Double‐tagging scores of seabirds reveals that light‐level geolocator accuracy is limited by species idiosyncrasies and equatorial solar profiles

Light-level geolocators (‘geolocators’) are one of the most popular and practical tools available to study animal movement, with well-established, open access standards techniques guide analyses geolocation data (see Lisovski et al., 2020). However, there can be considerable uncertainty associated accuracy location estimates derived from light-level data. Geolocators small (i.e. ~0.3–3.3 g) archival loggers that measure record solar intensity at regular intervals, some capability measuring archiving other information such as water temperature, wet/dry events barometric pressure. When retrieved, downloaded directed into astronomical equations estimate spatial locations based on timing twilight sunrises sunsets). Geolocator interpolated or two positions per day latitude estimated by length, longitude local midday midnight relative Greenwich Mean Time Julian (Hill, 1994). were first applied tracking movements marine vertebrates including elephant seals (Delong 1992), fish (Block 1998), seabirds (Croxall 2005; Egevang 2010; González-Solís 2007; Guilford 2009; Phillips 2006; Shaffer Tuck 1999) sea turtles (Fuller 2008). Recently, have undergone miniaturisation improvements onboard storage capacity, which has stimulated an increase in studies use infer about both terrestrial species otherwise too burdened devices (Bridge 2011). The number ecologists using seabird increased tandem these sensor newly designed methods implemented several R packages for processing analysing (e.g. Merkel 2016; Rakhimberdiev 2017; Sumner 2009). Despite growing volume geolocator data, used to-date been empirically tested more precise technologies only three albatross limited latitudinal breadth (Phillips 2004; 2005). These employed older sensors recorded light levels infrequently previous threshold method software that, unlike modern methods, did not incorporate movement models probabilistic algorithms. measured distances satellite Platform Terminal Transmitter (PTT) corresponding assessed mean accuracies ± standard deviation (SD) 186 114 km 2004) 202 171 (Shaffer estimations coarse those obtained fix orbiting Advanced Research Global Observation Satellites (ARGOS), a typical 1–3 (Burger & Shaffer, 2008) Positioning System (GPS), regularly average less than 10 m (Hulbert French, 2001) ~15 (Forin-Wiart 2015). instruments satellites tend large many may prohibitively expensive. Satellite typically power capacity on-board memory storage, depending species, place unreasonable burdens birds terms wing-loading hydrodynamic drag Furthermore, attaching feathers long-term deployments is suitable because they periodically moult. For harness attachment remedy this constraint recommended 2003) mortality device-induced behaviours (Barron 2010). therefore offered attractive year-round alternative tackles constraints larger, spatially accurate technology. inherently prone accuracy, particularly generally considered become under increasingly ‘equatorial’ profiles; is, either nearer Equator (spatial variation) equinox (temporal where when length changes shallowly (Ekstrom, Hill, 1994; inherent geolocations fluctuates date, even if amount shading remains constant (Lisovski 2012). Weather cloud cover) behavioural patterns roosting periods induce errors night thus thought affect In addition, collected during breeding stages often reduced due specific might curves 2012, example, roost ground (Corre Jouventin, 1997; Schreiber Chovan, 1986), brood their young twilights (Howell Bartholomew, 1969) shade sensors, nest underground burrows 2006). fitted go before last depart sites after could exhibit abnormal transitions between dark times curve (Gow, 2016). Another behaviour wide-ranging occur within single (Clay 2018; McDuie Such large-scale impact interpolations shifting noon, compressing elongating all speed direction travel time year seabirds, leg rings, likely idiosyncratic differences being stationary 2012; Welch Eveson, 1999). Accordingly, it suggested performance dependent 2005) choice algorithm position (Musyl 2001). well-known hypothesised limitations geolocation, satisfactory studying foraging ranges 2004), habitat preferences distributions pelagic (Egevang Halpin Lascelles Congdon, Pollet 2014; Quillfeldt Here, we sought evaluate algorithms diverse sample free-flying assess whether affected tracked behaviours. Past technology and/or al. 2005), evaluated static tags carried resident fixed environment Fudickar 2012), out sizes adequate statistical disentangle Our objectives (a) investigate reported applicable context large-scale, multi-species study; (b) test advanced models; (c) sea-surface temperature (SST) interpolation improved new d) model situational factors accuracy. To address aims, conducted field synchronous GPS individual around world. We geolocator-interpolated routes tracks, effects individuals, inter-twilight travelled geolocation. analysed 151 chick-provisioning concurrently ‘double tagged’) packages, FlightR (Rakhimberdiev 2017) probGLS (Merkel Tracking represent 