) (Brill, 1996; Brilland Lutcavage, 2001). Longline configuration,deployment and the length of time the hooks are inthe water will alter the way different species arecaught. It follows that the ETBF longline catch dataare a resource that can be used, with caution, to obtaina clearer understanding of the southern distributionallimits and habitat preferences of tropical pelagic pre-dators in the Tasman Sea.

The regional oceanography of the eastern Austra-lian fishing zone targeted by the ETBF is both complexand dynamic. The major large-scale oceanographicfeatures are the East Australian Current (EAC) andthe Tasman Front (Cresswell, 2001). Seasonal anddecadal periodicities in the mass transport of Coral Seawater control the penetration of the southward arm ofthe EAC, which brings warmer water from the CoralSea south to meet the cooler water of the Tasman Sea.This southern extension of tropical waters into higherlatitudes, promotes an expansion of potential habitatfor warm water species (Ling et al., 2009; Figueira andBooth, 2010). The location of the Tasman Front is atthe interface between these two water masses and ischaracterized by strong gradients of temperature,nutrients and productivity (Baird et al., 2008). Thethermal gradients also give rise to strong along-frontvelocities. Regions with strong gradients in theseocean characteristics can shape the realized nicheboundaries of pelagic species, including YFT. Thus thesouthern arm of the EAC, the Tasman Front andsurrounding water masses are suitable to investigatethe relationship between YFT catch and ocean char-acteristics. Our focus is on the biological and physicalsurface characteristics that constitute some of thecomplexity of these major oceanographic features(Baird et al., 2008). We use existing ocean dataproducts which are observable at the spatial andtemporal scales most relevant to YFT.

This is a region where a commercial fishery (ETBF)expends the greatest effort and records high qualitydata, and data on the surface characteristics of theoceanography are also readily available. We use aquantitative modelling framework, utilizing machinelearning and Bayesian generalized linear modelling tomodel these data in relation to YFT catch. Our aimswere to better understand YFT habitat preference in

384 J. Dell et al.

� 2011 Blackwell Publishing Ltd, Fish. Oceanogr., 20:5, 383–396.

the Tasman Sea and to identify the relative impor-tance of location, seasonality and environmentalvariables that underlie these preferences.

METHODS

Fisheries data

We used commercial ETBF longline vessel catch datacollected by the Australian Fisheries ManagementAuthority (AFMA). Catch data covered thearea 22–45�S, 144–173�E in the years 1999–2005.Although catch data were available across the entirearea of the ETBF and earlier than 1999, a large pro-portion of yellowfin tuna catches occur below 20�Sand data records prior to 1999 were often incomplete.The complex oceanography of the area includes thesouthern edge of the East Australian Current andthe Tasman front. This region is where the majority ofthe Australian commercial longline effort currentlyoccurs (Campbell, 2008).

At a minimum, the catch data comprise date, timeand location of set, weather and ocean observations,basic gear structure and composition of catch. Thelogbook data were screened to remove records thatwere clearly in error or insufficient. As a minimumrequirement, only those records that included date,location of start of longline operation, number ofhooks set and a list of the species caught, wereincluded in analyses. The location of each observationwas included as the latitude at the beginning of eachlongline set. Year and month were represented aspositive whole numbers, as they occur in the calendar.Season was included in two alternate forms, as detailedin the model selection section.

Ocean data

Data describing ocean conditions at the start locationof all longline operations were either drawn fromexisting satellite data sets (primary), derived from theprimary satellite data, or derived from modelled out-put. The primary data were sea surface temperature(SST) data (4 km · 3-day composite) measured bythe NOAA ⁄ AVHRR satellite, the sea surface height(SSH) data (0.33� · 3-day composite) from theTOPEX ⁄ Poseidon and ERS-2 altimeters and chloro-phyll a data (4 km · 7-day composite, units mg m)3)from sea-viewing wide field-of-view sensor (SeaWiFS).Modeled data were derived from the CSIRO Atlas ofRegional Seas (CARS, Condie and Dunn, 2006,http://www.marine.csiro.au/~dunn/cars2006/) and theBluelink oceanographic model (Hobday et al., 2004;Oke et al., 2008) and accessed using purpose-built

MATLAB functions (Hobday et al., 2006). The deriveddata products included CARS mixed layer depth(MLD; defined as the depth where there is a minimum0.2�C change in temperature and 0.03 in salinity,starting from 10 m below the surface), current speedand direction, and sea surface height anomaly. Weused derived products based on both the modelled andthe measured products in an attempt to summarize thecharacteristics of mesoscale ocean features includingeddy kinetic energy (EKE) and frontal density (algo-rithm as defined in Hobday et al., 2006; followingCayula and Cornillon, 1992). [EKE was calculatedusing the following formula: EKE = 0.5 · (u2 + v2),where u is the East–West component of the velocity ofthe surface ocean and v is the North–South compo-nent. The vertical velocity component, z, is ignored inthis simplified formulation of EKE as its contribution isminor. Units are in cm2 s)1]. The influence ofbathymetry on YFT catch was considered in initialinvestigations of the data but it was found to be of lowimportance and was excluded from subsequent analy-sis. We also investigated a simple binary index tosummarize the location of EAC water. The index wasbased on modelled temperature at depth and currentspeed and direction. On review, it had a very lowpredictive power for YFT catch or vessel locations andwas not considered further. Maury et al. (2001) suggestthat salinity and dissolved oxygen are useful fordetermining YFT concentrations at temporal andspatial scales associated with ocean basins; however,the range of these variables are not physiologicallylimiting to tunas (Brill, 1994) in the system we studied(Condie and Dunn, 2006) and were not consideredfurther.

Variable importance using random forest algorithm

As a preliminary exploration of the available covari-ates we used a recursive partitioning product, calledRandom Forest, available in the package PARTY

(Hothorn et al., 2006) in the statistical language R

(R-Development-Core-Team, 2007). The aim was torobustly limit the number of variables to the mostimportant variables in order to reduce model com-plexity. This statistical approach to regression treeanalysis is based on significance tests to partition data(Hothorn et al., 2006; Strobl et al., 2007). Bootstrap-ping of data allows for the random selection of cova-riates and production of multiple trees (‘a forest’), fromwhich a value of variable importance is calculatedbased on the number of times a variable appears in thebootstrapped models. Covariates identified as impor-tant were assessed for correlation based on Pearson’scorrelation using all of the available and complete

Yellowfin habitat derived from catch data 385


records. Highly correlated covariates were excludedfrom the general linear modelling phase to preventco-linearity and bias (Venables and Dichmont, 2004;Maunder et al., 2006).

