Our study area is approximately 1,200 km north of Perth, in the western Pilbara inshore region (Figure 1). The 10,600 km² area encapsulates the wide protected embayment of Exmouth Gulf and coastal waters north to the north-east of Barrow Island, and out to the 20 m isobath (Figure 1A). An ancient, submerged coastline (exposed during the Last Glacial Maximum), the study area is situated on the inner shelf and is predominantly flat and shallow (Wilson, 2013). Depth gradually increases at the outer, seaward, edge of the study area with the first drop in depth at the 20 m bathymetric contour (Figure 1B). Habitats in the area include patchy mangroves, broad expanses of intertidal algal mats (in the south), ephemeral seagrasses, and both intertidal (fringing islands and on limestone platforms) and subtidal tropical coral communities. Ningaloo Marine Park (MP) and Muiron Island Marine Management Area (MMA) (both within the Ningaloo Coast World Heritage Area) are in the south-west, and Barrow Island MP and MMA are in the north-east (Figure 1B). Strong tidal currents between the North West Cape and Muiron Islands flush internal waters of Exmouth Gulf, which maintains normal oceanic salinity levels and facilitates an exchange of nutrient rich waters and associated fish assemblages between deep and shallow environments. The major currents that facilitate productivity in the western Pilbara are the Holloway Current and the Leeuwin Current (Wilson, 2013), the former feeding into the latter. The southward flowing Leeuwin Current is WA’s prevailing oceanographic feature. It is strongest in May–July and during La Niña years; weakest between October to March and during El Niño years (D’Adamo et al., 2009). The northward Ningaloo Current is dominant on the inner shelf from September to mid-April and supports the retention of planktonic biomass in the ecosystem (Taylor and Pearce, 1999). Seasonal cyclonic activity is high between October and March and is expected to increase because of climate change (Abbs, 2012).

The study area encompasses a large portion of the newly recognised Ningaloo Reef to Montebello Islands Important Marine Mammal Area (IMMA), identified by the IUCN for cetacean biodiversity including coastal dolphins.

Dolphin sightings were opportunistically recorded (as a second priority taxa) during aerial surveys designed for dugongs, following the protocol of Marsh and Sinclair (1989) as refined by Pollock et al. (2006). Ten surveys were conducted over a four-year period: May, July, October, and December 2012, and then May (Austral autumn) and October (Austral spring) in 2013, 2014, and 2015. Survey flights were conducted using a fixed-high-wing, six-seater Partenavia P68B aircraft, at an altitude of 500 ft (152 m) above sea level, and at a groundspeed of 100 knots. The flights were restricted to weather conditions suitable for detecting dugongs and, by default, coastal dolphins (wind speeds ≤10 knots). A systematic, parallel strip transect survey design followed the same transects used for previous dugong studies (Preen et al., 1997; Prince et al., 2001; Hodgson, 2007) (Figure 1A). Three blocks were flown within the study area – Exmouth Gulf, Onslow and a repeated area within the Onslow block that included the development footprint for which the dugong monitoring programme was conducted.

Experienced observers and a double platform approach [i.e., two observers on either side of the aircraft as outlined by Marsh and Sinclair (1989) and Pollock et al. (2006)] were used to maximise accuracy and enable corrections for perception bias in recording dugongs. A block-out curtain and separate intercom systems visually and acoustically isolated observers on the same side of the aircraft so that sightings were made independently. Marker rods attached to pseudo wing struts on both sides of the aircraft delineated 200 m (at ground level) strips consisting of four 50 m zones. Observers focussed on recording sightings within the strip, however, if they saw any animals as well as those inside (i.e., within the narrow space between the aircraft and the first marker rod) and just outside (i.e., beyond) the strip they also recorded those (largely to ensure observers weren’t tempted to record sightings as within the strip when they were not). Sighting data were recorded by observers in real time onto a portable two-track audio recorder and later processed using time stamps and zone information to merge double sightings from the teams of two observers. Sighting conditions (glare, visibility, cloud cover, and sea state) and effort data were entered into a customised iPad app (Marsh, 2015) by the survey team leader approximately every 2 min or whenever conditions changed. The location of the aircraft was recorded by two Garmin eTrex 10 Global Positioning System units every 2 s (up to 15 m positional accuracy).

