Jenna Hounslow

Habitat suitability models (HSMs) are popular statistical tools used to inform decision-making for conservation planning, using species location data to characterize species-environment relationships and identify important habitats. Suitable habitats may vary according to behavior-specific resource requirements (e.g., foraging, resting), yet HSMs generally ignore behavior because obtaining spatially explicit behavioral data from wild animals is challenging. As such, suitable habitats may be incorrectly identified, and processes determining habitat selection may be misinterpreted. Despite offering unprecedented behavioral insight, contemporary multi-sensor biologgers remain underutilized in this context. We incorporated behavior into HSMs using biologging data collected from adult flatback turtles Natator depressus (n = 42) at a macrotidal study site in Western Australia and subsequently identified and characterized suitable habitat for key in-water behaviors. Foraging and resting locations derived from high-resolution motion sensor data (e.g., accelerometer, magnetometer) coupled with animal-borne video and GPS data were combined with 10 environmental features (i.e., bathymetry, aspect, slope, terrain ruggedness, distance from the coast and currents from a bespoke hydrodynamic model of the study site). A series of random forest HSMs were implemented for each behavior, accounting for temporal variation in habitat use. Bathymetry, distance from the coast, and currents best determined both foraging and resting suitability, with observed differences in habitat selection between behaviors. Overall, spatiotemporal patterns of most suitable foraging and resting habitat were similar, with shallow (10-15 m deep) nearshore (5-10 km from the coast) waters most suitable for both behaviors; however, habitats nearest to the coast (<5 km) were more suitable for foraging than resting. Overall, for foraging and resting, as water level increased turtles selected increasingly nearshore habitats where current speed was low and more variable direction. Overlap between most suitable habitats and current spatial zoning at the study site varied both seasonally and with water level, likely reflecting strong tidal influence on distribution and hence highlighting considerable opportunity for dynamic management. Our approach facilitates mechanistic insight into habitat selection and is generalizable across behaviors, taxa, and study systems, advancing the application of biologging tools to enhance the utility of HSMs and providing crucial context for decision-makers in threatened species management.

Home range size and metabolic rate of animals are theorized to scale in relation to body mass with similar exponents. This expectation has only been indirectly tested using lab-derived estimates of basal metabolic rate as proxies for field energy requirements. Therefore, it is unclear if existing theory aligns with observed patterns of home range scaling since field metabolic rates may scale differently than basal metabolic rates. We conducted the first direct field test of the relationship between home range and metabolic rate allometry. Using acoustic telemetry, we simultaneously measured the home range sizes and field metabolic rates of lemon sharks (Negaprion brevirostris) spanning one order of magnitude in body mass and compared the allometric scaling exponents of these traits. Similarity between allometric scaling exponents confirmed an expected strong association between metabolic rate and home range size. However, a nonsignificant but negative association between standard metabolic rate (SMR) and home range size suggests a complex relationship between metabolism and home range, contrasting previous assumptions of a positive relationship. Nevertheless, an overall positive association between home range size and total metabolic rate persisted, driven by a strong association between active energy expenditure and home range size. These findings underscore the intricate relationship between energetics and home range size, emphasizing the need for additional direct field investigations and the potential for modern tagging technologies to gather relevant data.

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.

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.

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.

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.

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.

Diving behaviour of ‘surfacers' such as sea snakes, cetaceans and turtles is complex and multi-dimensional, thus may be better captured by multi-sensor biologging data. However, analysing these large multi-faceted datasets remains challenging, though a high priority. We used high-resolution multi-sensor biologging data to provide the first detailed description of the environmental influences on flatback turtle (Natator depressus) diving behaviour, during its foraging life-history stage. We developed an analytical method to investigate seasonal, diel and tidal effects on diving behaviour for 24 adult flatback turtles tagged with biologgers. We extracted 16 dive variables associated with three-dimensional and kinematic characteristics for 4128 dives. K-means and hierarchical cluster analyses failed to identify distinct dive types. Instead, principal component analysis objectively condensed the dive variables, removing collinearity and highlighting the main features of diving behaviour. Generalized additive mixed models of the main principal components identified significant seasonal, diel and tidal effects on flatback turtle diving behaviour. Flatback turtles altered their diving behaviour in response to extreme tidal and water temperature ranges, displaying thermoregulation and predator avoidance strategies while likely optimizing foraging in this challenging environment. This study demonstrates an alternative statistical technique for objectively interpreting diving behaviour from multivariate collinear data derived from biologgers.