Hierarchical modelling using Bayesian mixture models

To model YFT catch, a Bayesian hierarchical gener-alized linear model framework was used. Our aim wasto find associations between YFT catch and thecharacteristics of the surface ocean, rather thandetermining quantitative inputs for catch standardi-zation. For this reason GLMs were chosen over GAMs,as they can provide clearer relationships betweenvariables and catch data. The catches of YFT weremodelled directly as a discrete count, incorporating anoffset term in the statistical model to accommodatenominal effort. We considered a range of competingmodels under this framework, where each modelincorporated different combinations of seasonal, spa-tial and oceanographic covariates. Due to computa-tional limitations, models were fit to a set of 5000randomly selected longline records. A further 5000longline sets were randomly selected and set aside formodel validation and testing. Many of the longlineoperations included in the analysis contained zero YFTcatch records. To account for the large number of zerocatches, we modelled the data using a Poisson bino-mial mixture model, often called a zero-inflated Pois-son (ZIP) (Lambert, 1992). We modelled the catch offish in a location as a Poisson random variable, con-ditional on a Bernoulli random variable. The outcomeof the Bernoulli trial either determines that zero fishare caught (1–pi), or the quantity of fish caught isdrawn from a Poisson process (pi). The ZIP calculatedthe conditional probability of catching c fish on the ithlongline set, given the presence of fish (ci) at thatlocation as:

PrðC¼ cijpi �bÞ ¼PoisðkiÞ;with probabilityðpiÞ

0;with probabilityð1� piÞð1Þ

given

InðkiÞ ¼ bYi ð2Þwhere ci is the rate of the Poisson influenced by Yi, thevector of the oceanographic, temporal and spatialcovariates, and b, the complementary vector ofcoefficients.

Estimation of model variance and errors

Estimation of the parameters in the model wasachieved via Markov Chain Monte Carlo (MCMC)sampling using WINBUGS 1.4 (Spiegelhalter et al.,2002). The convergence of the Markov chains wasassessed using the CODA package in R, with the diag-

nostics described in Cowles and Carlin (1996) andMengersen et al. (1999). Based on CODA output forpreliminary model runs, we ran 35 000 iterations anddiscarded the first 10 000 iterations as the burn-inphase. To ensure convergence, every fifth MCMCsample was retained, resulting in chains of 5000 iter-ations from which we formed the posterior distributionfor each coefficient.

Model structure and selection

We explored two options for the structure of theseasonal variable included in the rate of the Poisson.Season was represented as four positive numbers(1–4), where 1 summarized the austral summer months(December through February), 2 = autumn (Marchthrough May), 3 = winter (June through August), and4 = spring (September through November). WINBUGS

allows season to be coded as either a three-elementcorner contrast factor or a two-element cyclic functionwhich assumes a cyclic nature for the seasonal signal.The corner contrast approach identifies one of theannual seasons as the mean state (summer) and con-trasts the remaining seasons, each defined by threebinary elements and matching coefficients, to uniquelyspecify autumn, winter and spring values. The cyclicseasonal variable was defined by a trigonometricfunction with two coefficients (3);

bcos � cosð2p� season

4Þ þ bsin � sinð2p� season

4Þ

� �ð3Þ

All oceanographic and spatial variables were stan-dardized to have a zero mean and unit variance toimprove MCMC performance and to allow directcomparisons of the relative influence each covari-ate ⁄ coefficient pair had on the modelled response.

Model fit was investigated using two approaches.First, based on the reported deviance and magnitude ofthe Bayesian information criterion (BIC, eqn 4) ameasure of model fit which is directly related to boththe negative log likelihood and the model deviance,whereby the likelihood (L) is corrected for parsimony(k = number of parameters) and sample size (N)(Schwarz, 1978). A lower BIC indicated a betterdegree of model fit.

BIC ¼ �2 � ln Lþ k � lnðNÞ ð4Þ

We identified the set of best models by looking atthose models that had BIC values overlapping the 2.5–97.5% quantiles of the best model. The BIC is bestinterpreted as a relative measure, where competing

386 J. Dell et al.


models are measured against the leading model. Sec-ondly, we evaluated the ability of all models to predictthe response variable. To do this, we randomly sam-pled parameter values for each model from the pos-terior predictions of parameter coefficients andpredicted the expected catch for 5000 randomlyselected longline operations distinct from those usedto parameterize the models. We compared catch esti-mates produced with the observed catches via themean of the squared deviations between predicted andobserved catches. We repeated this process using 1000randomly selected parameter sets from the Markovchain from each model. The mean-squared predictionerror of each model, referred to here as the predictiondiagnostic (eqn 5), was then used to produce a secondranking of the suite of models.

PD ¼

P5000

i¼1

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiP1000

j¼1

ðobs¼predÞ2

1000

s

5000ð5Þ

where i is of the set of 5000 randomly selected longlineoperations and j of the 1000 randomly selectedparameter sets from the Markov chain. This processwas repeated using multiple random samples from thecatch data to determine a median value for the pre-diction diagnostic with 95% confidence intervals. Theunits of the prediction diagnostic are number of fish,given that it is representing the mean number of fishby which the model prediction differs from theobserved catches.

Residual analysis

To investigate the possibility of including interactionsand higher order terms in the models we took the bestmodel, based on lowest BIC and the prediction diag-nostic, and plotted ranked-standardized model residu-als against values of potential covariates. We exploredall two-way interactions and higher order terms for allmain effects. Where obvious relationships betweencovariates and models were evident, extra terms wereadded to the best model. The fitting and modelselection process was repeated until no new covariateswere identified.

RESULTS

Distribution of longline fishing effort

The complete data set was a collection of 59 424individual longline fishing operations undertaken inthe ETBF between 1 January 1999 and 31 December2004. The most commonly caught species, YFT, rep-

resented over 30% of the total fish caught, withalbacore and big-eye tuna being the next most caughtspecies, each representing between 10 and 15% of thetotal catch. There were 11 875 instances where zeroYFT were caught; this represents �20% of the data.Conversely, �80% of longline operations caught atleast one YFT. Large catches of YFT, where morethan 30 fish per operation were caught, occurred in<3% of the data (1770 operations). The longlineeffort was concentrated along the Australian eastcoast with an eastward extension of activity occurringbetween 25� and 35�S (Fig. 1). There were concen-trations of fishing activity within this eastwardextension.

Preliminary variable selection

The random forest bootstrap analysis ranked thepotential covariates based on a scale of relativeimportance (Fig. 2) and identified season, chlorophylla concentration (SeaWiFS), sea surface temperature(SST), mixed layer depth (MLD), month, latitude,year and eddy kinetic energy (EKE) as the eight mostimportant variables. Additional variables (sea surfaceheight anomaly, moon-phase, frontal density, currentdirection and bathymetry) had lower relative impor-tance.

We based our decision on how to represent tem-poral information within the zero-inflated model on

−45

−40

−35

−30

−25

−20

−15

−10

Latit

ude

140 145 150 155 160 165 170 175 180Longitude

0

100

200

300

400

500

Figure 1. Total nominal effort (longline sets) for the Aus-tralian eastern tuna and billfish fleet from 1999 to 2004. Rawdata are aggregated by quarter-degree squares, legend scalerelates to the total number of longline sets recorded withineach square. Scale is capped at 500. The box defines the areaof focus in this study.



the performance of the temporal variables, month,moon phase, season and year, in the random forestanalysis. Season was retained in preference to monthand moon phase. Season was the temporal variablewhich appeared to best explain the intra-annualvariability in the data and was ranked higher thanboth month and moon phase on a scale of relativeimportance. We included a year variable to modelinter-annual differences.

The correlation analysis of the variables identifiedby the RF bootstrap analysis subsequently demon-strated that SST was strongly correlated with latitude

(r = 0.503), season (r = )0.497) and SeaWiFS(r = )0.598) (Table 1). To avoid co-linearity we ini-tially removed SST, but replaced it following theresidual analysis of the best models determined bymodel selection. The median and range over whichthe important ocean variables were measured are listedin Table 2.