For each dolphin sighting, the following data were recorded: taxon (species or genus); taxon identification certainty (as “certain,” “probable,” or “guess”), group size; calf presence; transect zone (“low,” “medium,” “high,” or “very high” – each being a consecutive 50 m further away within the observed strip width); and turbidity on a scale of 1–4 (“1”: clear water, shallow depth and the seafloor is clearly visible; “2”: variable water clarity, variable depth range and the seafloor is somewhat visible; “3”: clear water, depth greater than 5 m and the seafloor is not visible; “4”: turbid water, variable depth range and the seafloor is not visible; adapted from Pollock et al., 2006). Certainty in taxon identification was based on morphological characteristics and surfacing behaviour of the animal, and influenced by the quality of the observation which can be affected by factors including distance, duration of animal availability, aspect and sun reflection. Surveys were conducted in passing mode, meaning that the pilot did not circle back over animals to confirm species or group size estimates. Consistency amongst observers was achieved through pre-survey training and continual peer-moderation. We have considered dolphin species recorded as “unknown” or a “guess,” as unidentified dolphin species. Bottlenose dolphin sightings were assumed to be of Indo-Pacific bottlenose dolphins because the study area was bound to the 20 m bathymetric contour. A genetic investigation off northern WA by Allen et al. (2016) identified Indo-Pacific bottlenose dolphin occurrence within the 50 m bathymetric contour, and common bottlenose dolphin (Tursiops truncatus) occurrence beyond it. It is possible that common bottlenose dolphins were also encountered but in such low proportions that they would be unlikely to skew the data.

Dolphin locations were adjusted from the aircraft’s position on the transect line, to the middle of the zone in which they were recorded, and plotted in ArcGIS v 10.1 (ESRI, 2012), assuming dolphins were directly abeam of the aircraft at the time recorded. Dolphins recorded just “inside” or “outside” of the strip were retained for analysis with their location adjusted to 25 m inside or outside of the strip, accordingly. For each species recorded, we created point distribution maps with depth data at lowest astronomical tide (LAT) (mean tidal range at Onslow = 1.8 m) (Whiteway, 2009) to present sighting locations. We went on to generate habitat suitability models for humpback and bottlenose dolphins, as they had high enough sample sizes to allow us to do so.

