Shallow seamounts are “oases” and activity hubs for pelagic predators in a large-scale marine reserve

The study explored all major known seamounts located within the Ascension Island EEZ, including the Harris Stewart Seamount, the Grattan Seamount, and the Young Seamount (Fig 1). Grattan and Young are sister peaks located ca. 80 km apart, on or adjacent to the mid-Atlantic ridge, and are jointly known as the “Southern Seamounts.” All the features studied lie 260 to 320 km offshore from Ascension Island, which itself lies more than 2,000 km from Africa—the nearest continental land mass (Fig 1). Detailed bathymetric mapping carried out during the study shows that Ascension’s seamounts can be divided into 2 groups along geographic lines. Grattan and Young are both shallow features rising to within 101 m and 77 m of the surface, respectively, with summits that intersect the deep chlorophyll maximum (the zone of peak primary production in tropical oceanic waters) and are subject to strong internal tides that drive semi-diurnal upwellings of cooler water from deeper strata (S1 Fig and Table A in S1 Text). The summit of the Harris Stewart seamount, in contrast, lies entirely within the mesopelagic “twilight zone,” consisting of a broad plateau at ca. 500 m depth punctuated by 2 distinctive domes that constitute the shallowest points at depths of 260 to 300 m (S1 Fig and Table A in S1 Text). These differences had a profound impact on their pelagic ecology.

Strong evidence of a predator-aggregating effect was apparent in baited remote underwater video (BRUV) and visual surveys conducted over the shallow Southern Seamounts, but not around the deeper Harris Stewart seamount (Fig 2). In total, 29 species of pelagic sharks, fish, and cetaceans were observed in BRUV transects (Table B in S1 Text). When considering only large-bodied predators (i.e., sharks and teleosts at trophic level >4), species richness was found to increase significantly within 5.4 km of the summits of the Southern Seamounts (GAM, p 5.4 km, to 4.6 species per deployment (95% CI: 3.5 to 6.0) at the summit (Fig 2A). Predator biomass was also substantially elevated within 3.4 to 5.3 km of the Southern Seamounts (GAM, p Fig 2B).

Fig 2.

Variation in (A) species richness and (B) estimated relative biomass of sharks and predatory teleosts (trophic level >4) observed in stereo BRUV surveys as a function of distance from Ascension Island’s outlying seamounts. Estimated relative biomass is the maximum biomass of each taxa observed in a single video frame based on photogrammetric fork length measurements of individual animals converted using established length–weight relationships (see Methods). Solid trend lines and shaded envelopes are fitted smooths and associated 95% confidence intervals from negative binomial GAMs. Broken lines represent regional oceanic baselines from a reference set of 56 BRUVs surveys conducted >50 km from any seamount or island (see S1 Fig). Vertical gold lines represent the estimate radius of influence of the seamount where biomass or species richness began to singificantly increase and exceeded the regional baseline (see Methods). Data are plotted separately for the 2 Southern Seamounts but are modelled jointly based on AIC model selection which favoured a single global distance smoother for these adjacent shallow features in all cases. The data underlying this figure can be found in S2 Data. Illustrations Marc Dando (sharks) and Diane Rome Peebles (fish). AIC, Akaike’s information criterion; BRUV, baited remote underwater video.