Model selection

The leading model (A, Table 3) was 17 BIC pointslower than the next best (B, Table 3) and 2363 BICpoints lower than a model with only a mean (model T,Table 3). All models shared the hierarchical zero-inflated Poisson structure, differing only in the cova-riates governing the Poisson rate parameter. Thesecovariates can be grouped into four classes based onthe common variables that were included: mean,temporal, spatial and ocean variables. All combina-tions of these covariate sets were investigated and thebest are presented here. All models that containedvariables drawn from the temporal (e.g. season), spa-tial (e.g. latitude) or ocean variable sets (e.g. EKE,SeaWiFS, MLD and ⁄ or SST) had a lower BIC valuethan the model where the Poisson rate parameterwas informed only by the mean variable. Modelsincorporating only spatial and temporal variables hadhigher BIC values than those models that also

Var

iab

le im

po

rtan

ce -

z-s

core

0

5

10

15

20

25

Year

Month Lat

Seaso

nMoon

SST

CurrDir

Front

EKE

SeaW

iFS

MLDSSHa

Bathym

Figure 2. Importance of variables as determined by therandom forest algorithm. Columns represent the meanimportance z-score calculated from 1000 tree models.Additional error bars represent the 95% interval of thez-score. Z-score is a standard error statistic calculated fromthe variable importance and the standard deviation of theimportance, with a correction factor for the number of treesin the Random Forest.

Table 1. Pearson’s correlation table of important variables in the study region. Correlation scores (r) are listed in the lowerdiagonal and probability values (P) are shown in the upper diagonal.

YFTcnt Hooks Year Season Lat SeaWiFS EKE MLD SST

YFTcnt 1 0 0 0 0 0 0 0.084 0Hooks 0.08 1 0 0.836 0.0824 0.0616 0 0.0003 0Year 0.1152 0.0983 1 0.0001 0.0634 0.2811 0 0.0003 0Season 0.1387 0.0029 )0.056 1 0.009 0 0.014 0 0Lat )0.0585 0.0246 0.0263 0.037 1 0 0 0.0013 0SeaWiFS 0.1708 )0.0264 0.0152 0.3547 )0.4985 1 0 0 0EKE 0.104 )0.0873 0.0602 )0.0348 )0.2032 0.0593 1 0 0MLD )0.0244 0.0507 )0.0512 0.3769 0.0454 0.1419 )0.1171 1 0SST )0.0851 )0.0616 0.1138 )0.4973 0.5032 )0.5982 0.1048 )0.3621 1

Table 2. Range and median values of ocean variables sam-pled from the location of longline operations in the studyregion between 1999 and 2004.

SeaWiFS(mg m)3)

SST(�C)

EKE(cm)2 s)2)

MLD(m)

Min 0 9.3 )0.029 5Mean 2.5 22.3 0.853 129Max 4.9 35.2 1.735 253

388 J. Dell et al.


included ocean variables, indicating that ocean vari-ables explained substantial amounts of variation whenmodelling the YFT catch. The five best models allincluded three ocean variables (EKE, SeaWiFS andMLD), a spatial variable (latitude) and two temporalvariables (year, cyclic representation of season). Aftermodel improvement based on residual analysis, thebest model, based on BIC, also contained an addi-tional ocean variable (SST), a higher order spatialterm (latitude2), and an interaction term between twoocean variables (SeaWiFS and EKE). The latitude2

term was the only higher order term that was identifiedin the residual analysis and the improvement in BICindicates that the non-linear representation of latitudebetter fits the pattern of catch of yellowfin tuna in theTasman Sea.

The 2.5 and 97.5% quantiles for posterior distri-bution of BIC estimates for each candidate modeldetermined from the MCMC provided an initialindication of how different the fits were between eachmodel. Models of similar variable structure tended tobe grouped together on the BIC scale and there arethree main groupings of models in this scale. The twomodels with the lowest mean BIC score (models A and B

in Table 3) showed substantial overlap in the distri-bution of their BIC estimates. The next lowest-rankedgroup of models by BIC, models C and D, was sub-stantially different from models A and B. Models Cand D shared a similar range of BIC estimates and asimilar variable structure, only differing by the pres-ence ⁄ absence of the SST term. The next group ofthree models, based on their BIC distributions (modelsE, F and G), were markedly different to models A and B.This third group contained models that differed in theway seasonal information was incorporated. Themodel with a cyclic seasonal structure (model E) hadfewer variables than the other two models in thisgroup, but had the lowest BIC. The level of similarityand difference between the highest-rated models ismore evident when the entire distributions of BICwere considered (Fig. 3).

Model structures that did not incorporate oceanvariables all had higher BIC values than models thatincluded the key ocean variables. The best performingmodel without SeaWiFS, EKE and MLD variablesincluded the ocean variable SST, a spatial term andtemporal information (model M). Removing all oceanvariables further increased the BIC (model N). It was

Table 3. Competing model structures ranked by the Bayesian information criterion (BIC).

Modellabel

Modelstructure dk

2.5%bound BIC

97.5%bound

RescaledBIC

2.5%bound

Predictiondiagnostic

97.5%bound

A ov, c, y, t, Lat2,SeaWiFS*EKE

10 22 447 22 497 22 547 0 10.40 11.55 12.45

B ov, c, y, t, Lat2 9 22 463 22 513 22 563 16 10.37 11.32 12.31C ov, l, c, y, t 9 22 555 22 603 22 651 106 9.77 10.65 11.94D ov, l, c, y 8 22 562 22 610 22 657 113 9.73 10.65 11.95E ov, c, y 8 22 621 22 670 22 718 173 9.92 10.63 11.97F ov, l, f, y, t 10 22 634 22 682 22 730 185 10.61 11.55 12.93G ov, l, f, y 9 22 640 22 688 22 737 191 9.76 10.66 12.01H ov, l, c 7 22 815 22 861 22 906 364 9.93 10.77 12.07I ov, l, f 8 22 899 22 945 22 990 448 10.51 11.52 12.51J ov, l, y, t 7 23 021 23 066 23 111 569 9.98 10.83 12.12K ov, y, t 6 23 028 23 072 23 117 575 9.99 10.83 12.13L ov, l, y 6 23 148 23 192 23 236 695 9.99 11.06 12.28M l, c, y, t 6 23 237 23 281 23 325 784 9.97 10.81 11.96N l, c, y 5 23 314 23 357 23 400 860 9.90 10.88 12.01O l, f, y, t 7 23 332 23 376 23 420 879 10.80 11.77 13.04P l, f, y 6 23 399 23 442 23 486 945 10.77 11.52 12.78Q l, c 4 23 629 23 670 23 710 1173 10.16 10.90 12.11R l, f 5 23 728 23 768 23 809 1272 10.73 11.49 12.75S c, y, t 5 23 737 23 778 23 820 1282 10.08 10.80 11.98T Mean only 1 24 811 24 844 24 876 2347 10.22 10.99 12.19

ov, ocean variables (EKE, MLD, SeaWiFS); c, cyclic seasonal information; f, factor ⁄ corner contrast seasonal information;l, latitude; Lat2, second-order latitude term; y, year; t, SST; SeaWiFS*EKE, interaction term; dk, effective number of parameters;N, number of samples drawn from both the base data and the test data in order to calculate the respective diagnostics; delta BIC,relative difference in BIC between ranked models. The grey section highlights models without EKE, MLD or SeaWiFS.



notable that removing the spatial information andretaining the correlated SST variable led to a markedloss in fit. We observed the general pattern thatexclusion of ocean variables had the greatest effect onthe fit (BIC), followed by temporal information (par-ticularly seasonal information) and subsequently spa-tial information (Table 3).

