Volume 8, Issue 2 p. 161-173
Research Article
Free Access

Analysis of animal accelerometer data using hidden Markov models

Vianey Leos-Barajas

Corresponding Author

Vianey Leos-Barajas

Department of Statistics, Iowa State University, Snedecor Hall, Ames, IA, 50011 USA

Correspondence author. E-mail: [email protected]Search for more papers by this author
Theoni Photopoulou

Theoni Photopoulou

Department of Statistical Sciences, Centre for Statistics in Ecology, Environment and Conservation, University of Cape Town, Cape Town, Rondebosch, 7701 South Africa

Department of Zoology, Institute for Coastal and Marine Research, Nelson Mandela Metropolitan University, Port Elizabeth, 6031 South Africa

Search for more papers by this author
Roland Langrock

Roland Langrock

Department of Business Administration and Economics, Bielefeld University, Postfach 100131, 33501 Bielefeld, Germany

Search for more papers by this author
Toby A. Patterson

Toby A. Patterson

CSIRO Oceans and Atmosphere, PO Box 1538, Hobart, Tas., 7000 Australia

Search for more papers by this author
Yuuki Y. Watanabe

Yuuki Y. Watanabe

National Institute of Polar Research, 10-3, Midori-cho, Tachikawa, Tokyo, 190-8518 Japan

SOKENDAI (The Graduate University for Advanced Studies), 10-3, Midori-cho, Tachikawa, Tokyo, 190-8518 Japan

Search for more papers by this author
Megan Murgatroyd

Megan Murgatroyd

Animal Demography Unit, Department of Biological Sciences, University of Cape Town, Cape Town, Rondebosch, 7701 South Africa

Percy FitzPatrick Institute of African Ornithology, Department of Biological Sciences, University of Cape Town, Cape Town, Rondebosch, 7701 South Africa

Search for more papers by this author
Yannis P. Papastamatiou

Yannis P. Papastamatiou

School of Biology, Scottish Oceans Institute, University of St Andrews, St Andrews, KY16 8LB UK

Department of Biological Sciences, Florida International University, 3000 NE 151st, MSB 350, North Miami, FL, 33181 USA

Search for more papers by this author
First published: 30 September 2016
Citations: 98


  1. Use of accelerometers is now widespread within animal biologging as they provide a means of measuring an animal's activity in a meaningful and quantitative way where direct observation is not possible. In sequential acceleration data, there is a natural dependence between observations of behaviour, a fact that has been largely ignored in most analyses.

  2. Analyses of acceleration data where serial dependence has been explicitly modelled have largely relied on hidden Markov models (HMMs). Depending on the aim of an analysis, an HMM can be used for state prediction or to make inferences about drivers of behaviour. For state prediction, a supervised learning approach can be applied. That is, an HMM is trained to classify unlabelled acceleration data into a finite set of pre-specified categories. An unsupervised learning approach can be used to infer new aspects of animal behaviour when biologically meaningful response variables are used, with the caveat that the states may not map to specific behaviours.

  3. We provide the details necessary to implement and assess an HMM in both the supervised and unsupervised learning context and discuss the data requirements of each case. We outline two applications to marine and aerial systems (shark and eagle) taking the unsupervised learning approach, which is more readily applicable to animal activity measured in the field. HMMs were used to infer the effects of temporal, atmospheric and tidal inputs on animal behaviour.

  4. Animal accelerometer data allow ecologists to identify important correlates and drivers of animal activity (and hence behaviour). The HMM framework is well suited to deal with the main features commonly observed in accelerometer data and can easily be extended to suit a wide range of types of animal activity data. The ability to combine direct observations of animal activity with statistical models, which account for the features of accelerometer data, offers a new way to quantify animal behaviour and energetic expenditure and to deepen our insights into individual behaviour as a constituent of populations and ecosystems.


Accelerometers are becoming more prevalent in the fields of animal and human biologging (Bao & Intille 2004; Ravi et al. 2005; Shepard et al. 2008; Altun, Barshan & Tunçel 2010). The potential of accelerometers lies in the fact that they provide a means of measuring activity in a meaningful and quantitative way where direct observation is not possible (Shepard et al. 2008; Nathan et al. 2012; Brown et al. 2013). While these instruments are cheap and compact, recording acceleration at a high temporal resolution and in up to three dimensions quickly results in terabytes of data that present various challenges regarding transmission, storage, processing and statistical modelling.