All spatial data processing was carried out using ArcGIS v. 10.1. We generated a series of raster data layers for the study area at 500 × 500 m resolution. This resolution was chosen to keep a fine scale sensitivity when exploring distances from land and islands. Our five static environmental predictors were: depth, seabed slope, distance from the 20 m bathymetric contour, distance from intertidal areas and distance from major water ways (i.e., rivers and creeks). Our six dynamic predictor data layers were sea surface temperature (SST), ocean fronts, and chlorophyll a (chl-a), for both May and October. We assumed that because dolphins are apex predators their distribution is highly influenced by prey availability. In the absence of prey data, we used variables that could act as proxies for prey distribution. For instance, depth affects temperature, light attenuation, salinity, and the presence of benthic primary producers; and associated seabed slope can reflect upwelling areas that attract fish and cephalopods (Cañadas et al., 2002; Hastie et al., 2004; Redfern et al., 2006; Elith and Leathwick, 2009; Garaffo et al., 2011; Blasi and Boitani, 2012; Azzellino et al., 2012). Depth (m LAT) was sourced from the Whiteway (2009) gridded bathymetry data (0.0025 decimal degree resolution). We calculated slope from depth using the “Slope” tool, and distance from the slope at the 20 m contour (an ancient coastal feature, and most prominent bathymetric feature in the study area) using the “Euclidean distance” tool. As others have found, Australian coastal dolphin distribution to be influenced by distance from the mainland coast, islands and river mouths (Corkeron et al., 1997; Parra et al., 2006b; Hunt et al., 2020; Bouchet et al., 2021; Haughey et al., 2021). We calculated these distances at the mean high water mark from Geoscience Australia (2004)’s coastline and Crossman and Li (2015)’s surface hydrology data using the “Euclidean distance” tool. As the resulting “distance to islands” and “distance to the mainland” data were confounded when used separately in preliminary models, we created an alternative single data layer “distance to intertidal areas.” Intertidal areas were derived using the Euclidean Distance tool in GIS and the Geoscience Australia, 2004 “coast” dataset. Intertidal areas refer to areas around the mainland, islands and shoals that are exposed at low tide. Sea surface temperature (SST), ocean fronts (“fronts”) and chl-a are dynamic variables that can reflect seasonal pulses of productivity (Redfern et al., 2006). SST can be highly dynamic in coastal environments, with a seasonal effect on dolphin distribution (Dawson et al., 2008; Becker et al., 2010, 2012; Garaffo et al., 2011). We investigated the influence of SST within and between years because seasonal SST has been suggested as a driver of dolphin distribution in the offshore waters adjacent to our study area (Sleeman et al., 2007), and bottlenose dolphin distribution is influenced by El Niño-Southern Oscillation (ENSO) in southern WA (Sprogis et al., 2017a). We obtained daily SST at 1 km resolution from the JPL MUR MEaSUREs Project (2015), using Marine Geospatial Ecology Tools (MGET) (Roberts et al., 2010) and created a mean raster to represent each survey period. Exploratory analysis showed that SST in May and October 2015 (a strong El Niño year) were anomalous, with lower SST in May compared to other years and a convergence in SST between the 2 months (see Supplementary Material). Based on this, we created a mean of “non-El Niño” year data for May and October of 2012, 2013, and 2014, and kept May and October 2015 separate. Fronts were generated using the “focal statistics” tool calculating standard deviation in SST between grid cells. We sourced chl-a from NASA Goddard Space Flight Center (2014) using MGET (Roberts et al., 2010), while recognising that chl-a data are based on algorithms for the offshore oceanic environment and not validated for coastal waters.

We created distribution models for humpback and bottlenose dolphins using the Maximum Entropy (“MaxEnt”) software (Phillips et al., 2004, 2006), a machine learning method used in a wide range of ecological applications that performs well in its predictive accuracy (Elith et al., 2006). MaxEnt is a technique that uses presence-only data, along with background data, to model the probability of species presence, conditioned by environmental constraints (Phillips et al., 2004; Elith et al., 2011). We treated our data as presence-only because we knew there to be a high level of pseudo-absences, having removed the unidentified dolphin species sightings. MaxEnt generates background data in place of absence data by sampling environmental variables from random points throughout the landscape. MaxEnt’s underlying theory and assumptions are described in detail elsewhere (Phillips et al., 2006; Phillips and Dudik, 2008; Elith et al., 2011). MaxEnt uses an automatic regularisation setting that has been fine-tuned to suit the number of feature types and sample locations (Phillips et al., 2006). As with other forms of machine learning, regularisation is used to prevent models from being overfit by adding a complexity penalty to the loss function. Through regularisation and optimisation, a series of iterations are run which achieves high model gain. High gain reflects that the model is a good fit to the data, which is like the “deviance” goodness-of-fit measure used in general additive models (Phillips et al., 2004). During this process, MaxEnt generates a probability distribution and populates this information into cells of a geographic grid, based on the sampled background data.

We created a static model (i.e., without dynamic variables) for both humpback and bottlenose dolphins, which we used to compare distribution patterns between the species. In addition, we ran May and October, non-El Niño and El Niño models for bottlenose dolphins to investigate seasonal variation and response to dynamic variables. We were unable to create seasonal models for humpback dolphins, as there were fewer data available for sufficient sample sizes after subsetting. We accounted for greater sampling effort in the repeated “Onslow” block (Figure 1B) by including a spatial bias grid (Phillips et al., 2006, 2009). Cells within the repeated block (which had twice the survey effort) were assigned a value of “2” and all other cells were assigned a value of “1.”