Models were also compared based on their ability topredict the number of yellowfin tuna caught in 5000random commercial longline operations not used infitting the model (Fig. 4). All models considered,including the ‘mean only’ model, returned values forthe prediction diagnostic within a range of 9.76 and13.04 fish (Table 3). The models with the lowestmean prediction error (Models D and E, 11.15 fish)were among the five highest-ranked models based onBIC. Model D and E covariates were based on oceanconditions and seasonal information. Models withoutocean variables but containing the spatial and seasonalinformation (model N) resulted in a mean predictiondiagnostic of 11.34 fish. Adding SST as a covariate(model M) saw a minor improvement in the meanprediction accuracy (11.32 fish), whereas the additionof higher-order spatial terms and interactions between

ocean covariates improved the model fit but led topoorer predictions. These comparisons indicate thatmodels that utilize ocean variables lose predictivepower (with respect to the prediction diagnostic) buthave improved fit by BIC when SST, higher orderspatial terms and interactions between ocean covari-ates are also present. The mismatch between theprediction diagnostic and the BIC are a result ofmeasuring very different metrics. The BIC is measuredon the likelihood of the model, as it is based on theprobability of the residuals given the fitted parameters.The prediction diagnostic is measured on the scale ofthe observations, as it is the difference between themodel predictions and observations in numbers of fish.

The models that achieved the lowest BIC (model A)and the best prediction diagnostic (models D and E) allshared the same ocean variables (EKE, SeaWiFS andMLD). The coefficients of these ocean variables are ofcomparable size and share the same shape (linear), withMLD have fractionally more influence on the Poissonrate than the other two variables (Table 4). This indi-cates more fish are caught when the MLD is shallow andin an area of high EKE and high primary productivity.The positive linear relationship shows that withincreasing EKE and chlorophyll a concentration(SeaWiFS), an increase in the longline catch of YFT isexpected. The SST term and an interaction term(SeaWiFS*EKE), increased the number of ocean co-efficients in model A by two, but the relative influenceof these additional coefficients was an order of magni-tude smaller (0.02 and )0.04) than the other ocean

22 500 22 600 22 700 22 800 22 900 23 000

0.00

00.

005

0.01

00.

015

BIC

Den

sity

Figure 3. Posterior distributions of BIC for the leadingmodels taken from MCMC. Curves to the left of the plot arethe leading models. Reading the plot left to right comparesmodel A through to model H. Colours are used to identifythe curves (online version): green = model A (ov, c, y, t,Lat2, SeaWiFS*EKE); blue = model B (ov, c, y, t, Lat2);red = model C (ov, l, c, y, t); black = model D (ov, l, c, y, t);purple = model E (ov, c, y); pink = model F (ov, l, f, y, t);orange = model G (ov, l, f, y); brown = model H (ov, l, c).Abbreviations explain covariate structure of models: ov,ocean variables (EKE, MLD SeaWiFS); c, cyclic seasonalinformation; f, factor ⁄ corner contrast seasonal information;l, latitude; Lat2, second order latitude term; y, year; t, SST;SeaWiFS*EKE, interaction term.

YFT caught per longline operation

Fre

quen

cy

0 20 40 60 80 100 120 140

050

010

0015

00

Figure 4. Number of yellowfin tuna (YFT) caught perlongline operation in the ETBF from 5000 randomly sampledoperations used to parameterize the hierarchical models.Vertical line indicates the value of the mean error in themodel prediction when compared to the observed number ofYFT caught, as determined using the prediction diagnostic.

390 J. Dell et al.


variable coefficients in model A (SeaWiFS = 0.17,EKE = 0.13, MLD = )0.17). The squared latitude termis absent in model E (Table 4). Model E has the simplerstructure and summary of the relationship between YFTcatch and the important ocean variables. The evidencesupporting the choice of model E over model A is thatthe addition of the higher-order spatial term, interactiveocean term and the reintroduction of SST (model A)led to less than a 1% improvement in the fit, asdetermined by BIC, but a more than 8% loss in pre-dictive ability, suggesting that model A may be slightlyoverfit. Furthermore, we looked at the overdispersionin the zero-inflated Poisson as overdispersion oftenleads to over-parameterization. We used two post hocmethods to assess the extent of overdispersion. Thefirst approach investigated the pattern of the residualsfrom predictions of catch relative to observed catch(producing qq plots of expected catch versus theranked standardized residuals). Models showed evi-dence of lack of fit for catches close to zero and forvery large catches. The second approach investigatedthe cumulative density function (cdf) of the fittedmodel up to and including the value of each obser-vation, and was uniform between the probabilities of 0and 1. We used the Kolmogorov–Smirnov test todetermine whether the integrated cdf for the proba-bility of each observed catch was significantly differentfrom the uniform distribution. This hypothesis wassupported, suggesting that catch data were overdi-spersed in relation to the zero-inflated Poisson.Investigating the cdf plot further supported the initialassessment that the model underestimated the numberof zero catches and very large catches.

DISCUSSION

The aim of this study was to use commercial catch datato characterize ocean habitat associations of YFT in theTasman Sea. Remotely sensed ocean data and variablesderived from this information (e.g. EKE) were moreinformative than spatial or temporal variables in pre-dicting YFT catch. Specifically, chlorophyll a con-

centration, eddy kinetic energy and mixed layer depthwere responsible for the greatest improvements inmodel fit. Our results suggest that those areas of theTasman Sea with relatively high phytoplankton bio-mass (SeaWiFS) and current shear (EKE) and shallowmixed layer depths (MLD) yield higher catches of YFT.

These properties of the upper surface waters havebeen associated with mechanisms that either concen-trate tuna forage or increase the catchability ofsurface-oriented predators in other ocean regions(Laurs et al., 1984; Polovina et al., 2001; Zainuddinet al., 2006). The shallow MLD may restrict potentialforage and YFT activity to a depth that can be moreeffectively sampled by the arrays of hooks deployed bythe boats in the ETBF (Young et al., 2010). Longlinehooks can reach up 400 m in depth but a majority ofthe hooks are at shallower depths due to the catenarystructure and shoaling of longlines (Bach et al., 2009).Areas of elevated chlorophyll a are often associatedwith increased primary and secondary production.Increased eddy kinetic energy is characteristic of edgesof eddy features and their associated frontal structures.These mesoscale oceanographic features are known tolead to the aggregation of tuna forage by bothentrainment and encouraging vertical mixing, whichintroduces nutrients into the photic zone, promotingphytoplankton growth, and which in turn may attractpredators from higher trophic levels (Sournia, 1994;Bakun, 1996). The negative influence of the interac-tion between EKE and SeaWiFS on model predictionssuggests that the link between YFT catch and theseocean variables is not a simple linear association.Season and year effects were also found to be impor-tant covariates; however, these variables are likely tobe proxies for other unknown processes influencing thelocation of YFT. We found that a cyclic parameteri-zation for season yielded a better fit than usingindependent factors. Given that the cyclic parame-terization is more constrained in its fit to the data, thissuggests that the periodic structure reduces the vari-ability in the estimates of seasonal coefficients, relativeto the factor-based approach, during seasons where the

Table 4. Covariates and estimated coefficient values of the leading hierarchical models, as determined by BIC and the pre-diction diagnostic.

