Journal list menu

Volume 60, Issue 9 p. 1828-1840
Open Access

Behaviour-specific spatiotemporal patterns of habitat use by sea turtles revealed using biologging and supervised machine learning

Jenna L. Hounslow

Corresponding Author

Jenna L. Hounslow

Centre for Sustainable Aquatic Ecosystems, Harry Butler Institute, Murdoch University, Murdoch, Western Australia, Australia

Environmental and Conservation Sciences, Murdoch University, Murdoch, Western Australia, Australia


Jenna L. Hounslow

Email: [email protected]

Search for more papers by this author
Sabrina Fossette

Sabrina Fossette

Biodiversity and Conservation Science, Department of Biodiversity, Conservation & Attractions, Kensington, Western Australia, Australia

Search for more papers by this author
Wei Chong

Wei Chong

Centre for Sustainable Aquatic Ecosystems, Harry Butler Institute, Murdoch University, Murdoch, Western Australia, Australia

Environmental and Conservation Sciences, Murdoch University, Murdoch, Western Australia, Australia

Search for more papers by this author
Randa Bali

Randa Bali

Centre for Sustainable Aquatic Ecosystems, Harry Butler Institute, Murdoch University, Murdoch, Western Australia, Australia

Environmental and Conservation Sciences, Murdoch University, Murdoch, Western Australia, Australia

Search for more papers by this author
Anton D. Tucker

Anton D. Tucker

Biodiversity and Conservation Science, Department of Biodiversity, Conservation & Attractions, Kensington, Western Australia, Australia

Search for more papers by this author
Scott D. Whiting

Scott D. Whiting

Biodiversity and Conservation Science, Department of Biodiversity, Conservation & Attractions, Kensington, Western Australia, Australia

Search for more papers by this author
Adrian C. Gleiss

Adrian C. Gleiss

Centre for Sustainable Aquatic Ecosystems, Harry Butler Institute, Murdoch University, Murdoch, Western Australia, Australia

Environmental and Conservation Sciences, Murdoch University, Murdoch, Western Australia, Australia

Search for more papers by this author
First published: 29 June 2023
Handling Editor: Hedley Grantham


  1. Conservation of threatened species and anthropogenic threat mitigation commonly rely on spatially managed areas selected according to habitat preference. Since the impact of threats can be behaviour-specific, such information could be incorporated into spatial management to improve conservation outcomes. However, collecting spatially explicit behavioural data is challenging.
  2. Using multi-sensor biologging tags containing high-resolution movement sensors (e.g. accelerometer, magnetometer, GPS) and animal-borne video cameras, combined with supervised machine learning, we developed a method to automatically detect and geolocate typically ambiguous behaviours for the poorly understood flatback turtle Natator depressus. Subsequently, we evaluated behaviour-specific spatiotemporal patterns of habitat use.
  3. Boosted regression trees successfully identified the presence of foraging and resting in 7074 dives (AUC > 0.9), using dive features representing characteristics of locomotory activity, body posture, and three-dimensional dive paths validated by ancillary video data. Foraging was characterised by dives with longer duration, variable depth, tortuous bottom phases; resting was characterised by dives with decreased locomotory activity and longer duration bottom phases.
  4. Foraging and resting showed minimal spatial segregation based on 50% and 95% utilisation distributions. Expected diel patterns of behaviour-specific habitat use were superseded by the extreme tides at the near-shore study site. Turtles rested in areas close to the subtidal and intertidal boundary within larger overlapping foraging areas, allowing efficient access to intertidal food resources upon inundation at high tides when foraging was ~25% more likely.
  5. Synthesis and applications. Using supervised machine learning and biologging tools, we show the potential for dynamic spatial management of flatback turtles to mitigate behaviour-specific threats by prioritising protection of important locations at pertinent times. Although results are a species-specific response to a super-tidal environment, our approach can be generalised to a broad range of taxa and study systems, facilitating a conceptual advance in spatial management.