Maximum Entropy default settings can fit features (combinations of environmental variables with alternative transformations and model terms) to the model depending on the training data sample size (Phillips et al., 2004; Elith et al., 2006). While MaxEnt default settings perform well, Merow et al. (2013) suggests that user refine these based on their data to achieve clear and simple models and avoid over-fitting. We incrementally tested various combinations, checking the output curves to see that they were not overly complex for the sample sizes available (as reported in the section “Results”). We chose to specify linear, quadratic and product features for our final models, and deselected threshold and hinge features, which allow for steps in the fitted function. We set the prevalence and regularisation at the default value of 0.5, which represents a 50% chance of species occurrence in any cell (Phillips et al., 2004; Elith et al., 2006). The raw sightings data showed that humpback and bottlenose dolphins were broadly distributed throughout the study area. As such, we deemed the default setting of 0.5 to be appropriate for this preliminary study, with each cell having an equal chance that a humpback or bottlenose dolphin could be present. Each model was run using sampled background data set to a maximum of 10,000 random points over a maximum of 500 iterations for 1,000 bootstrapped replicates. This resampled, with replacement, 25% of dolphin data as test data, thereby conserving the data available for training purposes. Each final model was the mean of all replicates.

Our model outputs consist of estimates of habitat suitability, environmental variable contributions and response curves. Mean relative habitat suitability, as a probability distribution from zero to one based on favourable environmental conditions across the study area, was generated in logistic output. Environmental variable contributions (relative to each other and expressed as a percentage) to the model were determined during regularisation. We used non-marginal response curves which are clearer to interpret than marginal response curves as they reflect the influence of the variable itself, and not the interactions it has with other variables. Diagnostic plots of commission and omission rates were used to assess model performance. The Receiver Operating Characteristic (ROC) curve represents sensitivity (true positive rate) against the specificity (fraction of negative instances) (Phillips et al., 2006) and is reflected by the Area Under the Curve (AUC) metric. The AUC metric ranges from 0 to 1 in presence-absence models. An AUC of 1.0 is unable to be reached by MaxEnt models because they are trained on background (a closely related alternative to pseudo-absence) data (Phillips et al., 2006; Elith et al., 2011). The models present a landscape of potential distribution, rather than a “true” distribution. An AUC score of >0.5 suggests that a model performs better than random (i.e., when it has 50% chance of being correct) (Fielding and Bell, 1997). We judged model performance following Hosmer and Lemeshow (1989): <0.5 = none; 0.5–0.7 = poor; 0.7–0.8 = acceptable; 0.8–0.9 = excellent; >0.9 = outstanding. This aligns with Elith (2000)’s assertion that MaxEnt models >0.75 have potentially useful management applications. The omission plots show training omission rate and predicted area as a function of the cumulative threshold of suitable habitat. A good model, in terms of omission, is indicated by the training data trend line falling below the predicted rate (Phillips et al., 2006).

We surveyed 21,156 km of transect line during May (9,116 km), July (2,279 km), October (7,482 km), and December (2,279 km) from 2012 to 2015, with July and December surveyed only in 2012 (Table 1). Fewer transects (i.e., those in the repeat block) were flown in October 2014 and 2015 due to weather constraints. The aerial surveys were conducted in good weather conditions (wind speed mode = 5 knots; cloud cover mode = 0 okta; Beaufort sea state mode = 2; visibility mode = >10 km) (Supplementary Table 1). Glare was arguably the most limiting sighting condition (mode = 3 on a scale of 1–4), largely dependent on the time of day.

