Introduction

Unprovoked shark attacks (or shark bites) are dramatic events that command considerable public attention, and can increase tension between policy mandates for the protection of public safety and marine animal conservation (e.g. Neff 2012; Hazin and Afonso 2014; Gibbs and Warren 2015). In this context, unprovoked shark attacks are defined as those inflicted upon humans who have not attempted to deliberately interact with a shark (e.g. by feeding, touching, pursuing or capturing the animal), or those who are not otherwise engaged in activities likely to attract sharks (e.g. spearfishing or cleaning fish) (West 2011; McPhee 2014; Ricci et al. 2016). The word ‘attack’ may not always accurately represent the motivations underpinning these interactions (Neff and Heuter 2013), but is used here for consistency. Both globally and in Australia, over half of all unprovoked attacks are attributed to white (Carcharodon carcharias), bull (Carcharhinus leucas) and tiger (Galeocerdo cuvier) sharks (West 2011; McPhee 2014). These three species are also implicated in almost all fatal attacks (West 2011; McPhee 2014).

Although shark attacks remain rare, the total number of global attacks increased 3-fold between 1982 and 2011 (McPhee 2014). Some of this increase reflects growing human population and greater overall participation in aquatic recreation (West 2011). The internet and widespread use of smartphones have also facilitated public awareness and reporting of shark attacks (West 2011). However, even after allowing for these factors, 30 years of data extracted from the Global Shark Attack File indicates that the proportional increase in attacks has outstripped population growth, suggesting a real, albeit small, increase (McPhee 2014). The mechanisms underpinning this rise are unclear, but may involve increased abundance of some prey species (especially marine mammals) (McPhee 2014; Chapman and McPhee 2016), changing human recreational behaviours (West 2011) and sharks’ shifting habitat selection in marine ecosystems undergoing anthropogenic change (Hazin et al. 2008; Chapman and McPhee 2016). Greater shark abundance has also been raised as a possible cause, and although some local increases have been reported (e.g. Carlson et al. 2012; Froeschke et al. 2013), these appear insufficient to account for the global rise in attacks (West 2011; McPhee 2014). Regardless of the underlying causes, spatially and temporally clustered attacks, such as those along the coastlines of Recife (Hazin et al. 2008, 2013), Reunion Island (Blaison et al. 2015; Lemahieu et al. 2017), North and South Carolina (Amin et al. 2015) and off the Australian states of New South Wales (NSW; Neff 2012; Pepin-Neff and Wynter 2017) and Western Australia (WA; Gibbs and Warren 2015), usually result in public demands for risk mitigation (Hazin et al. 2008; Cliff and Dudley 2011; Hazin and Afonso 2014).

Shark attack clusters led to the establishment of the three longest-standing shark control programs globally: the NSW Shark Meshing Program (established 1937); the KwaZulu–Natal shark control program in South Africa (established 1952); and the Queensland Shark Control Program (established 1962) (Cliff and Dudley 2011). Initially, these programs focused solely on localised reduction of shark numbers using mesh nets and a system of large baited hooks suspended from surface floats, known as ‘drum lines’, to catch and kill sharks (Cliff and Dudley 2011; Reid et al. 2011; Sumpton et al. 2011). Gradually, increasing public concern over the programs’ impacts upon both by-catch (typically non-target elasmobranchs, marine mammals and reptiles) and target species has seen governments searching for measures that improve protection for swimmers while minimising the impacts on sharks and other marine wildlife (Cliff and Dudley 2011; Hazin and Afonso 2014; Engelbrecht et al. 2017).