200 across five 11 separate colonies 2011 2019 (Table 1). originally had 278 double-tagged (some individuals tagged once years), but dataset excluding insufficient produce calibrations, poor. Breeding located regions Southern Europe (Mediterranean Sea), West Africa (east Atlantic Ocean) Australia (south Pacific Ocean). northern hemisphere, Cape Verde Shearwaters (Calonectris edwardsii, n = 11; 2014 2018), Cory's (C. borealis, 100; 2013–2018), Scopoli's diomedea, 61; 2014–2018) Red-billed Tropicbirds (Phaethon aethereus, 7; 2017–2018) latitudes 15°N–40°N. southern White-necked Petrels (Pterodroma cervicalis, 21; 2018 2019) colony 29°S. immersion sensors: BAS_MK19 (British Antarctic Survey) Biotrack_MK3005 [formerly BAS_MK19] (Biotrack Ltd), every minute maximum value 5 min immersed continuously 25 min; Intigeo-C330, Intigeo-C250 C65-SUPER (Migrate Technology minute, storing 20 min. conductivity (wet/dry) sampling rate was 6 s models. Devices mounted tarsus mounting darvic metallic ring plastic cable tie, Velcro© (38 mm; Paskal) hook-and-loop harness. Tesa© tape (4651; Tesa Tape Inc.) taping contour scapulae, base four central rectrices shearwaters petrels six tropicbirds. All processed R, version 3.5.1 (R Core Team, 2020), measurements calculated World Geodetic (WGS 1984) ellipsoid. analyst who no knowledge attributes paired so decisions parameterising influenced prior birds' underlying movements. This done ensure our would comparable practitioners bird travelled. tracks standardised package adehabitatLT (Calenge, 2006) resampling equal interval rates varied among colonies. gap-filled except 1 hr occurred fixes. account erroneous caused poor reception, allowable flight velocity 27.8 m/s (100 km/hr) consecutive taxa. realistic (Lascelles Depending brand geolocator, imported raw functions readMTlux TwGeos (Wotherspoon 2016) ligTrans GeoLight Hahn, then automated event sunsets) annotation function preprocessLight level 1, presented above differentiate noise log-transformed Following guidelines (2020), visually reviewed identify any areas series manually inspected each event, subsequently deleting deemed falsely annotated procedure, light. Indistinguishable unclear becoming shaded weather, plumage. procedure resulted transition exclusion 33.6% Shearwaters, 29.1% 33.9% Tropicbirds, 32% 14% Petrels. expected see greater proportion excluded phase. Contrary non-breeding, migratory phase visit nests, raft visiting nests cause obscured times. analysis seabirds: probGLS. Using produced ‘TAGS’ files TwGeos2TAGS FLightR preparation analyses. 185 probGLS, included 156 same datasets (15 collect SST exclusively device water). Data While on-bird calibration possible centrally placed 2017), foraging. Calibrations ‘rooftop calibrations’ packages. parameterised (FlightR probGLS) calculate bounding box extending 35° direction, 25° nearest pole 50° Equator. package, likelihood movements, uses hidden Markov true unobserved state. Inference performed particle filter, template-fit allow 2017). also incorporates biologically relevant parameters improve estimates. function, requires measures relationship observed data) theoretical current elevation angles executing algorithm, calibrated deployment seabird. Analyses run without masks explore how land masking set daily 1,500 50 grid. locations, ran filter million particles median posterior probability distribution relocations. SST, 2016), geolocations. iterative forward step selection process, computing weighted potential (10,000 point cloud) producing path iterations track dry (probable SD (m/s), supplementary metadata) Spear Ainley (1997) dry-logger m/s, matching relocations; wet modest drift ocean long (fastest 1.3 m/s). Geolocations mask prevent estimating inland coasts. encountered bird, computed 4 matched satellite-derived (0.25° × 0.25°, NOAA OI V2 High-Resolution Dataset). masks. discrepancy positions, distance geographic fixes ±30 min, respectively, given pair sunset–sunrise vice versa) geolocator-estimated noon/midnight period. expressed great-circle kilometres individual's its events. potentially nonlinear predictor variables estimates, constructed generalised additive mixed-effects (GAMM) gamma log link function. separately modelled predictors accuracy: displacement (kilometres) successive (from GPS) period duration daylength geosphere (Hijmans, 2019). univariate thin-plate regression splines. initially error: closeness March September equinoxes; consider independent strong concurvity predictor, equatorial profile index proximate mechanism governing effect. identity nested type random model. Both contain avoid over land. masked covariates error, restricted relatively Mediterranean Sea. Therefore, (both packages), (probGLS) inbuilt outlier detection (FLightR). correlogram examine residual autocorrelation Some evidence evident lag, thinning include second third observation effect overall results. Thus, thin mgcv (Wood, checks converged basis dimension sufficiently (using permutation presence pattern along predictors). residuals followed assumption structure heterogeneity variances against candidate predictors. initial results applying (±SD) 432 