A total of 1,169 dolphin sightings (i.e., groups, which may consist of one or more individual animals) were recorded over the full survey period (Table 2 and Figure 2). Bottlenose dolphins were the most frequently encountered species (n = 661; 55.7%; rate of 2.90 sightings/100 km), seen across all water depths measured at LAT (Geoscience Australia, 2004) (min – max LAT = 0–37 m; mean LAT = 10.5 m). A quarter of all sightings (n = 290; 24.8%; rate of 1.27/100 km) were of unidentified dolphin species, also seen across all water depths. Humpback dolphins constituted 16.3% of the sighting data (n = 191; rate of 0.83/100 km), their locations trending toward shallower water depths (min – max LAT = 0–24 m; mean LAT = 10.5 m).

The remaining sightings were of snubfin dolphin (n = 9; 0 certain; all probable), Stenella sp., (n = 16; 4 certain; 12 probable), killer whale (Orcinus orca, n = 1; certain) and false killer whale (Pseudorca crassidens, n = 1; certain) groups (Tables 2, 3). Snubfin dolphin sightings were mainly within Exmouth Gulf, in depths of 5 m or less (min = 0 m LAT; max = 12 m LAT; mode = 5 m LAT). Most (n = 13). Stenella sp. sightings were in the north-eastern portion of the study, as were killer and false killer whale sightings. A group of two adult killer whales were recorded in December near Barrow Island in 14 m LAT depth. A group of false killer whales (n = 50–60 with calves present) was recorded in October, west of Thevenard Island in 9 m LAT water depth.

Model output should always be interpreted with an understanding of the model assumptions and data characteristics on which they are based. Due attention should always be paid to understand model output type, level of certainty and generalisation, spatial resolution, and the accuracy and sample size of both species and predictor variable data on which the model has been based (Redfern et al., 2006; Araújo and Guisan, 2006; Jiménez-Valverde et al., 2008; Elith et al., 2011). Our maps present relative dolphin habitat suitability across the western Pilbara, based on the environmental data available to us to include in the MaxEnt models. They do not present relative density or abundance. MaxEnt maps present a probability distribution across the study area based on favourable environmental conditions (i.e., habitat suitability) (Phillips et al., 2006), in contrast to a relative index of environmental suitability based on species occurrence that can be obtained by presence-absence techniques. Also, they depict a generalised distribution which is broader than the species’ realised niche (Phillips et al., 2006; Araújo and Guisan, 2006; Hirzel and Le Lay, 2008). We also wish to highlight that they cannot be directly compared across models. The AUC score for MaxEnt models will always be penalised because the model is based on background data rather than absence data; it can never reach “1,” and “0” does not indicate absence (Phillips et al., 2006; Elith et al., 2011).

Our models provide new insight into potential humpback and bottlenose dolphin habitat suitability. The performance metrics are not high, and we suggest that the main reason for this is a lack of informative environmental variables across the study area, followed by imperfect sightings data (see section “Study Limitations”). However, at such a broadscale they are still useful for conservation and management planning (Elith, 2000). As outlined by Redfern et al. (2006) habitat models for cetaceans can be useful as a starting point for understanding distribution patterns when knowledge of a species’ ecology is lacking, which we advocate for the coastal dolphins in WA’s Pilbara region.

Across our broad study area, we have found that humpback dolphins are most likely to inhabit intertidal areas and we suggest that these could represent important habitat requiring research and conservation focus. Most areas of high habitat suitability coincided with shallow waters and coral reefs, even when far from the mainland. This aligns with findings from aerial surveys over the Great Barrier Reef Marine Park (Corkeron et al., 1997) and from smaller scale studies in the region (Raudino et al., 2018a,b; Hunt et al., 2020). Humpback dolphins are listed as conservation values that are protected within Ningaloo MP, Muiron Islands MMA and Barrow Island MMA. Further, the North West Cape, situated within Ningaloo MP, has been identified as being of habitat importance for humpback dolphins (Brown et al., 2012; Hunt et al., 2017, 2020). Our model supports these claims but also suggests that these areas are part of a group of important (and likely connected) humpback dolphin habitats, most of which are outside of protected areas. The region supports a mosaic of tropical coral reefs in shallow areas around islands and on limestone platforms, such as Barrow Island Shoals and Mary Ann Passage, and patch reefs such as those within Exmouth Gulf (Wilson, 2013). We speculate that humpback dolphins in this region prey upon reef-associated fish and cephalopods, but this is yet to be confirmed through behavioural and stable isotope studies. Within the Ningaloo MP, humpback dolphins have been observed foraging and resting near reefs (D. Hanf pers. obs.) They may also benefit from positioning themselves with access to both shallow and deep water around offshore islands and North West Cape, for opportunities to hunt prey associated with offshore currents, whilst being able to use shallow water to avoid or escape predators.