https://doi.org/10.1371/journal.pbio.3003016.g002

However, shallow seamounts did not aggregate all pelagic predator species equally. Of the 8 most frequently observed species in BRUV surveys (those present in ≥6 deployments), 5 were found to be significantly more abundant within 1 to 5 km of the Southern Seamounts, including silky sharks (Carcharhinus falciformis), Galapagos sharks (Carcharhinus galapagensis), yellowfin tuna (Thunnus albacares), rainbow runner (Elagatis bipinnulata), and wahoo (Acanthocybium solandri) (GAM, all p ≤ 0.037; Fig 3A and Table C in S1 Text). Abundances of blue shark (Prionace glauca), mahi mahi (Coryphaena sp.), and sailfish (Istiophorus albicans) did not vary significantly with distance from the summits of these features over the scales sampled by our radial transects (GAM, all p > 0.1; Table C in S1 Text and S3 Fig). However, when considering all deployments made within 20 km of the Southern Seamounts (n = 41), the probability of occurrence of sailfish was an order of magnitude higher than regional oceanic baselines, which may indicate a looser association (seamount probability = 0.20; oceanic probability = 0.018; Z test: = 6.9, p = 0.009) (S3 Fig and Table B in S1 Text). Similarly, mixed results were found in vessel-based visual surveys. Of the 7 seabird species observed in visual transects (Table D in S1 Text), only 2—the Ascension frigatebird (Fregata aquila) and the sooty tern (Onychoprion fuscatus)—showed evidence of local enhancement around seamounts, with elevated abundances detected within ca. 2 km of the summit (GAMM, all p ≤ 0.035; Fig 3B and Table E in S1 Text). Other commonly observed species such as storm petrels (Oceanodroma sp.) and shearwaters (Procellariidae sp.) exhibited no significant associations with seamounts (GAMM, all p > 0.7; S4 Fig and Table E in S1 Text). No cetacean species were observed in visual transects and only a single individual was detected in BRUV surveys suggesting that, unlike some seamounts [10], our study features are not important aggregation areas for these species (at least during the sampling period).

Fig 3.

Variation in relative abundance of predatory fish, sharks, and seabirds observed in (A) BRUVs surveys and (B) vessel-based visual transects as a function of distance from seamount summits. Relative abundances are expressed as the maximum number of individuals observed in a single video frame for each BRUV deployment (MaxN) and as the number of individuals observed in a 300 m belt transect per 5-min sampling window for visual transects. Solid trend lines and shaded envelopes are fitted smooths and associated 95% confidence intervals from negative binomial GAM(M)s. Only species exhibiting significant seamount associations are shown (see S3 and S4 Figs for all species). Broken lines and gold bars in (A) represent the regional oceanic baseline and approximate radius of influence of the seamount for each species as per Fig 2. The data underlying this figure can be found in S2 Data (A) and S3 Data (B). Illustrations Marc Dando (sharks), Diane Rome Peebles (fish), and Peter Harrison (seabirds). BRUV, baited remote underwater video.

https://doi.org/10.1371/journal.pbio.3003016.g003

No evidence of elevated predator species richness, total predator biomass, or predator abundance were found in comparable BRUV surveys and visual transects conducted over the deeper Harris Stewart seamount (GAM, all p > 0.11; Figs 2 and S6 and Table G in S1 Text), although only a single species, blue shark (Prionace glauca), was observed sufficiently frequently to analyse abundance trends. A weak positive trend in biomass of predatory teleosts was detected in BRUV transects conducted over this feature (GAM, p = 0.006); however, the predicted biomass at the summit was below the regional oceanic baseline (Fig 2).

To investigate how the dominant pelagic predators observed in our surveys interact with seamounts, a combination of passive acoustic and satellite telemetry was used to study the movements of individual Galapagos sharks (n = 15), silky sharks (n = 14), yellowfin tuna (n = 9), and bigeye tuna (Thunnus obesus, n = 9) captured from summit aggregations (Tables H and I in S1 Text). Although the latter were not observed sufficiently frequently in shallow diurnal BRUV surveys to model abundance trends (probably because they spend the day foraging in mesopelagic waters [26]), this deeper-schooling species is known to associate with seamounts [10,22] and was regularly caught during dawn/nocturnal fishing around our study features.

Sharks.