Some of the most recent alternative measures are based on modifications of existing technology. For example, the SMART (Shark Management Alert in Real Time) drumlines developed in Reunion Island incorporate satellite communications that automatically alert operators to a hooked animal via email, phone call and text message, enabling rapid release and/or relocation, usually after acoustic or satellite tagging (McPhee and Blount 2015). Other bather protection measures include: (1) physical barriers that aim to provide an impassable obstacle that separates sharks from people; (2) visual barriers, such as bubble curtains, that are not physically impassable but present aversive visual stimuli; and (3) barriers that aim to exploit sharks’ aversion to particular electrical and electromagnetic stimuli (O’Connell et al. 2014a, 2014b, 2014c; McPhee and Blount 2015). Combined deterrents are also under investigation, such as the SharkSafe Barrier (Sharksafe Barrier^TM, Stellenbosch, South Africa;  https://www.sharksafesolution.com/, accessed November 2019), which combines an electromagnetic deterrent with PVC piping that mimics dense kelp beds, creating a threshold that large sharks may be unwilling to cross (McPhee and Blount 2015; O’Connell et al. 2018). Personal deterrents, designed to be worn or carried by an individual swimmer or surfer, are not considered in detail here, but are reviewed by O’Connell et al. (2014c) and Hart and Collin (2015). Most of the above measures have reported some degree of success, but none represents a complete solution at either the individual or whole-of-beach scale. For example, physical barriers are susceptible to damage during rough seas (Cliff and Dudley 2011), and the species-specific nature of elasmobranch responses to electromagnetic stimuli may preclude universal application of electrical and electromagnetic deterrents (O’Connell et al. 2011, 2014c; Hart and Collin 2015).

The difficulty of identifying and implementing broadly applicable and cost-effective shark deterrents has given rise to measures aimed at shark detection (with subsequent warnings to swimmers) rather than deterrence. Detection approaches include trained observers working from elevated coastal vantage points (Engelbrecht et al. 2017), automated sonar systems (McPhee and Blount 2015; Parsons et al. 2015), acoustic tagging and monitoring programs (McAuley et al. 2016) and manned aerial patrols using fixed-wing aircraft or helicopters (McPhee and Blount 2015; Robbins et al. 2014).

Methodologically, aerial shark patrols represent a specialisation of the aerial surveys used to estimate wildlife distribution and abundance (e.g. Fleming and Tracey 2008; Rowat et al. 2009; O’Donoghue et al. 2010; Poole et al. 2013; Kleen and Breland 2014; Fuentes et al. 2015). Although aerial shark patrols aim simply to locate sharks and advise the public, rather than attempting formal estimates of abundance or distribution, high detection likelihoods and accuracy are important if the patrols are to effectively protect bathers (Robbins et al. 2014). Unfortunately, detecting sharks via aerial surveys can be difficult, because unlike air-breathing aquatic animals that spend a considerable portion of time at the water’s surface, and therefore can be readily sighted from the air, white, bull and tiger sharks spend most of their time completely submerged (Holland et al. 1999; Dewar et al. 2004; Bonfil et al. 2005; Carlson et al. 2010). Aerial surveys have been used for locating sharks, particularly for species inhabiting shallow, clear water (Kessel et al. 2013), or when experimental designs allow flights to be undertaken only during periods of optimal visibility (Dicken and Booth 2013). However, these studies have not included formal estimates of detection likelihood, instead assuming that favourable conditions would facilitate detection of a high, but unspecified, proportion of the sharks present. Furthermore, manned aerial shark patrols are flown in a range of sea states and environmental conditions that may reduce detection rates by an unknown amount (Robbins et al. 2014). Studies quantifying detection likelihood in aerial shark patrols are scarce, but indicate low sighting rates. For example, Robbins et al. (2014) reported sighting rates of 12.5% and 17.1% for analogue sharks seen by observers from fixed-wing aircraft and helicopters, respectively. These rates suggest that, although aerial shark patrols may provide some psychological reassurance to bathers (Neff 2012), improved sighting rates and periods of surveillance ‘cover’ are required before they can be considered an effective bather protection measure (Robbins et al. 2014).