Background

Tri-axial accelerometers have been used to remotely describe and identify in situ behaviours of a range of animals without requiring direct observations. Datasets collected from these accelerometers (i.e. acceleration, body position) are often large, requiring development of semi-automated analyses to classify behaviours. Marine fishes exhibit many “burst” behaviours with high amplitude accelerations that are difficult to interpret and differentiate. This has constrained the development of accurate automated techniques to identify different “burst” behaviours occurring naturally, where direct observations are not possible.

Methods

We trained a random forest machine learning algorithm based on 624 h of accelerometer data from six captive yellowtail kingfish during spawning periods. We identified five distinct behaviours (swim, feed, chafe, escape, and courtship), which were used to train the model based on 58 predictive variables.

Results

Overall accuracy of the model was 94%. Classification of each behavioural class was variable; F1 scores ranged from 0.48 (chafe) – 0.99 (swim). The model was subsequently applied to accelerometer data from eight free-ranging kingfish, and all behaviour classes described from captive fish were predicted by the model to occur, including 19 events of courtship behaviours ranging from 3 s to 108 min in duration.

Conclusion

Our findings provide a novel approach of applying a supervised machine learning model on free-ranging animals, which has previously been predominantly constrained to direct observations of behaviours and not predicted from an unseen dataset. Additionally, our findings identify typically ambiguous spawning and courtship behaviours of a large pelagic fish as they naturally occur.

Predation is a primary selection pressure contributing to both the morphological and behavioral adaptations of organisms (Brodie 1983, Lima and Dill 1990). However, studying the anti‐predator behaviors of aquatic taxa such as sea turtles is currently limited by the difficulty of observing the natural behaviors of free‐ranging individuals at sea (Heithaus et al. 2008). Using an Animal‐borne Video and Environmental Data‐collection (AVED) biologging device, we captured a predatory interaction between a species of sea turtle, the flatback turtle (Natator depressus) and a tiger shark (Galeocerdo cuvier). This interaction occurred at Roebuck Bay, Western Australia, where we are studying the foraging ecology of flatback turtles. Given that tiger sharks are a major predator of sea turtles (Witzell 1987), this interaction was not entirely unforeseen. However, here the shark’s predatory attempt was countered by the turtle lunging multiple times toward the shark attempting to bite its attacker (Fig. 1 and video linked in Data Availability). Our unique vantage point from the perspective of the study animal led to this novel observation of behavior that might have otherwise been missed.

Understanding the behaviours of free-ranging animals over biologically meaningful time scales (e.g., diel, tidal, lunar, seasonal, annual) gives unique insight into their ecology. Bio-logging tools such as accelerometers allow the remote study of elusive or inaccessible animals by recording high resolution movement data. Machine learning (ML) is becoming a common tool for automatic classification of behaviours from these types of large data sets. These classifiers often perform best using high sampling frequencies; however, these frequencies also limit archival device recording duration through elevated battery and memory use. In this study we assess the effect of sampling frequency on a ML algorithm's ability to correctly classify behaviours from accelerometer data and present a framework for programming bio-logging devices that maintains classifier performance while optimizing data collection duration. Accelerometer data (30 Hz) were collected from juvenile lemon sharks (Negaprion brevirostris) during semi-captive trials at Bimini, Bahamas, and were ground-truthed to a discrete catalogue of behaviours through direct observation of sharks during trials. The ground-truthed data were re-sampled to a range of sampling frequencies (30, 15, 10, 5, 3 and 1 Hz) and behaviours (swim, rest, burst, chafe, headshake) were classified using a random forest ML algorithm. We demonstrate that as sampling frequency decreases, classifier performance decreases. Best overall classification was achieved at 30 Hz (F-score > 0.790), although 5 Hz was appropriate for classification of swim and rest (F-score > 0.964). For fine-scale behaviours characterised by faster kinematics (headshake, burst and chafe), classification performance was lower across the entire range of sampling frequencies (0.535–0.846, 1–30 Hz), though did not decrease significantly until sampling frequency was <5 Hz. We discuss the effects of signal aliasing and recommend that for best classification of fine-scale behaviours, frequencies >5 Hz are required. However, when seeking to maximise the available device memory and battery capacity and therefore extend deployment duration, 5 Hz is an appropriate sampling frequency for classifying behaviours in similar-sized animals.