Many species are facing major declines globally, with survival increasingly threatened by anthropogenic threats including habitat loss, overexploitation, and climate change (Butchart et al., 2010). Area-based conservation—spatial management—is fundamental to protecting threatened species and preventing such declines (Dudley & Stolton, 2010). Spatially managed areas are theoretically designated according to an understanding of species spatial ecology and habitat preference, in order to mitigate the impacts of anthropogenic threats (Watson et al., 2014). However, effective protection requires a more nuanced understanding of species behaviours within spatially managed areas (Camphuysen et al., 2012). This is partly because the impact of certain threats can be behaviour dependent; for example, bats Myotis sp. are particularly sensitive to noise pollution during foraging (Francis & Barber, 2013). Spinner dolphins Stenella longirostris were inflexible in the times and locations used for resting, with authors suggesting resting habitat should be protected via spatiotemporal exclusion of all anthropogenic activities (Tyne et al., 2015). Furthermore, the impact of vessel disturbance on endangered killer whales Orcinus orca was greatest when feeding. Compared to indiscriminately protecting larger habitats based only on presence, prioritisation of smaller but more important foraging areas for protection from vessel traffic was proposed to offer superior conservation benefits (Ashe et al., 2010).

Endangered worldwide, sea turtles face many threats where behavioural information offers critical context for spatial management planning. For instance, accidental by-catch by bottom trawlers or entrainment in dredging operations are more likely when sea turtles rest on the seafloor, compared to when turtles are swimming midwater or near the surface (Ramirez et al., 2017). Additionally, sea turtles are at risk of disturbance and injury from overhead vessel traffic, both at the surface and when resting on the seafloor in shallow waters. Management interventions to mitigate impacts from vessel traffic, such as ‘go-slow’ vessel zones, could therefore be aligned with the location and timing of resting (Shimada et al., 2017). Conversely, dredging indirectly impacts benthic foraging via habitat disruption and the release of toxins previously trapped by bottom sediment, while the risk of debris ingestion substantially increases during pelagic foraging (Goldberg et al., 2015; Schuyler et al., 2014). These examples provide clear support for incorporating behavioural information into spatial management practices, to maximise conservation benefits.

Collecting spatially explicit behavioural data in the wild remains challenging, however, especially for marine species whose lifestyles preclude direct observation for extended periods. High-use habitats can be identified for spatial management via a toolbox of telemetry (e.g. satellite, acoustic, radio) and survey (e.g. aerial surveys, mark recapture, photo identification) techniques (Block et al., 2011; Hays & Hawkes, 2018). Despite informing global conservation of sea turtles according to life-stage via ocean zoning (e.g. protecting migratory corridors or entire foraging areas; Baudouin et al., 2015; Shimada et al., 2017), contextually poor data generated from these conventional techniques offer incomplete perspectives of sub-surface behaviours.

Contemporary biologging tools have transformed behavioural studies for free-ranging animals, allowing cryptic behaviours to be remotely monitored and quantified with greater precision than ever before (Ropert-Coudert & Wilson, 2005). Combining multiple high-resolution movement sensors (e.g. accelerometer) with environmental (e.g. pressure) and location (e.g. GPS) data measured by one animal-attached device reveals typically ambiguous, naturally occurring behaviours (Wilson et al., 2008). For instance, multi-sensor data have confirmed the location and timing of spawning by wild kingfish Seriola sp. and prey attack by pumas Puma concolor (Clarke et al., 2021; Wang et al., 2015). Similarly, three-dimensional trajectories reconstructed using magnetometer data elucidated foraging by northern fur seals Callorhinus ursinus (Battaile et al., 2015). Already vastly improving our inferential capacity, behaviours deduced from movement sensor data can also be verified using ancillary video footage from animal-borne video cameras on wild individuals (Moll et al., 2007).

To that end, modern biologging tools offer a myriad of opportunities to quantify fine-scale behaviours and provide the contextual enrichment necessary to protect threatened species. This study aimed to use biologging tools to evaluate behaviour-specific patterns of habitat use by a species of sea turtle, the flatback turtle Natator depressus. Specifically, our objectives were to (1) develop a method to identify fine-scale behaviours from multi-sensor data in order to (2) describe and evaluate behaviour-specific spatiotemporal patterns of habitat use and (3) discuss implications for spatial planning and conservation management.


