Fine‐scale oceanographic drivers of reef manta ray (Mobula alfredi) visitation patterns at a feeding aggregation site

Abstract Globally, reef manta rays (Mobula alfredi) are in decline and are particularly vulnerable to exploitation and disturbance at aggregation sites. Here, passive acoustic telemetry and a suite of advanced oceanographic technologies were used for the first time to investigate the fine‐scale (5‐min) influence of oceanographic drivers on the visitation patterns of 19 tagged M. alfredi to a feeding aggregation site at Egmont Atoll in the Chagos Archipelago. Boosted regression trees indicate that tag detection probability increased with the intrusion of cold‐water bores propagating up the atoll slope through the narrow lagoon inlet during flood tide, potentially transporting zooplankton from the thermocline. Tag detection probability also increased with warmer near‐surface temperature close to low tide, with near‐surface currents flowing offshore, and with high levels of backscatter (a proxy of zooplankton biomass). These combinations of processes support the proposition that zooplankton carried from the thermocline into the lagoon during the flood may be pumped back out through the narrow inlet during an ebb tide. These conditions provide temporally limited feeding opportunities for M. alfredi, which are tied on the tides. Results also provide some evidence of the presence of Langmuir Circulation, which transports and concentrates zooplankton, and may partly explain why M. alfredi occasionally remained at the feeding location for longer than that two hours. Identification of these correlations provides unique insight into the dynamic synthesis of fine‐scale oceanographic processes which are likely to influence the foraging ecology of M. alfredi at Egmont Atoll, and elsewhere throughout their range.

Extensive targeted and bycatch fisheries of M. alfredi, driven in part for their gill plates [prebranchial appendages, used to filter their zooplankton prey from the water (Paig-Tran et al., 2013), which are utilized in the Asian medicinal trade ], have led to dramatic subpopulation declines in recent decades (Couturier et al., 2013;Lawson et al., 2017;Marshall et al., 2019;Rohner et al., 2017). Population recovery from such exploitation is hindered by their conservative life-history traits; the species are slow-growing, late to mature, and only have a few offspring in their lifetime (Dulvy et al., 2014;Stevens, 2016).
Mobula alfredi are particularly vulnerable to exploitation and changes in climate at feeding aggregation sites. For example, anthropogenic disturbance may reduce individual M. alfredi fitness by driving them away from productive feeding areas (Murray et al., 2019;Venables et al., 2016), which has been highlighted as a major conservation concern for the species (Harris et al., 2020;Murray et al., 2019). Feeding behavior may also be disrupted by enhanced stratification driven by rising sea surface temperatures, which can decrease marine phytoplankton (Roxy et al., 2016), and with it zooplankton biomass (Richardson, 2008).
Studies which investigate M. alfredi aggregation behavior have associated their occurrence with various broadscale physical factors, such as wind speed, moon phase, sea surface temperature, and tidal phase (Couturier et al., 2018;Dewar et al., 2008;Jaine et al., 2012;. However, the fine-scale changes in the oceanographic environment that potentially drive feeding aggrega- tions have yet to be investigated. Situated in the central Indian Ocean, the Chagos Archipelago ( Figure 1) has been uninhabited for many decades (excluding Diego F I G U R E 1 The Central Indian Ocean with Chagos Archipelago; British Indian Ocean Territory indicated within the red box (left inset). The Chagos Archipelago with Egmont Atoll indicated within the red box (left). Egmont Atoll and the location of the oceanographic and acoustic receiver mooring in Manta Alley (red and yellow dots) and four acoustic receivers (green dots) (top right). Bathymetric view of Manta Alley obtained via multibeam survey (E. Robinson, P. Hosegood, A. Bolton, unpublished data) showing the location of the moorings (bottom right). Bottom right legend showing instrument configurations of the long thermistor string (red dot/pin) and subsurface ADCP moorings (yellow dot/pin) deployed 182 m apart anchored at a depth of 66 m. Z is the height above the seabed Garcia Atoll; Sheppard et al., 2012). Due to the lack of human influence, such as coastal development and anthropogenic pollution, the region is considered virtually pristine (Readman et al., 2013). Owing to the region's unique marine environment, a no-take marine protected area (MPA), which encompasses the entire exclusive economic zone (EEZ; 640,000 km 2 ) except for a 3 nm exclusion around the boundary of Diego Garcia Atoll, was established in 2010 (Sheppard et al., 2012). The archipelago supports a subpopulation of M. alfredi which is largely undocumented due to the remoteness of the location and strict protective measures; as is the region's physical oceanographic environment . Broadscale studies conducted in the region indicated that Egmont Atoll, situated in the southwest of the archipelago, provides key habitats for this M. alfredi subpopulation (Andrzejaczek et al., 2020;Harris, 2019). Feeding M.
alfredi are regularly observed around the atoll (Harris, 2019), behavior which is thought to be associated with shallow bathymetry, low current speeds, and cooler sea surface temperatures (Armstrong et al., 2016;Couturier et al., 2018;Harris, 2019;Jaine et al., 2012;. Together, these factors may act to induce upwelling of nutrients, increasing primary and secondary production (McManus et al., 2005). Furthermore, currents interacting with topography may aggregate zooplankton (Genin et al., 2005), resulting in highly productive feeding grounds for a range of species, including M. alfredi .
Field observations have identified an M. alfredi feeding aggregation "hotspot" at the north of Egmont Atoll (J. Harris and G. Stevens, unpublished data). However, M. alfredi feeding activity can be dramatically different from one day to another, with little apparent change in broadscale oceanographic conditions (Harris, 2019). Therefore, a greater understanding of how M. alfredi respond to fine-scale environmental drivers is needed. Here, passive acoustic telemetry and in situ oceanographic monitoring are used to investigate M. alfredi activity at Egmont Atoll, and assess what physical factors drive fine-scale (5-min) visitation patterns at the observed feeding aggregation hotspot. This study aims to enhance the current understanding of M. alfredi foraging ecology by providing detailed insight into their fine-scale movement patterns in response to natural changes in their oceanographic environment.

