Supervised versus unsupervised approaches to classification of accelerometry data

Sophisticated animal-borne sensor systems are increasingly providing novel insight into how animals behave and move. Despite their widespread use in ecology, the diversity and expanding quality and quantity of data they produce have created a need for robust analytical methods for biological interpretation. Machine learning tools are often used to meet this need. However, their relative effectiveness is not well known and, in the case of unsupervised tools, given that they do not use validation data, their accuracy can be difficult to assess. We evaluated the effectiveness of supervised (n = 6), semi-supervised (n = 1), and unsupervised (n = 2) approaches to analyzing accelerometry data collected from critically endangered California condors (Gymnogyps californianus). Unsupervised K-means and EM (expectation-maximization) clustering approaches performed poorly, with adequate classification accuracies of <0.8 but very low values for kappa statistics (range: -0.02 to 0.06). The semi-supervised nearest mean classifier was moderately effective at classification, with an overall classification accuracy of 0.61 but effective classification only of two of the four behavioral classes. Supervised random forest (RF) and k-nearest neighbor (kNN) machine learning models were most effective at classification across all behavior types, with overall accuracies >0.81. Kappa statistics were also highest for RF and kNN, in most cases substantially greater than for other modeling approaches. Unsupervised modeling, which is commonly used for the classification of a priori-defined behaviors in telemetry data, can provide useful information but likely is instead better suited to post hoc definition of generalized behavioral states. This work also shows the potential for substantial variation in classification accuracy among different machine learning approaches and among different metrics of accuracy. As such, when analyzing biotelemetry data, best practices appear to call for the evaluation of several machine learning techniques and several measures of accuracy for each dataset under consideration.

Biomechanically distinct filter-feeding behaviors distinguish sei whales as a functional intermediate and ecologically flexible species

With their ability to facultatively switch between filter-feeding modes, sei whales represent a functional and ecological intermediate in the transition between intermittent and continuous filter feeding. Morphologically resembling their lunge-feeding, rorqual relatives, sei whales have convergently evolved the ability to skim prey near the surface of the water, like the more distantly related balaenids. Because of their intermediate nature, understanding how sei whales switch between feeding behaviors may shed light on the rapid evolution and flexibility of filter-feeding strategies. We deployed multi-sensor bio-logging tags on two sei whales and measured the kinematics of feeding behaviors in this poorly understood and endangered species. To forage at the surface, sei whales used a unique combination of surface lunges and skim-feeding behaviors. The surface lunges were slow and stereotyped, and were unlike lunges performed by other rorqual species. The skim-feeding events featured a different filtration mechanism from the lunges and were kinematically different from the continuous filter feeding used by balaenids. While foraging below the surface, sei whales used faster and more variable lunges. The morphological characteristics that allow sei whales to effectively perform different feeding behaviors suggest that sei whales rapidly evolved their functionally intermediate and ecologically flexible form to compete with larger and more efficient rorqual species.

Two hundred years of zooplankton vertical migration research

Vertical migration is a geographically and taxonomically widespread behaviour among zooplankton that spans across diel and seasonal timescales. The shorter-term diel vertical migration (DVM) has a periodicity of up to 1 day and was first described by the French naturalist Georges Cuvier in 1817. In 1888, the German marine biologist Carl Chun described the longer-term seasonal vertical migration (SVM), which has a periodicity of ca. 1 year. The proximate control and adaptive significance of DVM have been extensively studied and are well understood. DVM is generally a behaviour controlled by ambient irradiance, which allows herbivorous zooplankton to feed in food-rich shallower waters during the night when light-dependent (visual) predation risk is minimal and take refuge in deeper, darker waters during daytime. However, DVMs of herbivorous zooplankton are followed by their predators, producing complex predator-prey patterns that may be traced across multiple trophic levels. In contrast to DVM, SVM research is relatively young and its causes and consequences are less well understood. During periods of seasonal environmental deterioration, SVM allows zooplankton to evacuate shallower waters seasonally and take refuge in deeper waters often in a state of dormancy. Both DVM and SVM play a significant role in the vertical transport of organic carbon to deeper waters (biological carbon sequestration), and hence in the buffering of global climate change. Although many animal migrations are expected to change under future climate scenarios, little is known about the potential implications of global climate change on zooplankton vertical migrations and its impact on the biological carbon sequestration process. Further, the combined influence of DVM and SVM in determining zooplankton fitness and maintenance of their horizontal (geographic) distributions is not well understood. The contrasting spatial (deep versus shallow) and temporal (diel versus seasonal) scales over which these two migrations occur lead to challenges in studying them at higher spatial, temporal and biological resolution and coverage. Extending the largely population-based vertical migration knowledge base to individual-based studies will be an important way forward. While tracking individual zooplankton in their natural habitats remains a major challenge, conducting trophic-scale, high-resolution, year-round studies that utilise emerging field sampling and observation techniques, molecular genetic tools and computational hardware and software will be the best solution to improve our understanding of zooplankton vertical migrations.