All animal use was approved by Animal Ethics Committees (MU: 653-R3164/19 & DBCA: 2016-18/2019-12-B) under relevant licences (DBCA: 08-009604-1 and Department of Primary Industries and Regional Development U6-2017-2019/2020-2022).

2.1 Study species and site

Listed as data-deficient (IUCN, 1996), flatback turtles have a distribution limited to northern shelf waters of Australia and face significant in-water threats associated with increasing coastal development (Limpus, 2007; Whittock et al., 2016). Information on the flatbacks' foraging life-stage is lacking, mainly due to remote access difficulties. Since 2018, flatback turtles have been reliably accessible during this life-stage at Yawuru Nagulagun Roebuck Bay (YNRB; Hounslow et al., 2021). YNRB is a 360 km2 embayment of the Indian Ocean in the Kimberley region of Western Australia (see Figure S1), with a sub-tropical climate of wet summers (October–March) and dry winters (April–September). YNRB is lined by mangroves and tidal creeks and dominated by semi-diurnal tides ranging up to ~10 m, inundating up to 200 square km intertidal sand and mud flats (Pearson et al., 2008).

2.2 Field protocol and tag deployment

Fieldwork was conducted between August 2018 and May 2021. Adult flatback turtles were equipped with either a CATS-Diary (Customised Animal Tracking Solutions; CATS, Moffat Beach, Australia) or CATS-Cam (aka Diary plus video camera) tag on an ad-hoc basis for up to 1 week each. For detailed capture, tag attachment, programming, and retrieval information, see Hounslow et al. (2022).

2.3 Automatically identifying behaviours

All analyses were performed in R (v 4.0.3) (R Core Team, 2020) unless otherwise stated. Incomplete deployments (e.g. pressure sensor or video camera failure) were discarded. The first 2 h data from complete deployments were discarded to account for post-capture recovery (Thomson & Heithaus, 2014). Only ~3.3% of the total multi-sensor diary data collected were associated with ancillary video, either because the tag did not include a camera (CATS-Diary), or due to duty-cycling required for battery and memory intensive camera operation (CATS-Cam) (Table S1). Therefore, we developed a predictive supervised machine learning model to automatically identify behaviours from multi-sensor diary data without associated video.

2.3.1 Video processing

Videos from CATS-Cam deployments were watched using BORIS event-coding software (v.7.12; Friard & Gamba, 2016), with behaviours (start and end) logged following a custom-written ethogram when visibility was adequate (Table S2). Behavioural categories (states) included ‘Swim’, ‘Rest’, ‘Surface’, ‘Forage’, ‘Dig’ and ‘Other’ (e.g. antipredator responses) (Movie S1; Table S2). We retained only foraging and resting as behaviours of interest for automatic identification, since they determine energy gain and conservation and ultimately fitness of animals, therefore are of high conservation value (Stephens & Krebs, 2019). Within foraging, we only retained the subcategory expressed adequately for further analyses–benthic foraging (hereafter foraging) (sensu Jeantet et al., 2018). All other observed behaviours were not used in subsequent analyses.

2.3.2 Feature extraction and labelling

Supervised machine learning requires input of labelled features to predict classes (behaviours) of interest (Valletta et al., 2017). Features (n = 14) were calculated by first segmenting the multi-sensor data into individual dives. Dives and their phases (descent, bottom, ascent, post-dive surface interval) were identified from zero-offset-corrected depth time-series resampled to 1 Hz, (depth threshold >1 m for ≥30 s) using the diveMove package (Luque, 2019). Dive-phase specific features were extracted, characterising kinematic (locomotory activity and body posture) and three-dimensional (tortuosity, duration and depth) aspects of behaviour (Table S3), following methods described in Hounslow et al. (2022).

The dive features were then assigned binary behaviour labels according to the presence or absence of foraging and resting during the dive, using corresponding encoded video. Behaviour was labelled present (1) if it occurred, whereas absence (0) was only labelled if it did not occur and 100% of the bottom phase was associated with video. For each behaviour, unlabelled dives not meeting the decision rule for assigning presence or absence labels or not associated with video (e.g. all CATS-Diary dives and CATS-Cam dives when duty-cycled video camera was not recording), were assigned an NA label (85.3% of dives for foraging and 60.1% for resting).