460 372 290 respectively 2). (for (probGLS only) 408 473 304 413 km, As additional equinoxes 3 weeks (21 days) (km SD) 227 250 369 GPS-derived package-specific 38.6% 84.5% (parameter 4, Table 2) 3, 2), fell 23.7% 88.8% geolocations, respectively. Estimated uncertainties provided material. provide material ≤21 outside ≥21 periods. Results outputs different Supporting Information. found bell-shaped (Figure 1a, FLightR: F6.47 993, p < 0.001; Figure 2a, probGLS: F7.1 718, 0.001). demonstrated drastically declines approach 12 closer Equator) best approximately 9 15 (Figures 1a 2a). ≤10 ≥14 243 232 239 (with SST) detection) significant respective 1c, F154.7 25.7, 2c, F158.2 11.8, 1d, F4.0 5, 2d, 16.8, 1b, F2.3 40.3, 0.001), 2b, F0.5 0.35, 0.111). differed 3), poorest results, whereas Petrel Shearwater consistently better 3) runs applied. assessment conditions. advance understanding findings contextualises hypotheses past tests until now testing tags, sizes, outdated methods). emphasise need spatiotemporal considering temporal effects) expect travel. If former wandering, circuitous tropical path, practitioner should adequately aware what scale analysed. birds, do luxury enhance interpolation. lower method. Moreover, (as much 76.3% 15.5% probGLS). varies species. previously (2020) (2005), inconsistent result species-dependent performance, relates way combination species-specific behaviour, morphology, plumage use. It smaller larger models, reliably hypothesis confounding individuals. quality influence 2012 detailed discussion), important carefully planning study. particular, capture complete variability care must taken focal contributed seen tracked—particularly case very distance, looping trips away whence calibrated. ‘rooftop’ commonly biologists. directly biologists studies. modelling showed were. explained species' morphology extremely short tarsi) nesting habits, returning sunset sunrise, affecting performance. Conversely, suggest habits twilights. vastly supports assertion correction good possibly area Sea) compared ocean-foraging Spatial sunsets strongest FLightR. made short-range basin, hence little diminish package. application will forced area, improving mask. still highest appeared weakest suggests non-breeding seabirds. reasons, researchers working coastal-foraging basins achieve useful seems sometimes start deployments, short-term deployments. because, determined low recognise paths, state driven Petrels, largest (more double species). Bayesian priors incorporated algorithm. benefit inclusion optional prior. events, broadly consistent latitude) contrasted empirical geolocator-observed data). demonstrate why similar globe, animals move ever degree present order magnitude (94–1,043 km) (Beck 2002; Delong 1992; Hull, 1999; Teo but, considerably (186–202 km; Based Gunn Le Boeuf 2000; essential variation types, resulting differently parametrised mask, etc.) validate recommendations concerning reporting parameters. Specifically, clearly unambiguously report assumptions compute unquestionably ecology organisms, cases options sensitive urge caution consideration draw inferences regional exact science combinations parameterisation, yield uncertainties. key message criticism uncertainty, demonstration adopt dynamic vary valid irrespective chosen, tracked. make profiles especially geolocator-derived precise, explicit conservation management actions implemented. L.R.H. received funding support Endeavour Postgraduate Leadership Award Australian Government. Fieldwork supported awards Holsworth Wildlife Endowment—Equity Trustees Charitable Foundation Ecological Society Birdlife Stuart Leslie Bird Award. New South Wales Office Environment Heritage Parks Australia. R.R. acknowledges postdoctoral Ramón y Cajal contract Spanish Ministry Economy, Industry Competitiveness (RYC-2017-22055) Z.Z. doctoral grant University Barcelona (APIF-2012). grateful staff Norfolk Island National Park, volunteers, T. Chatwin, D. Dow, C. Lee, Terrington L. Blight assistance project logistics work. Monash Animal Ethics Permit BSCI/2017/42 wildlife research approvals. gathered valuable help students volunteers whom enormously grateful. following permits: Cabildo de Lanzarote num. I-15-16-T, Gran Canaria 1169, Excmo. 2107, 1948, 2018/2290, Generalitat Valenciana 249-18 SGENP, Gobierno Canarias 2011/0795, 2015/1170 2016/9887, 2018/6432, Govern Balear ANE-02/2017, ANE-05/2016, CEP-24/2015, CEP-30/2016, CEP-31/2014, Parque Nacional 160/15, RES-AUT I03/2015, Región Murcia AUF20140056. authors declare conflict interest. J.D.R. conceived idea co-designed research; acquired scientific permits, led writing manuscript; generated R.M. helped development N.C. ideas, permit acquisition assisted collection N.G. responsible collating data; J.G.-S. J.M.R.-G., T.M., Z.Z., M.C.-F., S.S., V.M.-P., L.Z. F.D.F. L.N.-H. R.H.C. development, permits manuscript. critically manuscript drafts gave final approval publication. code archived Dryad Digital Repository https://doi.org/10.5061/dryad.gb5mkkwpf (Halpin 2021). Please note: publisher content functionality supporting supplied authors. Any queries (other missing content) author article.