One option to improve the efficacy of aerial shark detection is to use drones, also known as remotely piloted aircraft systems (RPAS) or unmanned aerial vehicles (UAV), which are available in a variety of fixed-wing and multi-rotor configurations (Colefax et al. 2018). Drones increasingly present an alternative to manned aircraft for a variety of wildlife survey and detection tasks (Kudo et al. 2012; Martin et al. 2012; Linchant et al. 2015; Evans et al. 2015; Kiszka et al. 2016; Colefax et al. 2019; Kelaher et al. 2019). In the context of shark patrols, drones offer several potential advantages over manned aircraft because they can operate at individual beaches, and fly constantly at lower airspeeds and altitudes than are physically or legally possible for manned aircraft (Kudo et al. 2012; Robbins et al. 2014, but see Linchant et al. (2015) for possible exceptions). Drones can also carry polarising filters to facilitate viewing in real time, or hyperspectral sensors for digitally enhancing images for post processing so that we can ‘see’ further below the water’s surface than human observers in manned aircraft (Stein et al. 1999; Schoonmaker et al. 2011). Furthermore, drones are more cost-effective to operate than manned aircraft for small-scale surveys, produce fewer environmental impacts and pose a lower risk to both operators and bystanders (Kudo et al. 2012; Martin et al. 2012; Linchant et al. 2015).

Although recent technological advances have seen drones used in many wildlife survey applications, their capacity to detect predatory sharks remains untested. The primary aim of the present study is to quantify the effectiveness of detecting submerged shark analogues using drones across a range of environmental conditions representative of those likely to be encountered along the eastern Australian coastline. Beyond the realm of bather protection, the research also aims to determine the utility of drones as a shark survey method.

Materials and methods

Shark analogues

For the experiment, two-dimensional shark analogues were used within the survey area (following the procedures of Robbins et al. 2014). Although the use of live sharks would be preferred, the use of analogue or decoy animals is well established in the aerial survey literature as an effective means of rigorously testing detection efficiency, because the number and positioning of analogues is known to researchers, but not to the observers or pilot (e.g. Jones et al. 2006; Koski et al. 2009; Robbins et al. 2014). Each analogue in the experiment measured 2430 mm (total length) by 1190 mm (pectoral fin width), and was constructed of 11-mm marine plywood. Both sides of the analogues were painted grey to replicate large oceanic sharks (Robbins et al. 2014), with two shaped as white sharks and two as hammerhead sharks (Sphyrna sp.) (Fig. 1).

Fig. 1. 

Images of hammerhead shark and white shark analogues across different water turbidity (m) and substrates (sand and rock) in comparison with a live white-spotted guitarfish (Rhynchobatus australiae) seen during the trials. White square indicates the location of the analogue.

One to four analogues were deployed from a vessel at a given location 2 m below the surface in 2.0–6.3 m water depth. This depth was chosen because poor detection rates have previously been reported for shark analogues positioned more than 2.6 m deep (Robbins et al. 2014). The actual water depth is important because the change in colour or background may influence the ability to locate an analogue. To enable horizontal orientation in the water, each analogue was attached at four points on the underside using wire cables terminating at a single metal ring attached to a rope and anchor. For retrieval purposes, a 50-mm float (not visible to the pilot) was attached to the analogue’s upper surface with clear monofilament line. All analogues were positioned within 70 m of the back line of the surf break to ensure they were visible within the strip width of the drone camera.

Statistical analysis

The analysis accounted for the structure of the experimental design when examining the influence of environmental factors on sighting rates. The model used in the analysis accommodated for the dependence in the response variable through unmitigated, shared covariates; this reduces the likelihood of false associations caused by confounding environmental variables with plot factors. For the formation of the model, we used the approach of Smith and Cullis (2019), thereby defining the plot and treatment factors, their associated structure and the design function for the randomisation of treatments. This provided the terms, which were included in the generalised linear mixed model (GLMM), as well as determining the status (fixed or random) of all terms in the model. Observations for analogue sightings are not independent because their deployment remained the same for both flight directions (north and south); thus, we used a GLMM with random terms to allow for dependence between observations. A logistic GLMM with a logit link was fitted to the binary response of detectability, coded as 0 (for non-detection) and 1 (for detection). The pilot and laboratory observer data were analysed together, as is typical with double-observer data in the field of ecology (e.g. Bernard et al. 2013; Strobel and Butler 2014).