2.3.3 Model building

Two separate boosted regression tree (BRT) classification algorithms were built to predict foraging and resting during all dives, including dives labelled NA. BRTs are a series of decision trees (recursive splits), where predictive performance for each new tree is sequentially improved (boosted) by considering the prediction error of the previous fitted tree (Elith et al., 2008). For each behaviour, the binary labelled dive features were randomly split for model training (80%) and testing (20%). BRTs were implemented using the dismo package (Hijmans et al., 2021) and optimised by internal 10-fold cross-validation (Hastie et al., 2009). BRT performance was assessed against the test data and evaluated via area under receiver operating curve (AUC) (package pROC; Robin et al., 2011). A model predicting no better than chance returns AUC = 0.5, while AUC > 0.9 reflects strong predictive ability. Model building is explained in detail in the Supporting Information (Appendix S1).

2.3.4 Model application

Behaviours during dives labelled NA were predicted by separately fitting each BRT to dive features for all dives. Predicted probabilities required conversion to presence/absence labels (1/0). To avoid overestimating prevalence of each behaviour, we visually selected a conversion threshold that minimised false positives (foraging = 0.3, resting = 0.7; Figure S2).

2.4 Spatiotemporal analysis of foraging and resting behaviours

2.4.1 Spatial analysis

To quantify behaviour-specific habitat use, dives were geolocated from continuous location estimates derived from georeferenced tracks calculated via GPS-anchored dead-reckoning, a method particularly suited to species who spend the majority of time submerged (Bidder et al., 2015; Wilson et al., 2007). For each individual, raw GPS data were filtered (number of satellites >5, altitude >0 m and <100 m), and duplicates were removed. Naturally, the number of surface GPS locations varied per individual (e.g. satellite availability, frequency and duration of surface intervals). After filtering, no GPS data were available for two individuals with very short duration deployments (IDs 25 and 44; Table S1).

Georeferenced tracks were individually calculated using the TrackReconstruction package (Battaile, 2021). Dynamic acceleration (DA) from the turtles anterior–posterior body axis was used as a proxy for speed (Wilson et al., 2007). DA, separated from static acceleration via running mean length (RmL) = 4 s, was normalised from 0 to maximum speed = 1.5 ms−1. Georeferenced tracks were created by forcing dead-reckoned tracks through filtered GPS locations. Behaviour-labelled dives were geolocated by selecting the latitude and longitude coordinates from the georeferenced track corresponding to the date and time at the mid-point of the dive. While georeferencing helps adjust for some error from ocean currents (Battaile et al., 2015), acceleration-based speed proxies cause location error to accumulate during rest periods. To account for this, geolocated dives were discarded (n = 1752) if they were more than 48.4 min (95th percentile of dive duration for all dives) from the nearest GPS anchor point (Figure S3).

To quantify behaviour-specific habitat use, we selected geolocated dives where behaviour was present (1) to estimate utilisation distributions (UDs) from kernel density estimates (KDE) (foraging = 2658, resting = 1948). KDEs were estimated over different temporal covariates corresponding to the date and time at the mid-point of the dive; season (winter or summer), diel period (day or night) and tide height (m) represented by hourly water level. Continuous water level data were converted to a categorical variable to allow for comparisons between UDs, by examining the overall distribution of observed water level throughout the entire study period; low (0.2–4.0 m), mid (4.1–6.9 m) and high (7.0–10.6 m).

KDEs and UDs were estimated using the package adehabitatHR, with a consistent grid cell size (50 m), extent (0.5) and a 5 km buffer (v.0.4.19; Calenge, 2006). The smoothing parameter was set as the reference bandwidth (h). Core use and representative areas were defined as the areas containing 50% and 95% of a UD (UD-50 and UD-95 respectively). We compared the distribution and extent of each behaviour for each temporal covariate by quantifying UD area and indices of overlap as a measure of similarity between UDs. Indices of overlap (method = HR; the proportion of one UD covered by another UD) were calculated to compare UDs between temporal covariates (Fieberg & Kochanny, 2005). UDs were visualised in QGIS (v.3.18.1). Sample size was checked as per Shimada et al. (2021) (Figure S4). For details on temporal covariates and checking sample size, see Appendix S1.