Modellabel

Modelstructure

RescaledBIC

Predictiondiagnostic Mean Year

Seasoneta

Seasonomega Lat SeaWiFS SST EKE MLD Lat2

SeaWiFS*EKE

A ov, c, y, t, Lat2,SeaWiFS*EKE

0 10.31 )5.08 )0.16 )0.30 0.26 0.17 0.02 0.13 )0.17 0.12 )0.04

D ov, l, c, y 117 9.68 )4.61 0.15 )0.28 0.23 )0.08 0.14 0.11 )0.18

See Table 3 for explanation of abbreviations.



variability of catches is large, which in turn helps toimprove overall model fit. In addition, other unknownand un-modelled processes may be contributing to thevariability in the seasonal catch. Possible mechanismsmay be seasonal shifts in prey species, the north–southmovement of the population of YFT within the Tas-man Sea, and recruitment pulses of juveniles to theregion, possibly associated with spawning events in theCoral Sea and other regions of the western and centralPacific ocean (Langley et al., 2009). These sameevents may also be contributing to the overdispersionobserved in relation to the uni-modal zero-inflatedPoisson used in this study. The results of this analysisunderline that there is more to understand about theprocesses responsible for the concentration of YFT insurface waters.

The variable latitude and SST contributed less tomodel fit than would have been expected based onprevious studies that have used these variables tocharacterize the distributions of YFT (Barratt et al.,2002) and other pelagic species (Walsh et al., 2006).This difference may be due to the geographic scale ofour analysis. We limited our analysis to a region withinwhich YFT and other large pelagic fish are targeted, aswe were interested in defining the preferred habitatsand not the factors that determine the complete set ofrange limits for YFT. However, we did draw data fromareas that are considered to be at the southern marginof YFT range (Ward and Bromhead, 2004). Workingwith data at a scale and within a region where SSTand latitude are not the dominant drivers of YFThabitat provided the opportunity to better understandthe association of the other ocean characteristics withcatches of YFT. This approach was chosen over onemore heavily focused on space and time, incorporatedas either spatial blocks or space-time grid effects, asthese approaches risk aliasing the environmentalcovariates included in the analysis (Walsh et al.,2006).

Sea surface temperature was initially excluded fromthe analysis based on the correlations the variableshared with season, latitude and chlorophyll a con-centration (SeaWiFS). Once the best model wasdetermined in the absence of co-linearity, SST wasreintroduced based on the outcomes of the residualanalysis. The addition of the SST covariate did furtherreduce the BIC, suggesting that SST is important toYFT catch or that it aliases other ocean conditions notcaptured in the models presented here. Interestingly,the model structure that excluded SST data (model D)predicted YFT catch better than those models withadditional and more complex covariates. The loss ofpredictive power with improved model fit raises the

possibility that the models with lower BIC values mayhave been overfit, even though BIC does penalizemodels based on numbers of covariates. Model D hadthe fewest variables and lacked higher-order andinteraction terms, and returned the best predictionsand could, therefore, be considered the most parsi-monious model for prediction purposes.

The measures used to rank model performance, theBIC and the prediction diagnostic, the graphicalreview of the model data and the two diagnostic testsindicated that data are over-dispersed with respect tothe model, despite using a zero inflation term. Themain determinants of the overdispersion werethe inability to effectively capture the excess zeros andthe rare events when more than 40 YFT were caughtduring a single longline operation. Although the rarelarge catches represent a small proportion of the totalobservations (<2%), there are sufficient numbers tocompromise the fit of a zero-inflated Poisson, whichdoes not fully account for the level of overdispersionthat is evident. Resolving this overdispersion mayrequire additional information, either as informativepriors or covariates, with explanatory power for theseobservations or, alternatively, the use of a morecomplex multimodal distribution that can accommo-date both zero inflation and the rare and randomextreme observations. Effectively modelling these rarelarge catch events in relation to ocean conditions andthe ETBF longline effort will improve the under-standing of the ecology, management and sustainableharvest of YFT in the Tasman Sea and would be aproductive direction for future work.

Remotely sensed ocean products, used here toquantify mesoscale ocean features (EKE) and primaryproductivity (SeaWiFS), have been used as coarseproxies to represent the unknown mechanisms thataccumulate tuna, and have also been identified in otherocean regions (Mugo et al., 2010; Polovina et al., 2001,Zainuddin et al., 2008). The inclusion of mixed layerdepth data in our best models further emphasizes theimportance of vertical temperature structure in deter-mining the distribution of YFT, a result supported byother models of tuna habitat (Bigelow and Maunder,2007). Statistical modelling of longline catches of YFTin relation to remotely sensed ocean data, and in situoceanographic measurements have occurred in thecentral Atlantic and Indian Oceans (Maury et al.,2001; Zagaglia et al., 2004; Song et al., 2008). Thesestudies found that vertical temperature structure of thesurface ocean explained YFT longline catches,although they included various representations of thesedata: vertical temperature at set depths, the thickness ofthe mixed layer or surface layer, or the structure of the

392 J. Dell et al.


thermocline (Maury et al., 2001; Bertrand et al., 2002;Song et al., 2008). In general, vertical temperaturestructure is widely accepted as the major influence inthe distribution of pelagic fishes in the vertical andhorizontal plane (Sund et al., 1981; Brill and Lutca-vage, 2001; Boyce et al., 2008). YFT is most often foundin the upper mixed layer in tropical and temperate re-gions, spending between 70 and 80% of time in thewarmest water available and rarely visiting water >8�Ccooler than the water at the sea surface (Block et al.,1997; Schaefer et al., 2007). The MLD variable inte-grates more ocean processes than purely vertical tem-perature structure, as it results from a combination ofprocesses that shape the upper ocean (water mass,temperature, wind stress and baroclinic currents).Although we cannot explicitly state how MLD leads toenhanced catch of YFT, we show that it is useful at ageographic scale relevant and important for under-standing YFT capture and spatial management.

The best models presented here used chlorophyll aconcentration as a proxy for the standing stock ofphytoplankton, which may also influence tuna forage.Forage information is considered to be particularlyimportant in driving YFT concentration and activityand chlorophyll a concentration has been used as itsproxy in previous studies (Maury et al., 2001; Lehodeyet al., 2003, 2008). Chlorophyll a concentration waspositively associated with the capture of YFT in theETBF, which is consistent with the observations madein a previous ship-based investigation in the TasmanSea (Young et al., 2001). In the Atlantic, Zagagliaet al. (2004) reported a negative, nonlinear associationbetween YFT capture and chlorophyll a, suggestingthat the decrease in visibility in these areas interfereswith the visual feeding by YFT. Song et al. (2008)determined that relatively low concentrations ofchlorophyll a were optimum for longline capture ofYFT in the Northern Indian Ocean. Conversely,anomalously large purse seine catches in the equatorialWestern Indian Ocean were linked to a high con-centration phytoplankton bloom associated with alarge anticyclonic eddy (Fonteneau et al., 2008).Chlorophyll a is a proxy for phytoplankton, at leasttwo trophic levels below the prey of YFT (Young et al.,2010), and the trophic pathway from phytoplanktonto tuna forage may vary substantially in differentpelagic systems.