Galapagos and silky sharks tagged with acoustic transmitters over the Southern Seamounts consisted predominantly of juveniles and subadults (mean ± SD fork length, Galapagos: 127 ± 23.5 cm, range = 105 to 196 cm; silky: 125 ± 14.3 cm, range = 102 to 152 cm; Table S8), which is consistent with size ranges estimated from BRUV surveys (silky sharks: 138 ± 22.5 cm; Galapagos sharks: 158 ± 21.0 cm; Table S8). Detections recorded by summit hydrophone arrays reveal that Galapagos sharks were almost continuously resident over the 595-day tracking period and had significantly higher mean residency indices (RI = proportion of tracking days with ≥1 detection) than silky sharks, which were more likely to disperse away during early phases of the study (Galapagos: mean RI = 0.97; Silky shark: mean RI = 0.68; quasibinomial GLM: F_1,28 = 23.7, p Fig 5A). The proportion of tagged silky sharks remaining on the Southern Seamounts declined rapidly over the first 90 days, but then stabilised at ca. 50% for the remainder of the study (Fig 5A). Partitional clustering around medoids identified 2 distinct groups of silky sharks which corresponded with this biphasic curve: those with RI ≤ 0.34 (“transients”; n = 7) and those with RI ≥ 0.76 (“residents”; n = 7) (Fig 3). Neither fork length nor sex influenced the observed RIs (quasibinomial GLM: all F p > 0.31). Galapagos and silky sharks also behaved differently during periods of residency on seamounts. While the former exhibited high and relatively constant mean detection probabilities across the diel cycle on both Southern seamounts (GAM, Grattan: F = 0.53, p = 0.042: Young: F = 0.0, p = 0.39), silky sharks displayed a strong circadian rhythm, with detections declining shortly after dusk and returning before dawn (GAM, all F > 16.2, p Fig 5C).

Unfortunately, locations from sharks double-tagged with fin-mounted GPS tags (n = 9 per species) were generally too sparse to reconstruct their movements during periods of absence from seamount summits. However, regular transmissions over a 137-day period by a single silky shark were potentially revealing. Following an initial period of residency on the Grattan seamount lasting 46 days, this individual ranged extensively over the south-eastern quadrant of the Ascension EEZ, making 2 return visits to the Young seamount after absences of 20 to 30 days during which it travelled up to 140 km from the seamount summits (Fig 6A). Similar movements between Grattan and Young were also apparent in detections of acoustically tagged animals. While fidelity to the seamount on which tagging originally occurred was generally high, 5 (36%) silky sharks and 2 (13%) Galapagos sharks moved between the Southern Seamounts at least once during the study (S10 Fig).

Fig 6. Movements and site fidelity of satellite-tagged sharks and tuna captured at the Grattan, Young and Harris Stewart seamounts.

(A) Movements of a single silky shark (C. falciformis) tagged with a fin-mounted Argos-GPS SPOT transmitter on the Grattan Seamount in June 2017. (B–D) Movements of (B) 9 yellowfin tuna (T. albacares) and (C, D) 9 bigeye tuna (T. obesus) fitted with PSATs at the Southern Seamounts (B, C) and Harris Stewart Seamount (D) between May 2017 and February 2018. In each case, upper panels show the posterior mean tracks of tagged individuals along with the overall “utilisation distribution” (UD), of the population derived from tag-specific SSMs (see Methods). UDs are presented as volume contours enclosing a given proportion of the joint SSM posterior probability density and were calculated using a weighted average that gives greater influence to individuals with longer deployment periods. Lower panels show how distance from the tagging seamount(s) varied with time for each posterior mean track, along with the population mean (blue line; tuna only) and associated 95% confidence interval (shaded regions) derived from SSM posterior simulations. Broken lines in lower panels (B–D) are isopleths describing the probability (p) that a hypothetical individual would have been located within a given distance of their tagging seamount on each relative tracking day if they dispersed randomly. Probabilities were estimated using correlated random walk simulations (n = 10,000) and constitute a test of the null hypothesis that tagged animals showed no fidelity to seamounts. The data underlying this figure can be found in S6 Data. Illustrations: Marc Dando (sharks) and Diane Rome Peebles (tuna). Bathymetry basemaps were sourced from GEBCO. PSAT, pop-up satellite archival tag; SSM, state-space model.