| Study site
The Chagos Archipelago is comprised of seven atolls, several large submerged banks, and more than 60 low lying islands, located at

| Oceanographic moorings
Two instrumented oceanographic moorings were deployed in Manta Alley ( Figure 1) from a research ship on 30th December 2019. Both moorings were positioned within Manta Alley. The first was a subsurface taut-line mooring deployed in 66 m, with the uppermost buoyancy element at a depth of 20 m. Temperature was measured by RBRSolo 3 T temperature sensors positioned at 2 m intervals from 4 to 48 m above the seabed. In addition to the temperature sensors, RBR Concerto conductivity-temperature-depth (CTD) sensors with a sampling interval of 5 s were positioned at 2 and 50 m above the bed. An acoustic receiver (see acoustic receiver array section below) was positioned approximately 7 m below the near-surface CTD, at 43 m above the bed. The second mooring, deployed 182 m southeast of the first, comprised an upward-facing Nortek Signature 500 kHz acoustic Doppler current profiler (ADCP), mounted on a subsurface buoy 3 m above the seabed. Both moorings were recovered on 17th March 2020; however, the Nortek Signature 500 kHz ADCP had ceased sampling on 10th March 2020. F I G U R E 2 Reef manta rays (Mobula alfredi) engaged in feeding activities at the Manta Alley feeding aggregation site in north Egmont Atoll. Photo by Simon Hilbourne, Manta Trust

| Acoustic tag deployment
Tagging activities were carried out at Egmont Atoll between November 19, 2019 and December 3, 2019 while freediving. Twenty VEMCO V16-4x acoustic transmitter tags (Vemco Inc.), each tethered to a titanium anchor (Wildlife Computers) with a small diameter cable, were deployed on the right dorsal musculature using a modified Hawaiian hand sling while swimming behind the M. alfredi.
Each tag was set to operate at 69 kHz and transmit a unique acoustic signal at random intervals between 30 and 90 s. Before being tagged, the ventral side of each M. alfredi was photographed to capture their unique spot pattern for identification purposes , and their sex and size class (a proxy of maturity status) were recorded (Stevens, 2016). Five of the twenty M. alfredi that were tagged were re-sighted at their tagging locations between three and 12 days after deployment. All five were observed to be engaged in normal feeding activities (Stevens, 2016). All activities were approved by the University of Plymouth Animals in Science Ethics Committee under permit ETHICS-24-2019.