Eddy kinetic energy (EKE) explained substantialamount of variation in YFT catch in the Tasman Sea.EKE represents both current strength and a generalmeasure of shear. Although mesoscale features (fronts,eddies and jets) have long been associated with tunaactivity (Fiedler and Bernard, 1987), we could not find

any other catch-based studies on YFT that includedEKE as an explanatory variable. Zagaglia et al. (2004)investigated sea surface height anomaly, a primaryvariable used in determining EKE, but found that ithad relatively weak association with YFT catch in theEquatorial Atlantic.

We would expect that improving the quality ofthe commercial catch data would provide theopportunity to assess more closely the influence ofthe oceanography on YFT catch. Having accurateposition and time information for both the start andcompletion of a longline operation would allow moreaccurate representation of the volume of watersampled, which would reduce the spatial errorsassociated with matching catch data to the hori-zontal and vertical structure of mesoscale oceanfeatures (Ward and Myers, 2005; Bach et al., 2009).The continued improvements in the spatial resolu-tion of modeled ocean data and increased coverageof in situ observations will also help to reduce thesepotential errors.

Overall, environmental variables were shown to begood covariates of YFT catch in the Tasman Sea.Specifically, areas with shallow mixed layer depth,elevated EKE and chlorophyll a concentrations aremore likely to yield higher catches of YFT. Areas ofshallow mixed layer depth may restrict YFT activityinto a smaller volume of water that is likely to increasetheir availability to longline hooks, and areas withelevated EKE and chlorophyll a are likely to be zonesof increased primary and secondary production withinthe Tasman Sea. The fishing strategies of boats withinthe fleet, the configuration of longline fishing gearsand bait selection should all be considered beforeextrapolating predictions of these models beyond thewaters of the Tasman Sea.

In this study we have quantified the preference ofYFT for particular oceanic conditions in the TasmanSea at a scale that is relevant to the regional ocean-ography and day-to-day operations of the commerciallongline fishery. Our approach used rigorous statisticalprocedures that allowed the most informative physicalvariables and parsimonious prediction model to beselected. The modelling framework is general andflexible and further refinements or elaborations can bemade as more data become available. For example,higher resolution in situ ocean data and more com-prehensive catch information, and the possibility ofincorporating expert opinion and belief as priorinformation, within the Bayesian context, would mostlikely further increase model fits to the data. Also, thesame approach, developed using YFT catch, can easilybe applied to the other valuable species caught by the



longline fleet operating in the ETBF. Furthermore, therelationships between YFT catch and spatial, seasonaland ocean covariates explored in this study can beused as correlates for spatially explicit population- orecosystem-based models, where there is great value inimproving the spatial and temporal scale of predatorand prey interactions (Fulton et al., 2004). Finally, themost direct application of the models would be toinform spatial management discussions, with fisheriesmanagers and resource management agencies beingable to identify regions of high potential catch andbetter balance the protection of the resource with thesustainable economic management of the fishery(Hartog et al., 2011) and, where necessary, imple-ment spatial closures or marine-protected areas insituations where the pelagic ecology requires greaterprotection from the impacts of longline fishing(Howell et al., 2008; Griffiths et al., 2010; Trebilcoet al., 2010).

ACKNOWLEDGEMENTS

Catch data for the ETBF were supplied by AFMA.Initial data extraction and processing was assisted byRobert Campbell, Jason Hartog and Scott Cooper.Suggestions from Karen Evans, Marinelle Basson andCampbell Davies and three anonymous reviewersimproved and clarified the ideas presented here. Thefirst author was supported by a postgraduate scholar-ship under a quantitative marine science program(QMS) co-funded by CSIRO and University of Tas-mania and the CSIRO Wealth from Oceans flagship.

REFERENCES

Alverson, F.G. (1959) Geographical distribution of yellowfintuna and skipjack catches from the Eastern Tropical PacificOcean, by quarters of the year, 1952–1955. Bull. Inter-Am.Trop. Tuna Comm. 3:167–213.

Bach, P., Gaertner, D., Menkes, C., Romanov, E. and Travassos,P. (2009) Effects of the gear deployment strategy and currentshear on pelagic longline shoaling. Fish. Res. 95:55–64.

Baird, M.E., Timko, P.G., Middleton, J.H., Mullaney, T.J., Cox,D.R. and Suthers, I.M. (2008) Biological properties acrossthe Tasman Front off southeast Australia. Deep-Sea Res. I55:1438–1455.

Bakun, A. (1996) Patterns in the Ocean: Ocean Processes andMarine Population Dynamics. La Jolla, CA: University ofCalifornia Sea Grant, in cooperation with Centro deInvestigaciones Biologicas de Noroeste, La Paz, Baja Cali-fornia Sur, Mexico, 323 pp.

Barratt, D., Bugg, A., Barry, S. and Wise, B. (2002) Modellingthe distribution of fish resources using oceanographic vari-ables, statistical methods and GIS technologies. In: Second

International Symposium on GIS ⁄ Spatial Analyses in Fisheryand Aquatic Sciences. Vol. 2, Abstracts p. 51; 2002. pp. 197–212.

Bertrand, A., Josse, E., Bach, P., Gros, P. and Dagorn, L. (2002)Hydrological and trophic characteristics of tuna habitat:consequences on tuna distribution and longline catchability.Can. J. Fish. Aquat. Sci. 59:1002–1013.

Bigelow, K.A. and Maunder, M.N. (2007) Does habitat or depthinfluence catch rates of pelagic species? Can. J. Fish. Aquat.Sci. 64:1581–1594.

Bigelow, K.A., Hampton, J. and Miyabe, N. (2002) Applicationof a habitat-based model to estimate effective longline fish-ing effort and relative abundance of Pacific bigeye tuna(Thunnus obesus). Fish. Oceanogr. 11:143–155.

Bishop, J. (2006) Standardizing fishery-dependent catch andeffort data in complex fisheries with technology change. Rev.Fish Biol. Fish. 16:21–38.

Block, B.A., Keen, J.E., Castillo, B. et al. (1997) Environmentalpreferences of yellowfin tuna (Thunnus albacares) at thenorthern extent of its range. Mar. Biol. 130:119–132.

Boyce, D.G., Tittensor, D.P. and Worm, B. (2008) Effects oftemperature on global patterns of tuna and billfish richness.Mar. Ecol. Progr. Ser. 355:267–276.

Brill, R.W. (1994) A review of temperature and oxygen toler-ance studies of tunas pertinent to Fish Oceanogr, movementmodels and stock assessments. Fish. Oceanogr. 3:204–216.

Brill, R.W. (1996) Selective advantages conferred by the highperformance physiology of tunas, billfish, and dolphin fish.Comp. Biochem. Physiol. 113:3–15.