https://doi.org/10.1371/journal.pbio.3003016.g006

Midwater BRUV surveys were conducted using drifting camera assemblies that have been described in detail elsewhere [3,46]. Each rig consisted of a stereo pair of high-definition cameras (GoPro HERO3) trained on a bait canister filled with 1 kg of macerated tuna sourced as waste from a local fish processing facility. Rigs were deployed in sets of 5, suspended at a depth of 10 m and spaced 200 m apart along short, floated longlines. All deployments took place during the daytime (07:00 to 18:00) with a soak time of 2 h, during which each rig gathered continuous video footage. Video data were analysed using EventMeasure (https://www.seagis.com.au/event.html) to identify, count, and measure individuals crossing the field of view. Camera rigs were photogrammetrically calibrated to estimate the lengths of individuals, which were in turn converted to mass estimates using published weight-length relationships (see [46] for details). To avoid double counting (within and between rigs), the relative abundance and biomass of each species encountered was calculated as the maximum number (MaxN) or maximum combined mass (of individuals observed in a single frame when pooled across all rigs in a set). These metrics therefore represent the minimum abundance and biomass of each taxa present in the survey area. For each set, we also calculated the species richness as the total number of species for which at least one individual was observed. The majority of BRUV surveys were carried out during the principal seamount expedition in May to June 2017. However, 9 of 59 (15%) deployments (3 per seamount)—primarily over summit areas—occurred during scoping and follow-up visits between 20th January to 20th February 2017 and 2018. We therefore included “season” as a factor in all subsequent models. These visits formed part of larger research cruises during which more extensive BRUV sampling using identical methods was undertaken in oceanic waters (>50 km from any island or seamount) throughout the Ascension EEZ (n = 56 surveys; S2 Fig). We did not include oceanic samples in models of seamount distance effects due to the very large spatial scales involved (up to 560 km from summits). However, means and CIs across all oceanic deployments were used to define regional baselines for estimating seamount radius of influence (see Modelling).

Vessel-based counts of surface-orientated marine vertebrates, including seabirds and flying fish, were carried out as described in [43]. Surveys were performed by a single observer stationed on the bow of a 22 m vessel (MV Extractor) travelling at a constant speed (approximately 8 knots) and heading and involved counting all individuals observed in flight or at the surface in a belt transect approximately 300 m wide (i.e., visible to the naked eye with identifications confirmed by binocular if necessary). Individual transects were approximately 1 h in duration (range = 0.4 to 1.5) and covered an average distance of 14.6 km. To enable finer-scale analysis of seamount effects, within transects, counts of each species were pooled in 5-min observation intervals which formed the response variables for subsequent statistical analyses. We used time rather than distance to aggregate counts as it was easier to standardise in the field and provided a relevant measure of survey “effort” for mobile species, such as seabirds, that were almost always observed in flight rather than resting on the water.

Hydroacoustic estimates of relative faunal biomass were generated using a hull-mounted Simrad EK80 echosounder operating at 3 frequencies (38, 70, and 120 kHz) with a 2 s ping rate. The acoustic settings and calibration parameters used are listed in Table J in S1 Text. Surveys consisted of a combination of nocturnal (20:00 to 06:00) and diurnal (07:00 to 18:00) transects, supplemented with more focussed nocturnal “zigzag” surveys over the summits of the Southern Seamounts (Fig 1). Acoustic data analysis was performed by developing a series of algorithms in Echoview processing software (v12+, Echoview Software Pty Ltd, Hobart) and consisted of 2 main steps: (1) data cleaning to exclude various forms of noise; (2) echo classification to assign the observed backscatter into coarse taxonomic groupings (fish, zooplankton) based on characteristic frequency responses following the methodology of Ballon and colleagues [47] (see S1 Text). For each of the final classes obtained by the discrimination algorithm, we then exported the nautical area scattering coefficient (NASC; a proxy for biomass) integrated over 500 m distance intervals for the whole water column (0 to 300 m) and stratified by depth layers (every 50 m) which formed the response variables in subsequent analyses. The fish NASC was exported at 38 kHz using a minimum threshold value of −70 dB while plankton NASC was exported at 120 kHz using a minimum threshold of −80 dB. Spatial patterns in total water column NASC of each taxon was mapped using ordinary kriging performed using the gstat [48] package for R, taking the centroid of each distance interval as the measurement point.