Our findings contrast with those from other areas that indicate that the primary driver of humpback dolphin distribution is proximity to river mouths (Parra et al., 2006b; Palmer et al., 2014; Meager et al., 2018) which may reflect the difference in survey scale, and the different geographic area. Northern WA has dissimilar hydrology to other areas of northern Australia (the Kimberley region of WA is more like Northern Territory and Queensland) and is an example of why that EIA should be based on locally collected species data, rather than surrogate data from other populations. As humpback dolphins have been observed foraging in rivers within the Pilbara region (e.g., Fortescue River, D. Hanf pers. obs.), this habitat could be important at a local scale and used intermittently based on tides, which would not be recorded by broadscale aerial surveys.

The relatively high habitat suitability of western Pilbara coral-fringed islands, Mary Ann Passage and Barrow Island Shoals, in addition to the North West Cape (Hunt et al., 2020), support the recent nomination of the Ningaloo Reef to Montebello IMMA. We recommend that these areas be recognised as biologically important areas (BIAs) (Australian Government, 2012a), and thus be given special attention during EIA and conservation planning. This will aid strategic planning and the development of management strategies (e.g., areas of speed restriction, placement of infrastructure, oil spill contingency planning, and climate change resilience planning).

By comparing broadscale models (i.e., western Pilbara region) for humpback and bottlenose dolphins, we may have been able to recognise distribution patterns not evident at a narrower spatial extent. Resource partitioning, where species live in sympatry by taking advantage of different niches to reduce competitive pressure, has been widely observed amongst dolphins (Bearzi, 2005; Parra, 2006; Kiszka et al., 2011; Fernández et al., 2013; Heithaus et al., 2013; Cremer et al., 2018). As top-order predators, co-existence of sympatric dolphin species is likely to be enabled by spatial-temporal differences in prey preferences, diet composition and foraging strategies (Berta et al., 2015). We found bottlenose dolphin habitat suitability to be high near the slope at the 20 m isobath, whereas humpback dolphin habitat suitability was highest around intertidal areas. Similar spatial separation between the two genera exists elsewhere (Stensland et al., 2006; Palmer et al., 2014). Humpback and snubfin dolphins also exhibit resource partitioning in areas where they co-occur (Parra, 2006). Separation could reflect differences in feeding strategies and risk-taking behaviour, with bottlenose dolphins exhibiting a high level of plasticity throughout their range (Cribb et al., 2013), whereas humpback dolphins have a higher site fidelity (Brown et al., 2016; Meager et al., 2018). Capture-recapture studies at the North West Cape demonstrate that bottlenose dolphins display a low site fidelity, often venturing outside of the study area while humpback dolphins show a high site fidelity with high recapture rates (Hunt et al., 2017; Haughey et al., 2020). Slopes are often important foraging areas because prey is usually in high abundance (Hastie et al., 2004). The 20 m isobath marks the end of the inner shelf after which it gradually descends to 100 m. These waters between 20 and 100 m are deep are open to oceanic currents which transport zooplankton and finfish. Furthermore, cephalopods in high abundance (Jackson et al., 2008), representing an alternative foraging area which bottlenose dolphins may take advantage of to reduce competition with humpback dolphins.