The experiment can be considered a multi-phase experiment, where Phase I involved the construction of the shark analogues, with two analogues for each species, and Phase II was the deployment of analogues. The plot factor for Phase I was the shark ‘analogue’ (four levels), and the treatment factor was ‘species’ (two levels). The treatment structure was 1 + species, where 1 represents the overall mean, and the plot structure was ‘analogue’. For Phase II, the set of plot factors were ‘Flight’ (two levels), ‘Week’ (three levels), ‘Day’ (three levels), ‘Time’ (three levels), ‘Viewing’ (two levels) and ‘Location’ (four levels). We further defined an additional plot factor ‘Set’, which was a factor for levels that were the unique combinations of the three plot factors ‘Week’, ‘Day’ and ‘Time’. For each level of ‘Set’, analogues were deployed to each of the randomly chosen locations within the pre-determined grid. The number of analogues that were deployed varied from zero to four. This approach was adopted to mask the true number of analogues deployed for any given set from the pilot and the observer. There were three sets that had no analogues deployed. These sets were discarded from the analysis of detectability for analogues, but used in the overall data. The factor ‘Viewing’, with two levels, corresponds to the two viewings of the footage at each deployment by the pilot and then subsequently by the laboratory observer. Finally, ‘Flight’ has two levels, north and south, corresponding to the flight direction. The plot structure is given by:

where Flight : Week : Day : Time : Location : Viewing indexes the observational units.

These treatment and plot structures lead to what is termed the working GLMM, where all model terms in the plot structure are assumed to be random terms. Terms in the treatment structure were then fitted as fixed terms, except for those terms in the plot structure of Phase I (i.e. Analogue), which were fitted as random.

The working GLMM was then extended to the full set of environmental factors and variates (Table 1). To assist with the notation, we refer to ‘X’ as the matrix of covariates and dummy factors associated with the environmental factors and variates listed in Table 1. The GLMM was extended by including X, as well as the interaction of species and observer with X. Terms were then removed from this full GLMM using backward elimination, with respect to marginality. That is, all interaction terms were considered before examination of the main effects of X.

Table 1. 

Summary of environmental factors and variates considered in the generalised linear mixed model for detecting shark analogues

Where cells in the Factor column are False, range is provided, where they are True, levels are provided

All analyses were conducted in the R (R Development Core Team 2015) package ASReml-R, which fits GLMMs using the approach of Breslow and Clayton (1993). Inference for fixed effects was conducted using approximate Wald-type pivots (Butler et al. 2009).

Results

Twenty-seven flights (mean ± s.d., 12:50 ± 1:55 min : seconds) across 9 days were completed over 3 weeks. The 54 shark analogues (23 hammerhead and 31 white sharks) were set in water depths between 2.0 and 6.3 m (mean ± s.d., 3.6 ± 1.1 m), over sand (39 deployments) and reef (15 deployments) habitats. Only one flight at 0900 hours was delayed due to rain, while the rest went as scheduled across the variety (mean ± s.d., range) of environmental conditions: wind speed (14.6 ± 7.5 km h⁻¹, 2.0–30.0 km h⁻¹); turbidity (2.1 ± 1.1 m, 0.4–6.4 m); humidity (67.1 ± 6.8%, 52–80.0%); barometric pressure (1014 ± 3.3 hPa, 1008–1021 hPa); cloud cover (3.7 ± 3.1, 0–8); air temperatures (surface: 25.5 ± 1.2°C, 23.5–28.4°C); glare (1.2 ± 0.7, 0.0–3.0); seas (0.7 ± 0.3 m, 0.4–1.4 m); swells (1.8 ± 0.2 m, 1.4–2.5 m); and sea state (2.3 ± 1.2 Bf, 1.0–5.0 Bf).

Because a given analogue had two chances of being detected (north- and south-bound flight segments), there was a total of 108 possible analogue detections, of which 67 were not sighted by either the pilot or observer. In total, 31 analogues (29%) were sighted by both the pilot and laboratory observer, and an additional 10 analogues (9%) were detected by the laboratory observer but not the pilot. No analogues were detected by the pilot and not the observer. Fig. 2 presents a jittered scatter plot of the analogues detected and missed by the pilot and laboratory observer against water turbidity (m), which indicates that the detection probability is very low in conditions where the water visibility is less than 1.5 m. Specifically, of the 31 analogue sightings detected by both the pilot and observer, only two of the sightings for each group occurred when the water visibility was less than 1.5 m (Fig. 2). Deployments when the water visibility was less than 1.5 m provide no insight into additional factors that may affect detection. The remaining analyses were therefore conducted on those deployments for which water visibility was greater than 1.5 m (n = 78). Of these 78 shark analogue deployments in water visibility >1.5 m, 29 (38%) were detected by the pilot and 39 (50%) from post analysis by the observer. There were no false positives observed where either the pilot or laboratory observer recorded a shark analogue when there were none deployed.