To provide contemporaneous estimates of water column structure and primary productivity around seamounts at the time of our biodiversity surveys, conductivity-temperature-depth (CTD) profiles (≥100 m depth) were collected at varying distances from seamount summits using a Valeport MIDAS CTD+ (MV Extractor) and SeaBird SBE 911+ CTD (RRS James Clark Ross) with integrated fluorometers for assessment of water column chlorophyll-α concentration (Fig 1B). From each profile, we then extracted the depth of the surface mixed layer (isothermal layer) along with the depth and concentration of the deep chlorophyll maximum (DCM). Fluorometers were not calibrated to local conditions and thus provided relative estimates of chlorophyll concentration only. To account for different calibration settings, the instrument used was included as a covariate in all subsequent analyses.

Seamount-induced chlorophyll enhancement (SICE) has been shown to be an ephemeral phenomenon that may not be detected in snapshots from ship-based surveys [15]. To provide a longer-term integration of primary productivity around our study seamounts, we therefore used high-resolution satellite imagery (https://resources.marine.copernicus.eu/, product: OCEANCOLOUR-GLO-CHL-L3-REP-OBSERVATIONS-009-085) to calculate 5-year (2016 to 2020) climatological mean sea surface chlorophyll (SSCHL) anomalies for all 4 × 4 km cells located within a 40 km radius of the summit of each feature. To assist with the detection of weak seamount effects, we also computed a local “Chlorophyll Enrichment Index” (CEI) for each daily scene following methods described in [24] and calculated the 5-year mean CEI for each cell. The CEI is defined as the percentage difference between the chlorophyll concentration measured at each cell and the mean value of all cells in a reference area located within a 30 to 90 km radius. For all analyses, we excluded satellite scenes with >50% cloud cover to reduce artefacts resulting from spatially uneven sampling. For analysis of seamount distance effects, we then calculated the average chlorophyll concentration and CEI in 1 km increments from the nearest seamount summit, helping to eliminate spatial autocorrelation in the raw raster values.

SICE is expected to arise due to upwelling of cool, nutrient rich waters driven by interactions between seamounts and ocean tides and currents. To explore potential upwelling effects, we analysed high-resolution time series of water temperature (10 s) and relative current velocity (1 s interval) recorded over a 600-day period by acoustic telemetry receivers (see below) moored at depths of 84 to 306 m around the summits of Grattan and Young seamounts. Time series were de-trended using loess smoothers and processed using spectral density estimation (spectrum function in R v3.6.1 [49]) to assess the amplitude and frequency of any periodic oscillations that might indicate diapycnal mixing.

The influence of seamounts on all oceanographic and ecological variables measured during the study was evaluated using generalised additive (mixed) models (GAM(M)s) implemented in the package mgcv [50] for R v 3.6.1 [49]. To each response variable (y) we fit a model of the basic form:

Where E(y) is the expected value, f(distance) is a penalised smooth function (thin plate spline basis) of distance from the nearest seamount summit to capture any nonlinearity in the form of the relationship, β₁ is the coefficient of an optional parametric term x (“season” in analysis of BRUV and visual transect data and “vessel” for CTD data) and g is a link function to map values from the response scale to the scale of linear predictor (identity in the Gaussian case and natural logarithm in the non-Gaussian). The exception was the analysis of depth-stratified hydroacoustic data where we specified a 2D tensor product smooth of distance and depth, f(distance, depth), to model the interacting effects of these 2 variables. To allow for variable seamount topography, including seamounts with multiple subpeaks, we defined the “summit” as the area shallower than 200 m for the Southern Seamounts and shallower than 300 m for Harris-Stewart (S1 Fig). Distance was then measured to the deployment location (BRUVs, CTDs), centroids of transect segments (hydroacoustic and visual surveys) or centroids of raster cells (SSCHL, Enrichment Index). We used distance from summit rather than bathymetry to test seamount effects because: (1) global altimetry-derived bathymetry was coarse beyond the area mapped in this study and was shown to be inaccurate for some of our features; and (2) for epipelagic species, distance from topographic features is likely to be a more important predictor of habitat use than bathymetry once seabed depths exceed >1,000 m.