2.4.2 Temporal analysis

Generalised linear mixed models (GLMMs), allowing binary responses (i.e. behaviour-labelled dives) and both fixed and random effects (Zuur et al., 2009), were used to test the significance of season, diel period and tide height on the prevalence of foraging and resting during dives. For each behaviour, a global model was fitted including all terms and their interactions, then the final model was selected according to Akaike's information criterion (AIC) and parsimony (Barton, 2022). For detailed implementation of GLMMs, see Appendix S1.


3.1 Behavioural ethogram

We successfully deployed and retrieved 51 tags. We collected 102.2 days multi-sensor data from 44 complete deployments, with 3.3% (83 h) associated with ancillary video recorded from individuals equipped with CATS-Cams (n = 22) (Table S1). Total encoded video ranged from 0.2 to 7.4 h per individual (= 3.7 ± 2.0 h; Table S1). Combined, flatback turtles spent most of the time swimming (50.3%), resting on the sea floor (34.8%, Figure 1a) or at the surface (13.6%) (Movie S1; Table S4). Overall, proportion of time spent foraging was <1% of the combined ethogram. Foraging was predominantly benthic (79.8% of all foraging; Figure 1b), observed by all but three individuals (Table S4). Foraging was rarely pelagic (1.8%, Table S4), with three individuals observed to forage in the water column on jellyfish. Only one individual (ID 6) was observed to potentially forage at the surface, spending ~5 min (18.3% of total observed foraging) exploring patches of floating weed. Two individuals were not observed to forage at all (ID 44 and 52), but these deployments were <0.1 day in duration (Tables S1 and S4). Out of 22 turtles, 15 used their fore flippers to dig through the benthos, representing 0.3% of the total ethogram.

Details are in the caption following the image
Video screen shots and representative 3-D dives for (a) resting and (b) foraging by adult flatback turtles. Dive paths coloured by overall dynamic body acceleration (ODBA) as proxy for locomotory activity. Inset same dives in 2-D. Yellow arrow = dive path start.

3.2 Boosted regression trees

We identified 7074 dives for feature extraction (Table S1). The total number of ground-truthed dives used for model training and validation was 354 for foraging and 416 for resting. Observed prevalence rates of foraging and resting during dives were 14.7% and 39.9% respectively. Validated BRT models scored an AUC >0.9 for both behaviours (Figure S5), indicating strong predictive ability. Dive features with the highest influence on the foraging BRT were related to bottom phase duration (43.3%), bottom wiggle (23.1%) and tortuosity (7.0%) (Figure 2a). Foraging more often occurred during dives with longer duration bottom phases, increased depth variation (wiggle) and during dives with more tortuous bottom phases (Figures 1b and 2a). Meanwhile the most influential dive features for resting were locomotory activity during the bottom phase (38.9%) and duration of the bottom (13.2%) and ascent (8.5%) phases (Figure 2b). Turtles were more likely to be engaged in resting during dives when locomotory activity was low during longer duration bottom phases, and shorter duration ascent phases (Figures 1a and 2b). After fitting validated BRTs to all dives, the predicted prevalence rate for foraging during dives was 43.2% and 32.4% for resting. Both foraging and resting co-occurred during the same dive in 24.2% dives and neither behaviour occurred in 47.7% of dives.

Details are in the caption following the image
Marginal effects (fitted function) of dive features, ranked by predictive influence (%) on boosted regression tree model predictions for (a) foraging and (b) resting during dives by adult flatback turtles. See Table S3 for full explanation of dive features.