Fig. 2. 

Scatter plot of jittered shark-sighting rates from the pilot and laboratory observer (hollow circle, hammerhead shark analogue; black circle, white shark analogue) against water turbidity (m). Dashed line shows the 1.5 m-turbidity level and analogues that were excluded from the formal analysis of detection probability.

Detectability

For the GLMMs exploring the influence of environmental variables relating to probability of analogue detection by the pilot or laboratory-based observer, all treatment terms were non-significant and dropped from the final model except for the main effect of water depth (Table 2). The term ‘species’ was retained because of the treatment structure, although there was no significant difference (P = 0.50) in detection probability between hammerhead (0.34 ± 0.64) and white shark (0.47 ± 0.64) analogues (Table 2).

Table 2. 

Summary of conditional Wald test statistical model for detection probability, and associated probabilities of significance, based on an asymptotic Chi-squared reference distribution

d.f., degrees of freedom; F.con, F-statistic; pchisq, cumulative Chi-squared reference distribution

There was a significant difference (P = 0.04; Table 2) in detection probability between the pilot (0.30 ± 0.61) and the laboratory observer (0.51 ± 0.61), with a log odds ratio of 0.916. The odds of detecting an analogue were 2.5 times greater for the laboratory observer than for the pilot. Water depth was also significant (P = 0.04), with a log odds ratio of −0.874. Specifically, an increase in water depth by 1 m results in a 58% reduction in the odds of detection. The predicted detection drops below 50% when the water depth exceeds 3.5 m (Fig. 3). However, there were three detections by both the pilot and observer in water depth >5 m and on both flight directions. These detections involved the same white shark analogue, and occurred over a sandy substrate on a windless day with smooth sea conditions.

Fig. 3. 

The jittered binary outcomes (top, seen; bottom, not seen) with the approximate 95% pointwise coverage interval for the predicted probability for detection (shaded area) for shark sighting rates against water depth (m) when water turbidity was >1.5 m depth.

The estimates on the logit scale for the between-flight (Week : Day : Time : Location) and within-flight variance (Flight : Week : Day : Time : Location) were 0.65 and 3.38, respectively. An approximate measure of dependence between flights (0.84) was given by the within-flight variance divided by the sum of the within- and between-flight variance. There was considerable correlation in detection probabilities between flight directions, which was expected given that the location of analogue deployments remained the same for flights within a set. However, the random component of the GLMM fitted accounted for this dependence.

Discussion

The present study has quantified the diverse range of coastal conditions under which drones can operate, demonstrating their utility for detecting objects in the water (in this case shark analogues), particularly in conditions of reasonable water visibility and depth. Depending on the environmental conditions, drones can potentially be used to spot sharks or other marine fauna along coastal beaches, and could supplement existing aerial survey methods (Colefax et al. 2019). Furthermore, drones would provide a cheaper and more versatile platform than manned aircraft for beach authorities and various user groups for localised coastal areas (Colefax et al. 2018). The potential benefits that drones can offer for bather protection, and specifically shark attack mitigation, are discussed by considering the underlying mechanisms that may affect their use.

Water turbidity, which was high throughout the study, had a strong negative relationship with detection probability. Specifically, when the depth of sightability (governed by turbidity) was shallower (<1.5 m) than the depth of the shark analogues (2 m), detection probability was very low, with neither the pilot nor laboratory observer able to identify the shark analogues. Turbidity is recognised as a major factor affecting data reliability in aerial surveys because it decreases the depth in the water column at which an animal can be sighted, thereby decreasing detection probability proportionately to dive behaviour (Pollock et al. 2006; Hodgson et al. 2013; Hagihara et al. 2014; Fuentes et al. 2015). In clear water, the pilot and observer detected analogues, but detection probability across the survey was skewed because water clarity was poor for much of the study.