Conditional distributions of the response variables were assumed to be negative binomial for overdispersed counts (BRUV and visual survey data), Tweedie for nonnegative continuous variables (hydroacoustic NASC), and Gaussian for oceanographic variables (DCM, DCM depth, ML depth, SSCHL). For continuous sampling methods (visual and hydroacoustic surveys), we included a random effect of transect to reflect the hierarchical structure of the data and, where necessary, incorporated nested autoregressive error structures to avoid overfitting associated with residual spatial and temporal autocorrelation. To determine whether seamount-distance effects differed significantly between our study features, for each response variable we initially compared 3 sets of models in which smooth function, f, was (1) allowed to vary by individual seamount; (2) varied by seamount depth (i.e., separate smooths for the shallow, Southern seamounts and deeper, Harris Stewart seamount); or, (3) assumed a single, global smooth for all features. Distance effects were fit as factor-smooth interactions using the “by” command in “mgcv” and the best model selected by minimization of the Akaike’s information criterion (AIC_c). AIC is not appropriate for comparing GAMMs with non-Gaussian errors fit using penalised quasi-likelihood, so in this case we used the “smooth_difference” function in R package “gratia” [51] to compute the 95% confidence intervals (CIs) of the differences between the fitted splines at 100 regularly sampled points along the distance axis. Where 2 splines did not differ significantly along their entire lengths (i.e., 95% difference CIs included zero), we assumed that a single smooth was sufficient. Final tests of variable significance were based on Bayesian p-values estimated in package mgcv [52], as this provides a consistent approach for GAMs fit using maximum-likelihood and GAMMs fit using PQL. Full details of model fitting, selection, and checking procedures and associated R code are provided in S1 Text and S1 Code.

Since one of our objectives was to assess the spatial extent and magnitude of any seamount aggregating effect, wherever a significant distance effect was detected, the “derivatives” function in the R package gratia [51] was used to identify regions of the fitted GAM(M) spline where the slope of the relationship was significantly nonzero (i.e., 95% CIs around the finite difference first derivatives did not overlap zero; see S1 Text). We defined the radius of influence (R) of the seamount as the furthest point at which a significant change was first detected and the magnitude of the effect as the ratio between the predicted value at the summit (i.e., E(y|distance = 0)) and the mean predicted value at distances > R (i.e., pelagic baselines). Uncertainty in the magnitude was estimated by repeatedly simulating from the posterior of the fitted GAM(M) (n = 1,000 simulations), recomputing the percentage change and extracting the 95% CIs. Simulations indicated that this method accurately retrieved R except for non-Gaussian cases where the pelagic baseline approached zero, because the “seamount effect” could not be distinguished from a monotonically increasing, linear relationship on the scale of the linear predictor. This primarily affected MaxN abundances of some individual species in BRUV surveys. In these cases, we defined R as the distance at which model predicted abundance exceeded mean MaxN from a larger set of oceanic reference BRUVs deployed throughout the Ascension Island EEZ over a similar period, or 0.05, whichever was greater. The magnitude of the seamount effect was not estimated in these cases. Custom R code used for estimating extent and magnitude of seamount effects is provided in S1 Code.

Study animals were captured opportunistically with baited handlines (sharks), by trolling lures (tuna on Southern Seamounts) or using short, vertical longlines (tuna on Harris Stewart) after which they were immobilised and winched to deck level using a custom tagging sling. Throughout all tagging procedures, a saltwater hose inserted into the mouth was used to continuously irrigated the gills and a wet cloth was placed over the eyes to minimise distress. Galapagos and silky sharks were fitted with internal acoustic fish tags (Vemco V16-6x, Bedford, Nova Scotia, Canada; dimensions: 16 × 95 mm) implanted into the abdominal cavity through a small incision made anterior to the anal fin and offset from the midline. Surgical implements and the insertion site were first sterilised with a Betadine solution and the wound was closed with surgical sutures. For a subset of individuals, Fastloc GPS-enabled Argos transmitters (SPOT-F-338, Wildlife Computers, Redmond, Washington, United States of America) were also mounted to the dorsal fin using plastic hex bolts and stainless-steel lock nuts such that the base of antenna was level with the apex of the fin. All tuna were equipped with pop-up satellite archival tags (MiniPAT, Wildlife Computers) that were secured through the dorsal pterygiophores using titanium darts and monofilament tethers as described previously [53]. Tags were programmed to release automatically after periods of 180 to 360 days or following >72 h at constant depth (±3 m). Complementary acoustic tagging of tuna was trialled but abandoned as the exceptionally high density of sharks was assessed to severely compromise the survival of fish released following surgery.