3.3 Spatiotemporal patterns of foraging and resting

Flatback turtles used most of YNRB, exhibiting minimal spatial segregation between core and representative foraging and resting areas but displayed seasonal, diel, and tidal changes in habitat use (Figure 3). Spatial overlap between seasons was low; during summer turtles used two discrete locations in the bay with only 5.2% of this area overlapped by the winter foraging area, which was similar for resting (9.4% overlap) (Figure 3a). Core areas used by turtles for both foraging and resting contracted during winter (47.4 and 33.4 km2, respectively) and were generally located in subtidal waters in close proximity to the intertidal zone (static subtidal-intertidal boundary delineated by red line in Figure 3; vector representation of median tide height (m) for the lowest interval of the observed intertidal range from Bishop-Taylor et al. (2019)) (Table 1; Figure 3a). Turtles generally used a larger area for foraging than resting under all temporal covariates (Table 1; Figure 3). During the day, core foraging area extent (88.1 km2) spanned more than double the core resting area (41.5 km2; Table 1; Figure 3b). Spatial distribution was similar between day and night for both behaviours; (85.7% and 99.6% day-night overlap for resting and foraging respectively; Table 1; Figure 3b). There was also a tidal influence on the spatial distribution of both behaviours. Both core and representative areas used by turtles for foraging and resting shifted further towards the intertidal zone as tide height increased, and furthest when tide height was 7.0 m (Figure 3c). The extent of areas used by turtles for both foraging and resting was smallest on intermediate tide heights (Table 1).

Details are in the caption following the image
Behaviour-specific spatiotemporal patterns of habitat use by adult flatback turtles (n = 42) at Yawuru Nagulagun Roebuck Bay, Western Australia. Core (50% UD; shaded polygons) and representative (95% UD; outline) resting and foraging areas according to (a) season, (b) diel period and (c) tide height. Red line represents static subtidal-intertidal boundary. Sample size (n dives) on each panel.
TABLE 1. Extent of core (UD-50) and representative (UD-95) areas used by flatback turtles (n = 42) for foraging and resting at Yawuru Nagulagun Roebuck Bay, Western Australia. UD, utilisation distribution.
Behaviour Temporal covariate UD-50 (km2) UD-95 (km2)
Forage Season Summer 73.9 307.6
Winter 47.4 303.1
Tide High 101.9 367.3
Mid 61.9 296.7
Low 97.1 379.0
Diel period Day 88.1 394.1
Night 86.9 317.8
Rest Season Summer 79.7 280.6
Winter 33.4 220.4
Tide High 76.2 297.3
Mid 39.7 238.3
Low 70.8 337.0
Diel period Day 41.5 279.6
Night 65.8 277.3

Model selection determined that tide was the main temporal driver of foraging during dives (Table S5). Turtles were more likely to forage during a dive as tide height increased, with the probability of foraging increasing by ~25% as tide height increased from ~2.5 to >7.5 m (Figure 4a). The effect of season, diel period and tide on resting during dives was more complex, with the top candidate model including all fixed terms and their interactions (Table S5). Overall, there was a similar effect of tide on turtles resting (probability increased with tide height) except during the day in winter when turtles were slightly less likely to rest as tide height increased (Figure 4b).

Details are in the caption following the image
Generalised linear mixed models (GLMMs) showing marginal temporal effects (shading = 95% confidence interval) on (a) foraging and (b) resting during a dive by flatback turtles (n = 44) at Yawuru Nagulagun Roebuck Bay, Western Australia.


The impact of certain threats on threatened species can be context dependent. As such, a strong spatiotemporal understanding of behaviour is critical for adequate and effective protection within spatially managed areas. However, understanding the timing and location of fine-scale behaviours of free-ranging animals remains challenging, particularly in aquatic habitats. Using supervised machine learning and high-resolution multi-sensor biologging data for a species of sea turtle, this study demonstrates the potential for incorporating behavioural information into spatial management practices for conservation. Supervised machine learning has become a popular tool to automatically identify and locate animal behaviours from multi-sensor datasets (Nathan et al., 2022). Here, boosted regression tree models performed exceptionally well, identifying both foraging and resting by flatback turtles with an AUC score >0.9, consistent with other studies automatically identifying sea turtle behaviours (e.g. AUC = 0.88; Jeantet et al., 2021).