Water depth had a significant effect on detection probability, particularly at depths >3.5 m. This is likely because light attenuation changes with water depth, altering contrast between the grey shark analogue and the substrate. Similarly, sighting rates for dugong (Dugong dugon) decreased with water depth and turbidity (Pollock et al. 2006; Hodgson et al. 2013). In deeper water, wavelengths of light are either scattered or absorbed before reflecting off the substrate, creating a darker ocean colour and masking the grey shark-like colour of the analogues (Bloom et al. 2019). Conversely, in shallower water, where the depth of sightability is greater than the water depth, light is reflected from the sand or reef substrate, giving the water a lighter appearance and allowing better contrast with the silhouette. Depth-related differences in light attenuation and reflection suggest that multispectral and hyperspectral sensors might allow better image contrasting through frequency band selection (Colefax et al. 2018).

Although glare (as sea surface reflectance) did not significantly affect sighting rates (through perception biasing; Pollock et al. 2006), glare effects were detectable, particularly around midday and at times when sea surface conditions produced wavelets. The effects of sea surface reflectance can typically be reduced in manned aerial surveys by polarisation (Zhang et al. 2017), such as having observers wear polarised sunglasses (Alves et al. 2013). In drone-based aerial surveys, the equivalent is achieved by applying appropriate circular polarising filters for the light intensity at a given location, as was done in this experiment. Further measures can also be taken to potentially reduce sea surface reflection, such as altering the drone’s orientation and making minor adjustments to the gimbal angle with regard to the sun’s position, particularly when near its zenith (Zhang et al. 2017). However, these measures may have consequences for the consistency of flight parameters (e.g. transect width), and perhaps depth of sightability due to altered sensor viewing angles.

The present study did not identify significant effects of wind and sea state on sighting rates. This counterintuitive result may be an artefact, reflecting the model’s inability to distinguish wind and sea-state effects from the high turbidity levels present throughout much of the study. Alternatively, strong winds and rough seas may actually impair marine fauna sighting rates less for drones than for manned aircraft, because drones are flying substantially slower and lower. Although other drone-based marine surveys have indicated likely declines in detection probability associated with increasing sea states (Koski et al. 2009; Hodgson et al. 2013), it is likely that the comparatively lower altitudes, smaller search area and slower speeds of drone-based surveys would allow longer scanning time over an area for the observer and pilot, and buffer some of the adverse effects on detection rates in comparison to manned aircraft (Colefax et al. 2018). To conclusively determine if detection probability from drone-based shark surveillance is less affected by environmental variables, including wind and sea state, an empirical investigation directly comparing it with manned aircraft is required.

Flights throughout the study were exposed to a range of environmental conditions, including fresh wind strengths of 30 km h⁻¹ and sea states up to 5 Bf. Manned aircraft can also conduct aerial shark surveillance during these conditions, but generally avoid doing so because sighting probability is significantly impaired (Rowat et al. 2009; Robbins et al. 2014). Manned aerial surveys are costly (∼AU$1200 per hour), and the reduced sighting probabilities characteristic of high winds and rough seas often result in decisions to restrict operations to sea states less than 3 Bf and winds <30 km h⁻¹) (Rowat et al. 2009; Kleen and Breland 2014; Fuentes et al. 2015). However, the array of real-time tracking buoys that detect dangerous sharks along 21 coastal beaches in NSW has shown that they are present across all environmental conditions (Paul Butcher, NSW DPI, unpubl. data).

The small but significant difference in sighting rates between the pilot and the observer indicates a degree of perception biasing in field-based analogue detections. Pilot perception error was noticeable between water depths of 3.5 and 4.5 m, where the observer detected more shark analogues than the pilot did. These differences in sighting rates usually manifested in conditions of low (but still >1.5 m) water clarity and over deeper water, so the shark analogues had only very faint background contrasting for the pilot to detect in real time. The results from the present study suggest that in shallower or clearer conditions, there would likely be negligible difference in detection probability between the pilot and observer. In scenarios where detection becomes increasingly difficult, perception biases in field detections may be explained by: (1) attention of the pilot spread across multiple tasks; (2) sun glare on the telemetry screen; (3) the telemetry screen size and/or decreased resolution compared with post-processing; or (4) the observer’s ability to pause, slow down or replay transect segments for sighting determinations.