| Acoustic tag analysis
All tag detection data were imported into VUE software (version 2.6.2) and filtered for active tags. The False Detection Analyser (VUE version 2.6.2) was then used to identify false detections, whereby the ratio of short and long periods between detections is calculated from the time between detections on each receiver (Simpfendorfer et al., 2015). Here, the default short to long periods of <30 min and >12 hr, respectively, were used (Simpfendorfer et al., 2015) and all detections suspected to be false were removed from analysis.
The percentage of sightings at each location was then projected in ArcGIS 10.7.
To assess whether Egmont Atoll can be considered a key habitat, residency indices (RI) were calculated using the following form , allowing comparison of residency patterns at Egmont Atoll between M. alfredi regardless of differences in tracking periods (Daly et al., 2014).
To assess the intensity at which locations were utilized, the amount of time each tagged M. alfredi spent within the detection range of each acoustic receivers was calculated using the VTrack R package (Campbell et al., 2012) in R 3.5.2 (R Core Team, 2018).
Briefly, each tag detection was classed as a resident or nonresident events. A resident event began when there were two or more successive detections (Nalesso et al., 2019) at the same receiver within 60 min. Termination of the resident event occurred at the time of the last detection when there were no further detections within 60 min, or when the tag was detected at least twice at another receiver (Campbell et al., 2012;Nalesso et al., 2019).

| Environmental influences: boosted regression trees
Boosted regression trees (BRT) were used to investigate the relationship between environmental variables and the visitation patterns of tagged M. alfredi to the feeding aggregation site at Manta Alley. The modeling technique is based on two algorithms: regression trees models and boosting, which build and combine large numbers of relatively small trees by fitting each new tree to the residuals of the last (Elith et al., 2008). Each tree is constructed through a series of binary splits of predictor variables (Hastie et al., 2009), which occur based on the homogeneity of their relationship to the response variable (Colin et al., 2017). Multiple splits are tested, and partitioning occurs when the greatest improvement of homogeneity is found (Colin et al., 2017). Advantages of this modeling technique include its ability to fit complex, nonlinear relationships, model interactions between response variables (Elith et al., 2008), and the appropriate data model does not require assumptions about the residuals of the model (Derville et al., 2016).
Detection data were divided into a time-series of 5-min bins starting from 1st December 2019 and ending on 10th March 2020.
The BRT was then constructed with a binomial response of present (1) or absent (0) within each 5-min bin. The final time-series contained 28,654 × 5-min bins of presence and absence observations. Nine predictor variables representing temperature (1-2), zooplankton biomass (3), ocean currents (4-8), and tide (9), all of which have been shown to influence M. alfredi occurrence (Anderson et al., 2011;Harris et al., 2020;O'Shea et al., 2010), were selected for inclusion (Table 1). Temperature variables included the following: temperature at 2 m above the seabed (temp 2 m) (1), and 50 m above the seabed (temp 50 m) (2), sampled every 5 s using RBR Concerto CTDs. Data were pooled into the same 5-min bins as the presence and absence data. The mean temperature for each 5-min bin was then calculated from the 60 data points. For zooplankton biomass converted to depth data using RSKTools inbuild conversion function.
Data were cleaned with a median filter and averaged with a running window (both size 501 points). The Matlab inbuilt find peaks function was then used and ran twice to pick out both high and low tides by inverting the data on one run. A 5.5 hr minimum peak spacing was specified to further reduce susceptibility to noise, and the resulting Note: All predictors are in 5-min means unless otherwise specified. All distances are meters above the seabed. Mean values show the value at which the predictor is held for partial dependency and interaction plots.