4.1 Optimising automatic identification of behaviours

Behaviour classification is frequently plagued with performance issues associated with animals' natural behavioural repertoires. Predictive models are often biased towards more frequently performed behaviours (Sakai et al., 2019), or behaviours with similar movement characteristics are misclassified as each other (Jeantet et al., 2018). Here, turtles were observed to disproportionately spend most of their time swimming or resting, and rarely foraging (<1% observed time). Behavioural time-budgets were also imbalanced for green Chelonia mydas, hawksbill Eretmochelys imbricata and loggerhead Caretta caretta turtles (Jeantet et al., 2018, 2020, 2021). Furthermore, turtles were observed to forage and rest while stationary and horizontal on the seafloor, expressing similar kinematics of locomotion and body posture. In anticipation of performance issues, we focused on two key behaviours fundamental to somatic growth, fitness and survival, that is, resting and foraging (Stephens & Krebs, 2019). In turn, this allowed us to treat proportionally dissimilar, yet kinematically similar behaviours as separate binary classification problems, avoiding model confusion between behaviours.

4.2 Dive characteristics and behaviour

Dives encompass fundamental aspects of sea turtle behaviour, yet it has proved challenging to reliably interpret behaviour from the characteristics of dives alone (Hounslow et al., 2022; Seminoff et al., 2006). We used ancillary video data to validate descriptive dive features characterising kinematic and three-dimensional aspects of dives, to provide reliable and objective biological insight that so far has remained elusive in dive classification from time-depth data. Unexpectedly, flatback turtles foraged during dives with longer duration bottom phases. Previously such dives were widely considered to represent resting in other species of sea turtle, resulting from lower oxygen consumption (Southwood et al., 2003). Alternatively, longer bottom durations could indicate turtles maximising time spent on the seafloor at profitable foraging patches, in accordance with optimal foraging theory (Charnov, 1976). Bottom phase wiggle was another distinguishing feature of foraging by flatback turtles, supporting studies that have indirectly ascribed foraging to wiggles in two-dimensional dive profiles for a range of aquatic taxa (Bost et al., 2007; Goldbogen et al., 2006). Foraging dives were also strongly associated with tortuous movement paths during the bottom phase of dives. Although tortuous movements have long been attributed to foraging (area-restricted search; Dorfman et al., 2022), this feature only made a minor contribution to the identification of foraging dives. In addition to clarifying features of foraging dives, we show that resting dives by flatback turtles were primarily characterised by less active bottom phases of increased duration, as predicted by theory (Southwood et al., 2003).

4.3 Behaviour-specific spatiotemporal patterns of habitat use

We found that foraging and resting by flatback turtles were minimally spatially segregated, which was surprising considering that spatiotemporal patterns of behaviour have been widely assumed to occur in sea turtles. For instance, green turtles commonly exhibit diel patterns of diurnal foraging and nocturnal resting, usually contingent with expanded movements by day associated with shallow sea grass beds and restricted use of deeper, complex habitats at night (Ballorain et al., 2013; Chambault et al., 2020; Christiansen et al., 2017; Taquet et al., 2006). This pattern was not exhibited by flatback turtles, who showed high diel overlap in areas used for both foraging and resting (~85%–99%). The extreme tides at the study site and ensuing tidal currents may best explain the lack of diel pattern in behaviour-specific habitat use. This lack of diel pattern was attributed to tide-mediated foraging, where the need to feed on intertidal resources at high tide supersedes expected spatiotemporal habitat use patterns (Pillans et al., 2021; Whiting & Miller, 1998). Here, tide was the only significant temporal driver of foraging by flatback turtles, with the probability of foraging during a dive increasing by ~25% as tide height increased above 7.5 m. Alongside evidence that turtles moved further into the intertidal zone on high tides, this strongly suggests flatback turtles forage intensely on intertidal resources. Despite sea turtles being visual predators relying on daylight to forage, we also found that flatbacks forage at night, not unfeasible on moonlit nights (Ballorain et al., 2013; Taquet et al., 2006).