Despite partitioning across the region, there were overlaps of suitable habitat for both species at certain locations. There could be a dietary divergence between the two species at these habitats, which has been found to occur with other dolphin species (Bearzi, 2005). More likely though, it could reflect plentiful prey availability at these sites. These areas of overlap may support higher productivity, greater prey species biomass or diversity at times. For instance, mixed species groups of humpback and bottlenose dolphin groups have been recorded at the North West Cape, where offshore and inshore waters are channelled and mixed (Brown et al., 2012; Hunt et al., 2017). This is plausible as both species are “opportunistic-generalist feeders,” consuming a wide range of benthic and pelagic fish and cephalopod species (Amir et al., 2005; Parra and Jedensjö, 2014; McCluskey et al., 2016; Smith and Sprogis, 2016; Sprogis et al., 2017b). As there are many unresolved questions, targeted observational studies are required to understand foraging strategies and the mechanisms underlying mixed species group formation and behaviour (Syme et al., 2021).

We found that bottlenose dolphin distribution is influenced by dynamic variables and theorise that response to warmer SST in May represents a shift in prey availability driven by the Leeuwin Current. A similar pattern for bottlenose dolphins was reported by Sleeman et al. (2007), with highest dolphin abundance between March and May in the area immediately offshore to our study area. This could also coincide with Sprogis et al. (2017a)’s finding that bottlenose dolphin distribution shifted offshore, toward the Leeuwin Current, during an El Nino period when the current is suppressed. Such plasticity and responsiveness to changes in environmental conditions may also be a driver of low bottlenose dolphin site fidelity at the North West Cape (Haughey et al., 2020). The response of bottlenose dolphins to SST was largely consistent between non-El Nino and El Nino years, but the mapped output indicates a more marked variation in habitat suitability which could be further explored. It is important to emphasise that the models present general patterns at a broadscale level. Bottlenose dolphins were found throughout the study area, yet the models suggest a potential preference to deeper areas which we suggest is driven by offshore oceanography that influences prey availability. Finer scale studies may find variation within the population. For instance, Sprogis et al. (2018) found female bottlenose dolphins to prefer deeper water depths than males during summer and spring. These hypotheses could be investigated through targeted in situ studies involving the collection of behavioural data and biopsy samples for stable isotope analyses.

Our ability to consider seasonal and annual variability highlights the strength of multi-year studies and cautions on single-year studies as there may be various confounding issues that can affect animal distribution and associated conclusions regarding presence and habitat. Longer-term studies (i.e., decadal) in the future would be preferable, to capture the effects of climatic events such as the La Nina event of 2011, which contributed to a marine heatwave and caused large-scale fish and benthic primary producer die-off, as well as the displacement of tropical species (Pearce and Feng, 2013). Sprogis et al. (2017a) were able to clearly detect abundance decline in their study area during an El Nino winter. Cyclones and associated flooding, prominent in the Pilbara region between October and March, are also set to increase with climate change (Abbs, 2012). As flood events effect coastal dolphin distribution (Cagnazzi et al., 2020), long-term data (ideally and decadal) are needed to build predictive models.

Snubfin dolphin occurrence in the Pilbara region is a subject of some conjecture and our study has not been able to confirm their presence. This location is approximately 650 km south-west of their recognised range (IUCN SSC Cetacean Specialist Group, 2017; see Supplementary Material). Occasional and compelling anecdotal reports of snubfin dolphin presence in the Pilbara region exist since at least 1995 when a Department of Conservation and Land Management Wildlife Officer P. Mawson filed an official memorandum describing his certainty in sighting of two or three individuals at the North West Cape. Allen et al. (2012) collated several reports including those made by seasoned marine mammal scientists. Nine snubfin dolphin sightings were recorded by observers during our surveys. However, none of these were made with certainty in species identification and there were no clear recaptures between front and rear observers. Misidentification between dugongs and snubfin dolphins can sometimes occur, particularly during aerial surveys platforms (Dunshea et al., 2020). It is worth noting that no opportunistic snubfin dolphin sightings have been made by marine mammal researchers conducting other boat-based studies in Exmouth Gulf and around NW Cape in recent times, which may be an indicator of their scarcity in the area (i.e., Hunt et al., 2017; Raudino et al., 2018a; Bejder et al., 2019; Haughey et al., 2020; Sprogis et al., 2020a,b). We suggest that vagrant individuals may occasionally occur within our study area. Parts of Exmouth Gulf (where most of our sightings were made), share similar habitat characteristics to Roebuck Bay, i.e., large shallow embayment fringed by mangroves and intertidal salt/mud flats connected to deep oceanic waters with a complex of seagrass, reef and sponge beds forming a rich prey environment (Australian Government, 2012a). Further studies are needed to investigate the species’ full extent of occurrence in Western Australia, including whether small discrete populations exist within the Pilbara region. Such information is required for their full assessment as a threatened species under the EPBC Act and is a critical information gap in marine spatial planning.