TA B L E 1 Description of the predictor variables used in boosted regression trees analysis of tagged Mobula alfredi occurrence at Manta Alley data points were visually validated against the raw depth data. The variable time relative to high tide (9) was then calculated with high tide as zero and negative hours before (flood) and positive hours after (ebb; . All models were fitted using the gbm.step() function of the dismo R package (Hijmans et al., 2017). Initial models were built to find suitable settings for four parameters: tree complexity (tc), which specifies the number of interactions that should be modeled, learning rate (lr), which regulates the contribution of each tree to the growing model, bag fraction (bf), which controls stochasticity by randomly selecting (without replacement) a specified subset of the data at each iteration and step size (ss), which controls the number of trees which should be added at each iteration (Elith et al., 2008). The following parameter settings were tested: tc = 1-6, lr = 0.01, 0.005, 0.001 and 0.0001, bf = 0.5, 0.7, 0.9, ss = 25 and 50, resulting in 144 models.
Ten-fold cross-validation (CV) was applied to assess model performance, whereby the model is fitted to training data and then is tested against a withheld portion (hold-out sample) of the dataset (Elith et al., 2008). The model's ability to fit the withheld data was then measured by comparing the area under the receiver operat- The percentage of deviance explained by the model was determined using the pseudo determination coefficient (D 2 ), calculated using the following form (Nieto & Mélin, 2017): The final model was fitted with tc = 6, lr = 0.005, bf = 0.7, and ss = 50 (Table S2). The relative contribution of predictor variables to the BRT model is measured by averaging the number of times a variable is chosen for splitting and the squared improvement resulting from these splits (scaled to 100 across all the variables; Elith et al., 2008). To ensure noninformative predictors were not hindering model performance, pairwise correlation coefficients and variance inflation factor (VIF) estimates (Jouffray et al., 2019) were calculated, all were in an acceptable range; coefficients <0.6 and/or VIF estimates <3.5 (Jouffray et al., 2019; Table S1; Figure S1).
Due to the complex nature of BRTs, model results cannot be easily visualized. Therefore, partial dependency and interaction plots were generated for interpretation. The plots display the results of the predicted effect on tag detection probability for a given predictor, or pair of predictors, after accounting for the mean effects of all other predictors (Elith et al., 2008;Hastie et al., 2009  used to generate a distribution under the null hypothesis of no interaction among predictors (Jouffray et al., 2019).
There were a total of 15,965 detections during the study period (Table 2). The highest percentage of detections occurred at the acoustic receiver deployed on the oceanographic mooring in Manta Alley (51.4%), followed by North IdR Cleaning Station (22.3%; Figure 4).
The overall distribution of detections by hour of the day shows 70.9% of detections occurred at Egmont Atoll during the day (06:00-18:00; Figure 5). For adults, only 18.3% of detections occurred at night (19:00-05:00), while 35% of detections occurred at night for juveniles.
The mean total time between tag deployment when the tags first began transmitting until the end of the study, when the detection data were downloaded, was 113 ± 5 day (range 106-119 days).
During this time, tagged M. alfredi were tracked (first to last tag detection) for a mean of 97 ± 32 days (range 3-116 days), with a mean of 50 ± 23 detection days (range 2-92 days). Residency indices show that tagged M. alfredi were detected at Egmont Atoll for a mean of 52% of the days they were tracked (RI = 52 ± 15.7%), with a minimum and maximum RI of 24% and 80.3%, respectively (Table 2). Mean residency indices were similar for both adults and juveniles (including sub-adults), which were 53 ± 16% and 51 ± 16%, respectively.
Overall, 2074 resident events were recorded for 19 M. alfredi