Brill, R. and Lutcavage, M. (2001) Understanding environ-mental influences on movements and depth distributions oftunas and billfishes can significantly improve populationassessments. In: Island in the Stream: Oceanography and Fish-eries of the Charleston Bump. G. Sedberry (ed.) Bethesda, MD:American Fisheries Society, pp. 179–198.

Brill, R.W., Block, B.A., Boggs, C.H., Bigelow, K.A., Freund,E.V. and Marcinek, D.J. (1999) Horizontal movements anddepth distribution of large adult yellowfin tuna (Thunnusalbacares) near the Hawaiian Islands, recorded using ultra-sonic telemetry: implications for the physiological ecology ofpelagic fishes. Mar. Biol. 133:395–408.

Brothers, N., Gales, R. and Reid, T. (1999) The influence ofenvironmental variables and mitigation measures on seabirdcatch rates in the Japanese tuna longline fishery within theAustralian Fishing Zone, 1991–1995. Biol. Conserv. 88:85–101.

Campbell, R. (2008) Summary of Catch and Effort InformationPertaining to the Australian Longline Fishing Operations in theEastern Tuna and Billfish Fishery. Hobart: CSIRO Marine andAtmospheric Research, Report.

Caton, A.E. and McLoughlin, K. (2005) Eastern tuna andBillfish fishery. In: Fishery Status Report – Fish Stocks Managedby the Australian Government. K.M. A Caton (ed.) Canberra:Bureau of Rural Sciences, pp. 53–79.

Cayula, J.-F. and Cornillon, P. (1992) Edge detection algorithmfor SST Images. J. Atmos. Oceanic Technol. 9:67–80.

Condie, S.A. and Dunn, J.R. (2006) Seasonal characteristics ofthe surface mixed layer in the Australasian region: implica-tions for primary production regimes and biogeography. Mar.Freshw. Res. 57:569–590.

394 J. Dell et al.


Cowles, M.K. and Carlin, B.P. (1996) Markov chain MonteCarlo convergence diagnostics: a comparative review. J. Am.Stat. Assoc. 91:883–904.

Cresswell, G. (2001) East Australian current. In: Encyclopedia ofOcean Sciences. G. Cresswell, J.H. Steele, K.K. Turekian andS.A. Thorpe (eds) San Diego: Academic Press, Inc., 525 BSt. Ste. 1900 San Diego CA 92101-4495 USA. p. 634.

Fiedler, P.C. and Bernard, H.J. (1987) Tuna aggregation andfeeding near fronts observed in satellite imagery. Cont. ShelfRes. 7:871–881.

Figueira, W.F. and Booth, D.J. (2010) Increasing ocean tem-peratures allow tropical fishes to survive overwinter in tem-perate waters. Glob. Change Biol. 16:506–516.

Fonteneau, A., Lucas, V., Tewkai, E., Delgado, A. and Demarcq,H. (2008) Mesoscale exploitation of a major tuna concen-tration in the Indian Ocean. Aquat. Living Resour. 21:109–121.

Fulton, E.A., Smith, A.D.M. and Johnson, C.R. (2004) Effectsof spatial resolution on the performance and interpretationof marine ecosystem models. Ecol. Modell. 176:27–42.

Garcia, S.P., DeLancey, L.B., Almeida, J.S. and Chapman, R.W.(2007) Ecoforecasting in real time for commercial fisheries:the Atlantic white shrimp as a case study. Mar. Biol.152:15–24.

Griffiths, S.P., Young, J.W., Lansdell, M.J. et al. (2010) Eco-logical effects of longline fishing and climate change on thepelagic ecosystem off eastern Australia. Rev. Fish Biol. Fish.20:239–272.

Gunn, J. and Block, B.A. (2001) Advances in acoustic, archivaland satellite tagging of tunas. In: Tuna: Physiology, Ecologyand Evolution. B.A. Block & E.D. Stevens (eds) San Diego:Academic Press, pp. 167–224.

Hartog, J.R., Hobday, A.J., Richard, M. and Feng, M. (2011)Habitat overlap between southern bluefin tuna and yellowfintuna in the east coast longline fishery – implications forpresent and future spatial management. Deep-Sea Res. II58:746–752.

Hobday, A.J. and Hartmann, K. (2006) Near real-time spatialmanagement based on habitat predictions for a longlinebycatch species. Fish. Manage. Ecol. 13:365–380.

Hobday, A.J., Hartmann, K. and Bestley, S. (2004) SpatialDynamics Ocean Data Explorer (SDODE). In: Pelagic Fish-eries and Ecosystems Research Group. Hobart: CSIRO MarineResearch.

Hobday, A.J., Hartmann, K., Hartog, J. and Bestley, S. (2006).SDODE: Spatial Dynamics Ocean Data Explorer. User Guide.Hobart: CSIRO Marine and Atmospheric Research.

Hobday, A.J., Flint, N., Stone, T. and Gunn, J.S. (2009) Elec-tronic tagging data supporting flexible spatial managementin an Australian longline fishery. In: Tagging and Tracking ofMarine Animals with Electronic Devices. J.L. Nielsen, H.Arrizabalaga, N. Fragoso, A. Hobday, M. Lutcavage & J.Sibert (eds) Dordrecht: Springer, pp. 381–403.

Hobday, A.J., Hartog, J.R., Timmiss, T. and Fielding, J. (2010)Dynamic spatial zoning to manage southern bluefin tuna(Thunnus maccoyii) capture in a multi-species longline fish-ery. Fish. Oceanogr. 19:243–253.

Holland, K., Brill, R., Ferguson, S., Chang, R. and Yost, R.(1985) A small vessel technique for tracking pelagic fish.Mar. Fish. Rev. 47:26–32.

Hothorn, T., Hornik, K. and Zeileis, A. (2006) Unbiasedrecursive partitioning: a conditional inference framework.J. Comput. Graph. Stat. 15:651–674.

Howell, E.A., Kobayashi, D.R., Parker, D.M., Balazs, G.H. andPolivina, J.J. (2008) TurtleWatch: a tool to aid in thebycatch reduction of loggerhead turtle Caretta caretta in theHawaii-based pelagic longline fishery. Endanger. Species Res.:Inter-Research, 12 pp.

Josse, E., Bach, P. and Dagorn, L. (1998) Simultaneous obser-vations of tuna movement and their prey by sonic trackingand acoustic surveys. Hydrobiologia 371 ⁄ 372:61–69.

King, J.R. and McFarlane, G.A. (2006) A framework forincorporating climate regime shifts into the management ofmarine resources. Fish. Manage. Ecol. 13:93–102.

Lambert, D. (1992) Zero-inflated Poisson regression, with anapplication to defects in manufacturing. Technometrics 34:1–14.

Langley, A., Briand, K., Kirby, D.S. and Murtugudde, R. (2009)Influence of oceanographic variability on recruitment ofyellowfin tuna (Thunnus albacares) in the western and centralPacific Ocean. Can. J. Fish. Aquat. Sci. 66:1462–1477.

Laurs, R.M., Fiedler, P.C. and Montgomery, D.R. (1984) Albacoretuna catch distributions relative to environmental featuresobserved from satellites. Deep-Sea Res. I 31:1085–1099.