The area used by flatback turtles for resting was generally smaller than and within the foraging area, but this may reflect the stationary characteristics of dives where resting occurred (Matley et al., 2021). Foraging and resting overlap was proposed as a mechanism increasing foraging efficiency by minimising transport costs between discrete foraging and resting areas, akin to central place foraging (Okuyama et al., 2013). Flatback turtles generally rested in the subtidal zone in closer proximity to the intertidal zone, allowing efficient access to intertidal foraging areas that are only available for limited periods on certain tides.

As well as increasing foraging efficiency by overlapping behaviours, resting in shallow, turbid waters near the intertidal zone likely reduces exposure to predators. Tiger sharks Galeocerdo cuvier predate on flatback turtles at the study site (Hounslow et al., 2021), and this strategy has previously been attributed to predator avoidance (Chambault et al., 2020; Christiansen et al., 2017). Any seasonal variation in habitat use by flatback turtles in this study was probably linked to seasonal variation in the distribution of benthic food resources, as well as seasonal changes in turtles' physiological performance capacity (Christiansen et al., 2017; Seminoff et al., 2020). Turtles used larger areas for both foraging and resting in summer (73.9 and 79.7 km2 respectively), compared to contracted areas during winter (47.4 and 33.4 km2), likely reflecting seasonal disparity in water temperature at the sub-tropical study site (~10°C). Warmer water temperatures increase physiological performance for ectotherms, thus enabling turtles to search for food over larger areas during summer (Seminoff et al., 2020).

4.4 Implications for conservation management

We demonstrate a technique to measure behaviour-specific spatiotemporal patterns of habitat use in the wild, providing contextual enrichment necessary for effective conservation of threatened species. Information on the timing and location of precise behaviours is critical for planning and prioritising when and where threat mitigation efforts should be applied within spatially managed areas. From a static management perspective, such protection could include area closures at foraging areas. Alternatively, dynamic management strategies that focus on specific smaller areas at pertinent times may offer adequate protection (Maxwell et al., 2015), such as temporary restrictions in intertidal areas on high tides, and proximate subtidal waters.

Our approach is not only applicable to other observed behaviours, but more notably to a broad range of taxa and study systems, facilitating a conceptual advance in spatial management. As Ashe et al. (2010) point out, flexible approaches are especially relevant where static protection (i.e. total exclusion of threatening anthropogenic activities) is unfeasible. Narrowly targeted conservation measures optimise protection but can also minimise stakeholder inconvenience. Consequently, we expect that incorporating behavioural information into spatial management will also facilitate stakeholder uptake and enforcement by managers, further maximising conservation outcomes.


Jenna L. Hounslow, Sabrina Fossette and Adrian C. Gleiss conceived the ideas. Jenna L. Hounslow, Sabrina Fossette, Anton D. Tucker and Scott D. Whiting collected the data. Jenna L. Hounslow conducted analyses with input from Wei Chong and Randa Bali. Jenna L. Hounslow led the writing of the manuscript. All authors contributed critically to the drafts and gave final approval for publication.


This research was carried out on the Sea Country of the Yawuru people, and we pay our respects to them as traditional custodians. This research falls under the Northwest Shelf Flatback Turtle Conservation Program, administered by the Department of Biodiversity, Conservation and Attractions (DBCA) in collaboration with Nyamba Buru Yawuru Pty. Ltd. We thank Yawuru Country Managers, Dean Mathews, Todd Quartermaine, Wil Bennet, DBCA West Kimberley District staff, James Gee, Broome Whale Watching, Quest Maritime Services and volunteers Oliver Jewell, Blair Bentley, Natalie Hill, Kate Salter, Malindi Gammon and Brodee Elsdon. Nikolai Liebsch, Lauren Peel, Olivier Friard and Brian Battaile generously provided prompt support during analyses. JLH was funded by the Australian Government Research Training Program and Murdoch University. Research funded by DBCA (MU:18841) and supported by the Holsworth Wildlife Research Endowment & Ecological Society of Australia (MU:20447). Open access publishing facilitated by Murdoch University, as part of the Wiley - Murdoch University agreement via the Council of Australian University Librarians.


    Authors declare no conflicts of interest.


    Data available from the Dryad Digital Repository (Hounslow et al., 2023).