| Environmental influences: boosted regression trees
Model performance evaluation for the BRT, including all nine predictors, had outstanding and excellent predictive performance for the training ( T AUC = 1) and cross-validated ( CV AUC = 0.89) data, respectively, with minimal evidence of overfitting (ΔAUC = 0.11). The estimated D 2 suggests that 72% of the deviance was explained (Table S2).
Detection probability was higher with greater downward vertical velocity (12.8%), and during the early stages of a flood tide (12.1%), approximately two hours following low tide when near-surface longshore current velocity (longshore (v) 48.5 m, 11.7%) was ap-   Figure S2). These interactions should not be considered in isolation, as they may arise separately or simultaneously, and may be affected by other variables. However, they provide insight into the estimated influence of several paired-environmental processes which can increase the probability of tag detections. For example, tag detections probability was highest when upward currents speed (vertical velocity) was increased, and the near-surface temperature ( Figure S2a) was warmer (temp 50 > 29.5°C). Tag detection probability also increased with cooler near-bed temperature (temp 2 m < 24°C) and increased near-surface temperatures (temp 50 m > 29.5°C; Figure S2b). It also increased when near-bed cross-shore currents

| D ISCUSS I ON
Overall, M. alfredi residency at Egmont Atoll, measured by the residency index (RI), was high (mean RI = 52%), which supports previous reports that Egmont Atoll provides key habitats for this species (Andrzejaczek et al., 2020;Harris, 2019). Similar high levels of residency have been observed in the Red Sea (mean RI = 65%; Braun et al., 2015) and at the Amirante Islands of Seychelles (mean RI = 62%; . Adult and juvenile M. alfredi displayed similar residency at Egmont Atoll, which is in contrast to patterns observed in Seychelles, where the RI was lower for adults , indicating that the M. alfredi habitat at Egmont Atoll is perhaps consistently important for all life stages.
Alternatively, the similar residency of adults and juveniles could be attributed to the acoustic array design .
To establish a more robust RI, future research would benefit from increased spatial coverage, including the deployment of acoustic receivers in locations which may be frequented by juveniles such as inside the lagoon.
Overall, detection data display diel behavior patterns, with the Note: Higher interaction size values indicate a more substantial interaction effect; near zero indicates negligible interactions. All interactions were significant (p < .01). The suggested influence of the interaction on the probability of detections is described along with the maximum detection probability estimated for each interaction. Setyawan et al., 2018). These patterns may be associated with the species use of cleaning stations, where cleaner fish are only active during the day (Côté, 2000). Diel movement patterns may also be associated with efficient foraging strategies. For example, M. alfredi may predominately frequent shallow reef habitats during the day to feed on reef-associated zooplankton which can accumulate in surface waters over shallow reefs when avoiding predation from reef-dwelling diurnal consumers (Alldredge & King, 2009;Leichter et al., 2013). At night, M. alfredi may then travel offshore to forage (Couturier et al., 2018;Dewar et al., 2008;Jaine et al., 2012) when diel vertically migrating zooplankton ascends into warmer water (Braun et al., 2014;Couturier et al., 2018;Dewar et al., 2008). This hypothesis is supported by stable isotope analysis, which indicates that a large proportion of M. alfredi diet is made up of both near-surface and demersal zooplankton (Couturier et al., 2013;. The diel M. alfredi movement pattern was less pronounced for juveniles, which were more frequently detected at night than adults, suggesting that juvenile M. alfredi remain in shallower reef habitats longer. This pattern has also been observed in other subpopulations and is likely a predator avoidance strategy by the more vulnerable juveniles (Cerutti-Pereyra et al., 2014;Stewart et al., 2018). Their smaller body size may also make it less energetically efficient for juveniles to travel offshore (Nøttestad et al., 1999;, and/or their foraging experience may be limited .