Pelagic species are likely to move into our study area from adjacent deeper waters where key ecological features (narrow shelf, canyons and the Exmouth plateau) facilitate upwelling and enhance local productivity (Australian Government, 2012a; Wilson, 2013). Stenella sp. sightings may include pantropical spotted (Stenella attenuata), and dwarf spinner (Stenella longirostris) dolphins, as North West Cape stranding records include the former (Groom and Coughran, 2012) and the latter has been identified from north WA biopsy samples (Leslie and Morin, 2018). Our killer whale sighting in December 2012 is one of the rare, documented sightings of this species in the region in summer outside of the humpback whale season when most killer whale research has been undertaken (e.g., Pitman et al., 2015). This suggests that the local population may not have a feeding niche that is as specialised as previously thought, or that resource partitioning exists with separate summer and winter groups feeding on different prey at different times (J. Totterdell pers.com.). We also recorded a large group of false killer whales, a “Near Threatened” species (IUCN, 2020), for which a paucity of data exists in Australia as well as globally (Woinarski et al., 2014). This species is generally regarded as a migratory, tropical to subtropical, deep-water, oceanic species that is sometimes encountered in neritic areas (Baird et al., 2012; Zaeschmar et al., 2014). Palmer et al. (2017) demonstrated that false killer whales may be regular visitors to coastal and estuarine environments in north Australia and that further research regarding their habitat use is warranted. As we only recorded the species once during 10 surveys, we suggest that they were likely encountered while undertaking broadscale movements and are more likely to be pelagic visitors to the coastal environment.

The major shortfall of our study was the uncertainty in species identification, as a result of observers focussing on searching for dugongs as their primary target, a limited “search” time (intrinsic to all aerial surveys and particularly passing mode surveys) and the similarity in appearance of humpback and bottlenose dolphin species from aircraft. These limitations reduced the data available for modelling by 25% and constrained our choice of modelling method. However, as surveys were conducted over multiple years, we were able to pool data for “static” and non-El Niño models, which assumed that relationships did not vary across years. The bottlenose dolphin El Niño models should be interpreted with caution as they encapsulate only one season’s worth of data. The low AUCs, weak patterns in environmental contributions (apart from distance from the 20 m slope for bottlenose), and non-monotonic response curves likely reflect noise in the model. Finer scale patterns than what we have presented are likely to exist across the area, with different levels of contribution by and interaction between environmental variables. Ideally, we would have a larger dataset which we could further partition to investigate this [e.g., using sampling instances as outlined by Matthiopoulos et al. (2011) in their generalised functional response approach]. This could be achieved through the simultaneous acquisition of photographic data that would allow species identification verification.

Another limitation in our study was the lack of data on benthic primary producers and prey distributions. These predictor variables would likely provide greater information to the models which may result in clearer patterns than models based on proxies. The negative response to chl-a could be a result of the ocean colour and raises several questions. It could suggest that the data are not a good proxy for productivity in the area, or that the distribution of these dolphins is not directly driven by primary productivity. Alternatively, the data could be masked or biassed by other factors, such as turbid terrestrial run off which is characteristic of nearshore Pilbara waters. Furthermore, it could be that there is a lag between productivity and dolphin response. Ultimately, we did not find chl-a as represented by ocean colour to be useful in the models.