Australian civil aviation regulations stipulate that drones must remain within the pilot’s line-of-sight (unless under special exemption), and that the drone’s condition and surrounding airspace must be monitored for hazards and air traffic throughout the flight. The multi-tasking nature of conducting these tasks while observing the video feed may have reduced the pilot’s detection performance (Adler and Benbunan-Fich 2012). One or two observers might significantly increase the probability of detecting objects on the screen. Additionally, although shading was used on the top and sides of the telemetry screen to reduce sun glare and enable a clearer picture, viewing drone footage on a larger, higher resolution screen in a controlled environment would likely facilitate object detection. Measures including the use of high resolution ‘first-person-view’ goggles (such as DJI goggles), a large protected screen showing the live telemetry to an observer, or a brighter screen (i.e. CrystalSky Ultra, DJI, China) could improve real-time object detection by drone pilots in the field. Alternatively, advancements in automatic software recognition may assist the pilot in identifying otherwise missed detections. Computer-based object recognition is a research focus in this field – and will inevitably lead to increased post-processing efficiency of drone-captured imagery – but is still largely in developmental stages, particularly in heterogeneous environments (Chabot and Francis 2016; Seymour et al. 2017; Gu et al. 2018; Saqib et al. 2018).

Despite all scheduled drone flights proceeding throughout the 3-week survey period, the presence of rain did delay one flight as the drone used in this study, like most drones, was not rated to fly in wet conditions. Although water-resistant drones exist, they generally have much shorter ranges and achieve shorter flight durations than equivalent non-water-resistant variants (Fiori et al. 2017). Marine fauna surveys from manned aircraft technically can operate during rain, but avoid doing so because it impairs the sightability of marine fauna (Pollock et al. 2006; Rowat et al. 2009; Kleen and Breland 2014). Similarly, rainy conditions would likely impair faunal sightability even if water-resistant drones were used. Precipitation is therefore potentially the main limitation on drone-based shark surveillance, because it can either force surveillance operations to cease or impede faunal sightings (Colefax et al. 2019).

The use of shark analogues to determine sighting rates in the present study provides an indication of performance, but, for a given shark size and water depth, a moving shark would undoubtedly be detected more easily than a still replica. The difference in detection probability between a shark analogue and live shark was assumed (but not tested) to be negligible from manned aircraft (Robbins et al. 2014). However, due to their low speed and narrow strip width, drones survey areas more intensively, resulting in a shark being captured within the sensor for a longer period (often >5 s). Movement may therefore be more important as a distinguishing factor for faunal detection using aerial surveys than previously assumed, particularly for surveys using drones.

The applicability of drones for aerial shark surveillance will depend partly on location- and jurisdiction-specific surveillance requirements. Demand is growing for reliable shark attack mitigation with minimal impact on marine animals (Cliff and Dudley 2011; Hazin and Afonso 2014; Meeuwig and Ferreira 2014).

Conclusions

Our results suggest that drones are a potential platform for non-invasive shark surveillance, and that further research into refining operational procedures and incorporating new technologies will allow for improvements in detection reliability across a broader range of operating conditions. In time, drones may offer better detection reliability than manned aircraft, with developments in real-time object recognition software likely to further improve detection probabilities. Although adverse environmental conditions impacted sighting rates, these conditions also likely correspond with few-to-no water users, arguably reducing the need for shark surveillance (de Freitas 2015). An empirical investigation to determine the effects of weather variables on the presence of water users could, in future, better define the environmental conditions under which reliable sighting rates are imperative. Under the range of environmental conditions seen during the present study, shark detection will be optimised: when the water turbidity state allows vision to a depth of at least 1.5 m; in shallow water depths (<3.5 m); and when a dedicated observer, rather than the drone pilot, is responsible for shark detection. This study has shown that drones can offer an alternative shark detection tool that could meet increasing demands for bather protection while minimising harm to marine life.

Conflicts of interest

The authors declare no conflicts of interest.