Lehodey, P., Chai, F. and Hampton, J. (2003) Modelling cli-mate-related variability of tuna populations from a coupledocean-biogeochemical-populations dynamics model. Fish.Oceanogr. 12:483–494.

Lehodey, P., Senina, I. and Murtugudde, R. (2008) A spatialecosystem and populations dynamics model (SEAPODYM)– modeling of tuna and tuna-like populations. Prog. Ocea-nogr. 78:304–318.

Ling, S.D., Johnson, C.R., Ridgway, K., Hobday, A.J. andHaddon, M. (2009) Climate-driven range extension of a seaurchin: inferring future trends by analysis of recent popula-tion dynamics. Glob. Change Biol. 15:719–731.

Maunder, M.N., Hinton, M.G., Bigelow, K.A. and Langley,A.D. (2006) Developing indices of abundance using habitatdata in a statistical framework. Bull. Mar. Sci. 79:545–559.

Maury, O., Gascuel, D., Marsac, F., Fonteneau, A. and Rosa,A.-L.D. (2001) Hierarchical interpretation of nonlinearrelationships linking yellowfin tuna (Thunnus albacares)distribution to the environment in the Atlantic Ocean. Can.J. Fish. Aquat. Sci. 58:458–469.

Mengersen, K.L., Robert, C.P. and Guihenneuc-Jouyaux, C.(1999) MCMC convergence diagnostics: a review. In:Bayesian Statistics 6. J.M. Bernardo, J.O. Berger, A.P. Dawid& A.F.M. Smith (eds) Oxford: Oxford University Press, pp.415–440.

Mugo, R., Saitoh, S.I., Nihira, A. and Kuroyama, T. (2010)Habitat characteristics of skipjack tuna (Katsuwonus pelamis)in the western North Pacific: a remote sensing perspective.Fish. Oceanogr. 19:382–396.

Nielsen, J.L., Sibert, J.R., Hobday, A.J., Lutcavage, M.E., Arri-zabalaga, H. and Fragosa, N. (eds) (2009) Tagging andTracking of Marine Animals with Electronic Devices. Dordrecht:Springer.

Oke, P.R., Brassington, G.B., Griffin, D.A. and Schiller, A.(2008) The Bluelink ocean data assimilation system(BODAS). Ocean Model. 21:46–70.

Pearman, P.B., Guisan, A., Broennimann, O. and Randin, C.F.(2008) Niche dynamics in space and time. Trends Ecol. Evol.23:149–158.

Pearson, R.G., Thuiller, W., Araujo, M.B. et al. (2006) Model-based uncertainty in species range prediction. J. Biogeogr.33:1704–1711.



Polovina, J.J., Howell, E., Kobayashi, D.R. and Seki, M.P.(2001) The transition zone chlorophyll front, a dynamicglobal feature defining migration and forage habitat formarine resources. Progr. Oceanogr. 49:469–483.

R-Development-Core-Team (2007) R: A Language and Envi-ronment for Statistical Computing. Vienna: R Foundation forStatistical Computing.

Schaefer, K.M., Fuller, D.W. and Block, B.A. (2007) Movements,behavior, and habitat utilization of yellowfin tuna (Thunnusalbacares) in the northeastern Pacific Ocean, ascertainedthrough archival tag data. Mar. Biol. 152:503–525.

Schwarz, G. (1978) Estimating the dimensions of a model. TheAnnals of Statistics 6:461–464.

Song, L.M., Zhang, Y., Xu, L.X., Jiang, W.X. and Wang, J.Q.(2008) Environmental preferences of longlining for yellowfintuna (Thunnus albacares) in the tropical high seas of theIndian Ocean. Fish. Oceanogr. 17:239–253.

Sournia, A. (1994) Pelagic biogeography and fronts. Prog.Oceanogr. 34:109–120.

Spiegelhalter, D.J., Best, N.G., Carlin, B.P. and Linde, A.V.D.(2002) Bayesian measures of model complexity and fit. J. R.Stat. Soc. Ser. B. 64:583–639.

Strobl, C., Boulesteix, A.L., Zeileis, A. and Hothorn, T. (2007)Bias in random forest variable importance measures: illus-trations, sources and a solution. BMC Bioinformatics 8:25.

Sund, P.N., Blackburn, M. and Williams, F. (1981) Tunas andtheir environment in the Pacific Ocean: a review. Oceanogr.Mar. Biol. 19:443–512.

Trebilco, R., Gales, R., Lawrence, E., Alderman, R., Robertson,G. and Baker, G.B. (2010) Characterizing seabird bycatch inthe eastern Australian tuna and billfish pelagic longlinefishery in relation to temporal, spatial and biological influ-ences. Aquatic Conserv. Mar. Freshw. Ecosyst. 20:531–542.

Venables, W.N. and Dichmont, C.M. (2004) GLMs, GAMs andGLMMs: an overview of theory for applications in fisheriesresearch. Fish. Res. 70:319–337.

Walsh, W.A., Howell, E.A., Bigelow, K.A. and McCracken,M.L. (2006) Analyses of observed longline catches of bluemarlin, Makaira nigricans, using generalized additive modelswith operational and environmental predictors. Bull. Mar.Sci. 79:607–622.

Ward, P. and Bromhead, D. (2004) Tuna and billfish fisheries ofthe Eastern Australian Fishing Zone and adjacent high seas.Noumea, New Caledonia. National Tuna Fishery Reportpresented at the 17th meeting of the Standing Committeeon Tuna and Billfish (8–18 August 2004, Majuro, Republicof the Marshall Islands), 17 pp.

Ward, P. and Myers, R.A. (2005) Inferring the depth distribu-tion of catchability for pelagic fishes and correcting forvariations in the depth of longline fishing gear. Can. J. Fish.Aquat. Sci. 62:1130–1142.

Young, J.W., Bradford, R., Lamb, T.D., Clementson, L.A.,Kloser, R. and Galea, H. (2001) Yellowfin tuna (Thunnusalbacares) aggregations along the shelf break off south-easternAustralia: links between inshore and offshore processes. Mar.Freshw. Res. 52:463–474.

Young, J.W., Lansdell, M.J., Campbell, R.A., Cooper, S.P.,Juanes, F. and Guest, M.A. (2010) Feeding ecology andniche segregation in oceanic top predators off eastern Aus-tralia. Mar. Biol. 157: (11).p.2347.

Zagaglia, C.R., Lorenzzetti, J.A. and Stech, J.L. (2004) Remotesensing data and longline catches of yellowfin tuna (Thunnusalbacares) in the equatorial Atlantic. Remote Sensing Environ.93:267–281.

Zainuddin, M., Kiyofuji, H., Saitoh, K. and Saitoh, S.-I. (2006)Using multi-sensor satellite remote sensing and catch data todetect ocean hot spots for albacore (Thunnus alalunga) in thenorthwestern North Pacific. Deep-Sea Res. II 53:419–431.

Zainuddin, M., Saitoh, K. and Saitoh, S.I. (2008) Albacore(Thunnus alalunga) fishing ground in relation to oceano-graphic conditions in the western North Pacific Ocean usingremotely sensed satellite data. [Fish. Oceanogr.] 17:61–73.

396 J. Dell et al.


Estimation of yellowfin tuna (Thunnus albacares) habitat in waters adjacent to Australia’s East Coast: making the most of commercial catch data

Documents