Acoustically tagged sharks were tracked using fixed receiver arrays (Vemco VR2AR-69khz) established on the summits of the Grattan and Young seamounts between 1st and 3rd June 2017 (n = 7 per seamount). Receivers were deployed encircling the summit plateau of Grattan and on prominent peaks and ridges of the more topographically complex Young seamount (S1 Fig). Mean (± SD) seabed depth at deployment locations was 161 ± 64 m on Grattan (range = 131 to 306 m) and 158 ± 48 m on Young (range = 84 to 202 m). Average minimum spacing between receivers was 1.36 ± 0.9 km on Grattan (range = 0.88 to 3.43 km) and 1.75 ± 0.3 km on Young (range = 1.48 to 2.17 km). Receivers were tethered to a ~100 kg anchor and suspended 10 m above the seabed attached to a syntactic foam buoy (HBF-13-750, DeepWater Buoyancy, Maine, USA) providing approximately 10 kg of lift to allow subsequent retrieval at the surface. Range tests were conducted on a subset of receivers (n = 5) using a matching acoustic tag programmed with a 10-s ping rate and suspended approximately 10 m below a slowly drifting vessel whose position was continuously logged using GPS. Binomial GAMMs were then used to model the relationship between detection probability (proportion of expected detections received per 180 s time bin) and distance from the receiver (including horizontal and vertical displacement from the test tag). Following Scherrer and colleagues [54], we defined the average maximum detection radius (AMDR) as the point at which mean predicted detection probability dropped below 5%, estimated as 839 m across the 5 sites tested. Receivers were recovered by triggering the integrated acoustic release mechanism with a Vemco VHTx transponding hydrophone: once in January 2018 to service and download data and again in February 2019 when the study was terminated.

Positional data from SPOTs and PSATs were processed using tag-specific state-space models (SSMs) to estimate most probable locations at regular 12-h intervals along with their associated posterior probability distributions. SPOT data were analysed using continuous time correlated walk (CTCRW) models implemented in the R package crawl [55], whereas PSAT data were processed using the manufacturer’s proprietary GPE3 model following methods described by Richardson and colleagues [53] (with a travel speed standard deviation of 1.4 FL s^-1). To avoid excessive interpolation associated with sparse data and long transmission gaps, SPOT tracks were first cut into segments with ≤14 days between successive locations and CTCRW models fit only to those segments with >10 locations. To investigate predator site fidelity and space use around seamounts, for each tagged individual and population (i.e., species-seamount cohort), we first derived an overall probability distribution describing the likelihood that a given area was used by merging the individual 12-hourly probability density surfaces on a common grid using a weighted average that gives greater influence to animals with longer deployment periods (see [53] for details). For each individual, we also calculated the distance travelled from their original tagging seamount at each interpolated point along their posterior mean track and computed the 95% CIs using 1,000 random draws from the posterior distribution of each location. Since location errors associated with PSAT geolocations are typically large, high-quality Argos positions (classes 1–3) at the point of tag release often provide the only confident estimate of position. To assess whether tagged tuna were located closer to seamounts at this release point than would be expected by chance, we simulated 10,000 correlated random walks with the same duration and movement parameters (i.e., step length and turning angle distributions) as the observed tracks (see S1 Text and S2 Code for details). The proportion of simulations that ended closer to seamounts than each observed track was then used as a test of the null hypothesis that an individual showed no site fidelity (α = 0.05).