Manta Alley had the highest number of detections, and there
were repeated resident events for 18 of the 19 tagged individuals, indicating a high level of site fidelity. Site fidelity is a well-reported characteristic of M. alfredi, having been observed in photographic identification, acoustic telemetry, and satellite tagging studies (Couturier et al., 2018;Deakos, 2011;Dewar et al., 2008;Jaine et al., 2014;Kessel et al., 2017;McCauley et al., 2014;Stevens, 2016).
Site fidelity has been attributed in part to the species' reliance on specific habitats, which provide a sufficient food resource, protection from predation, and opportunities to clean, socialize, and reproduce (Jaine et al., 2014;McCauley et al., 2014;Perryman et al., 2019;Stevens, 2016). Resident events were longer at Manta Alley than at any other location. The depth of the majority of the area within the range of the acoustic receiver is greater than 40 m. As 40 m is the maximum depth of occurrence for cleaner fish in the Chagos Archipelago (Kuiter, 2014), it is unlikely that these extended resident events at Manta Alley are associated with cleaning activities. A high tag detection probability of M. alfredi occurred with cold near-bed and warm near-surface temperature, and probability increased with increasing difference between these temperatures. Extreme short-term fluctuations in near-bed temperatures may be associated with the intrusion of cold water created by internal waves which disrupt the thermocline (Shanks et al., 2014).
Enhanced concentrations of zooplankton often occur at the thermocline, the thickness of which can be increased by internal waves (McManus et al., 2005). These internal waves break as they interact with the steep slope of an atoll leading to the formation of coldwater bores which propagate up the slope Woodson, 2018). Bores enhance the upward transport of organisms, and thus the concentration in surface waters (Stevick et al., 2008), which may provide efficient foraging opportunities for the zooplanktivorous M. alfredi. The upward propagation of cold-water bores has been observed to vary tidally Leichter et al., 1996), and can become more frequent during a flood tide leading to a pulsed delivery of organisms (Leichter et al., 1996;Woodson, 2018). Here, tag detection probability was high during the early stages of a flood tide and was also increased by the interaction effect between a flood tide and cooler near-surface temperature, which may indicate that cold-water bores propagate up the slope (Leichter et al., 1996). Plankton sampling and oceanographic measurements obtained inside the lagoon also indicate that increased zooplankton abundance is associated with the transfer of plankton into the lagoon from the intrusion of cold-water bores created by breaking internal waves (Sheehan et al., 2019). The intrusion of cold water may also provide metabolically advantageous feeding conditions for M. alfredi by reducing the energetic cost of feeding activities (Lawson et al., 2019).
In the current study, tag detection probability also increased with the interaction between high-velocity near-bed cross-shore currents flowing inshore and high levels of backscatter. This interaction may indicate that zooplankton is being carried from the thermocline into the lagoon during a flood tide and is likely pumped back out during ebb. Due to the partially enclosed morphology of the lagoon, water entering is likely to be restricted by the narrow subtidal passages. Even with a low tidal amplitude, strong jet-like currents can be generated (Dumas et al., 2012), which may increase the density of inflowing (outflowing) zooplankton approaching low tide (in the early stages of flood), as suggested by the in-water observations of the current study. During these events, mobile zooplankton may actively seek refuge zones to avoid predation or import into (export from) the lagoon (Pagano et al., 2017). Refuge zones include the thermocline and behind shallow back reefs (Leichter et al., 2013), where zooplankton become concentrated further, providing dense assemblages of prey for M. alfredi. Similar theories of zooplankton retention, which are also related to tide phase, have been suggested in other regions (Armstrong et al., 2016;Stevens, 2016).
The BRT also provided some evidence of the presence of Langmuir Circulation (LC), which can trap and concentrate particles in the water column (Smith, 2001). The process is driven by wind and waves which produce helical vortices that appear as rotating cells that rotate perpendicular to the wind direction (Smith, 2001). The interaction effect between high-velocity near-bed longshore currents flowing, when near-surface currents were flowing in the opposite direction, and high backscatter intensity could be evidence of LC cells. Alternating cells rotate in opposite directions leading to areas of convergence and divergence (Talley et al., 2011). Downwelling, which increased the probability of tag detections, occurs in areas of convergence where plankton, other organisms, and particles become trapped in highly concentrated bands (Kingsford et al., 1991;Thorpe, 2004). These bands may provide ideal foraging opportunities for M. alfredi. As LC can persist for hours or even days (Gargett et al., 2004), it could potentially be associated with resident events which last longer than the influence of the tide (>2 hr). The characteristic surface "slicks" which often accompany LC have also been regularly observed by authors in Manta Alley, further supporting this suggestion.
Under well-mixed conditions, LC can develop "super-cells" which extend the full depth of the water column (Gargett et al., 2004).
These super-cells can transport organisms and partials from depths up into the water column where they become concentrated in the narrow bands of the convergence zones (Gargett et al., 2004;Kukulka et al., 2012). alfredi were observed feeding where there were surface slicks and a high density of particles in the water column (Dewar et al., 2008), which is characteristic of LC convergence zones (Kingsford et al., 1991). Around Lady Elliot Island (LEI) in Australia, sightings of foraging individuals and increased acoustic tag detection were correlated with wind speed (Couturier et al., 2018;Jaine et al., 2012).
At LEI, M. alfredi sightings and tag detections peaked at wind speeds around 18 km/hr (approximately 5 m/s), an optimal speed for the development of LC (Langmuir, 1938;Plueddemann et al., 1996). At LEI, sightings were also associated with cooler sea surface temperatures, with a decrease in sightings and detections with increased temperature (Couturier et al., 2018;Jaine et al., 2012). Strong surface warming can lead to a breakdown of LC by disrupting the balance between wave-forcing and thermal convection (density-driven circulation ;Li & Garrett, 1994;Min & Noh, 2004), which may reduce the density of prey, leading to a lower number of sightings and tag detections of

M. alfredi.
There were some limitations to the current study. alfredi spending long periods of time at the location.

ACK N OWLED G EM ENTS
This study was made possible with funding from the Garfield Weston Foundation and the Bertarelli Foundation, and it contributes to the Bertarelli Programme in Marine Science. All work was approved by the British Indian Ocean Territory Administration (BIOTA, permit numbers: 0008SE19 and 0006SE20). We owe special thanks to Annie Murray, Adam Bolton, Craig and Mickael, and all the research vessel crew who provided invaluable support in the field. We also thank Jill Schwarz, Lauren Peel, Vinay Udyawer and Jean-Baptiste Jouffray who provide advice. Thank you to the two anonymous reviewers for offering constructive feedback which improved the manuscript.

CO N FLI C T O F I NTE R E S T
There are no competing financial, professional, or personal interests that might have influenced the performance or presentation of the work described in this manuscript. All authors have no conflict of interest to declare.

DATA AVA I L A B I L I T Y S TAT E M E N T
The data that support these findings are available from FigShare