Forced into an ecological corner: Round-the-clock deep foraging on small prey by elephant seals

Small mesopelagic fishes dominate the world's total fish biomass, yet their ecological importance as prey for large marine animals is poorly understood. To reveal the little-known ecosystem dynamics, we identified prey, measured feeding events, and quantified the daily energy balance of 48 deep-diving elephant seals throughout their oceanic migrations by leveraging innovative technologies: animal-borne smart accelerometers and video cameras. Seals only attained positive energy balance after feeding 1000 to 2000 times per day on small fishes, which required continuous deep diving (80 to 100% of each day). Interspecies allometry suggests that female elephant seals have exceptional diving abilities relative to their body size, enabling them to exploit a unique foraging niche on small but abundant mesopelagic fish. This unique foraging niche requires extreme round-the-clock deep diving, limiting the behavioral plasticity of elephant seals to a changing mesopelagic ecosystem.

Predicting suitable coastal habitat for sei whales, southern right whales and dolphins around the Falkland Islands

Species distribution models (SDMs) are valuable tools for describing the occurrence of species and predicting suitable habitats. This study used generalized additive models (GAMs) and MaxEnt models to predict the relative densities of four cetacean species (sei whale Balaeanoptera borealis, southern right whale Eubalaena australis, Peale’s dolphin Lagenorhynchus australis, and Commerson’s dolphin Cephalorhynchus commersonii) in neritic waters (≤100 m depth) around the Falkland Islands, using boat survey data collected over three seasons (2017–2019). The model predictor variables (PVs) included remotely sensed environmental variables (sea surface temperature, SST, and chlorophyll-a concentration) and static geographical variables (e.g. water depth, distance to shore, slope). The GAM results explained 35 to 41% of the total deviance for sei whale, combined sei whales and unidentified large baleen whales, and Commerson’s dolphins, but only 17% of the deviance for Peale’s dolphins. The MaxEnt models for all species had low to moderate discriminatory power. The relative density of sei whales increased with SST in both models, and their predicted distribution was widespread across the inner shelf which is consistent with the use of Falklands’ waters as a coastal summer feeding ground. Peale’s dolphins and Commerson’s dolphins were largely sympatric across the study area. However, the relative densities of Commerson’s dolphins were generally predicted to be higher in nearshore, semi-enclosed, waters compared with Peale’s dolphins, suggesting some habitat partitioning. The models for southern right whales performed poorly and the results were not considered meaningful, perhaps due to this species exhibiting fewer strong habitat preferences around the Falklands. The modelling results are applicable to marine spatial planning to identify where the occurrence of cetacean species and anthropogenic activities may most overlap. Additionally, the results can inform the process of delineating a potential Key Biodiversity Area for sei whales in the Falkland Islands.

Co-evolution of cerebral and cerebellar expansion in cetaceans

Cetaceans possess brains that rank among the largest to have ever evolved, either in terms of absolute mass or relative to body size. Cetaceans have evolved these huge brains under relatively unique environmental conditions, making them a fascinating case study to investigate the constraints and selection pressures that shape how brains evolve. Indeed, cetaceans have some unusual neuroanatomical features, including a thin but highly folded cerebrum with low cortical neuron density, as well as many structural adaptations associated with acoustic communication. Previous reports also suggest that at least some cetaceans have an expanded cerebellum, a brain structure with wide‐ranging functions in adaptive filtering of sensory information, the control of motor actions, and cognition. Here, we report that, relative to the size of the rest of the brain, both the cerebrum and cerebellum are dramatically enlarged in cetaceans and show evidence of co‐evolution, a pattern of brain evolution that is convergent with primates. However, we also highlight several branches where cortico‐cerebellar co‐evolution may be partially decoupled, suggesting these structures can respond to independent selection pressures. Across cetaceans, we find no evidence of a simple linear relationship between either cerebrum and cerebellum size and the complexity of social ecology or acoustic communication, but do find evidence that their expansion may be associated with dietary breadth. In addition, our results suggest that major increases in both cerebrum and cerebellum size occurred early in cetacean evolution, prior to the origin of the major extant clades, and predate the evolution of echolocation.

High accuracy at low frequency: detailed behavioural classification from accelerometer data

Accelerometers are a valuable tool for studying animal behaviour and physiology where direct observation is unfeasible. However, giving biological meaning to multivariate acceleration data is challenging. Here, we describe a method that reliably classifies a large number of behaviours using tri-axial accelerometer data collected at the low sampling frequency of 1 Hz, using the dingo (Canis dingo) as an example. We used out-of-sample validation to compare the predictive performance of four commonly used classification models (random forest, k-nearest neighbour, support vector machine, and naïve Bayes). We tested the importance of predictor variable selection and moving window size for the classification of each behaviour and overall model performance. Random forests produced the highest out-of-sample classification accuracy, with our best-performing model predicting 14 behaviours with a mean accuracy of 87%. We also investigated the relationship between overall dynamic body acceleration (ODBA) and the activity level of each behaviour, given the increasing use of ODBA in ecophysiology as a proxy for energy expenditure. ODBA values for our four 'high activity' behaviours were significantly greater than all other behaviours, with an overall positive trend between ODBA and intensity of movement. We show that a random forest model of relatively low complexity can mitigate some major challenges associated with establishing meaningful ecological conclusions from acceleration data. Our approach has broad applicability to free-ranging terrestrial quadrupeds of comparable size. Our use of a low sampling frequency shows potential for deploying accelerometers over extended time periods, enabling the capture of invaluable behavioural and physiological data across different ontogenies.