Much of the focus in the analysis of acceleration data has been on identifying patterns in the observed waveforms that correspond to a known behaviour or movement mode. This can be achieved by employing statistical classification methods and can entail observing the animal, manually assigning labels corresponding to behaviours to segments of the data and training a model using the labelled data in order to subsequently classify remaining unlabelled data. Many studies have shown the effectiveness of various machine learning algorithms for classification of human acceleration data (Bao & Intille 2004; Ravi et al. 2005; Altun, Barshan & Tunçel 2010; Mannini & Sabatini 2010). Algorithms such as support vector machines (SVM), classification trees, random forests, among others, have also recently been used for classification of animal acceleration data (Nathan et al. 2012; Carroll et al. 2014; Graf et al. 2015). For example, Nathan et al. (2012) compared the effectiveness of five machine learning algorithms to distinguish between eating, running, standing, active flight, passive flight, general preening and lying down, for griffon vultures.

Most machine learning algorithms assume independence between individual observations. However, in sequential acceleration data there is a natural dependence between observations of behaviour – once initiated, particular animal behaviours often last for periods longer than the sampling frequency. This fact has been largely ignored in most applications of classification approaches. The studies where serial dependence has been explicitly modelled have mostly relied on hidden Markov models (HMMs; Ward et al. 2006; He, Li & Tan 2007; Mannini & Sabatini 2010, 2011). HMMs are stochastic time-series models which assume that the observed time series, the so-called state-dependent process, is driven by an unobservable state process. In this scenario, the former corresponds to the acceleration data and the latter to the behavioural classes. Typically, and in common with the aforementioned machine learning approaches, in the training stage, the states of the HMM were known a priori, requiring corresponding data derived from direct observations.

There are two main difficulties with such a supervised learning approach. First, while there has been much success in classification of human acceleration data, where training data can usually be obtained with minimal effort, this may not be feasible for some animals. Humans can easily be observed in a laboratory setting, given instructions or monitored in more realistic settings, such as walking outdoors or in their home (Leenders, Sherman & Nagaraja 2000). In certain cases, animals can also be monitored in a laboratory setting (Wilson, Shepard & Liebsch 2008), but movement patterns recorded in the laboratory from free-ranging animals may not appear exactly the same as in data collected while in more natural settings. Conversely, many behaviours can only be observed in natural settings, although there has been success in using surrogate species for classification of behavioural modes (Shepard et al. 2008; Nathan et al. 2012; Brown et al. 2013; Campbell et al. 2014).

Secondly, human acceleration data have commonly been used as a tool for health monitoring and other situations where the focus is on (state) prediction, as opposed to learning how external factors drive the behaviours. Classification of behaviours alone, while certainly of interest in many scenarios, may not lead to biologically interesting inference. Once the classification has been done, the task of relating these states to environmental (and other) covariates in order to identify drivers in behaviours remains. Moreover, it is difficult to make appropriate inferential statements as the classifications are not without error, propagating the state uncertainty through to the modelled effect of the covariates.

In the supervised learning context, that is, when classification is the main purpose of an analysis, we train the HMM to recognize specific behaviours. Alternatively, HMMs can also be used in an unsupervised learning context, that is, when there are no labelled data. In an unsupervised learning context, the states are not pre-defined to represent a specific behaviour. Instead, the states are allocated such that the model captures as much as possible of the marginal distribution of the observations, that is, the distribution of an observation at a randomly chosen time point, not conditional on the previous history of the process, as well as their correlation structure. If biologically meaningful response variables from the acceleration data are considered, then the HMM states usually represent interpretable activity levels or even proxies of behavioural modes. Being data-driven, the states can be as, if not more, informative in the unsupervised learning setting than the alternatives. We can then incorporate exogenous or, where available, endogenous variable(s) of interest, to make inferential statements. HMMs and related state-switching models, in particular state-space models, have successfully been implemented to identify drivers of movement based on tracking data (Patterson et al. 2009) and can similarly be applied in the context of accelerometer data. For example, Phillips et al. (2015) applied HMMs in an unsupervised learning context to understand the behaviour of free swimming tuna from vertical movement data collected by data storage tags. We implement an unsupervised learning approach for another difficult to observe marine species, the blacktip reef shark and a volant species, the black eagle.

In this paper, we review HMM-based approaches to the analysis of animal accelerometer data. In section ‘2’, we provide an overview of accelerometer data and connect the data processing step to the HMM-based approaches described in section ‘3’. We refer to the term behavioural class, rather than differentiate between identification of specific movements (e.g. wing flapping) and behaviours (e.g. foraging). In section ‘4’, we demonstrate the use of HMMs with real data examples from marine and aerial systems.

Accelerometer data

Accelerometer devices measure up to three axes, which can be described relative to the body of the animal: longitudinal (surge), lateral (sway) and dorsoventral (heave). Acceleration recorded along one or two axes can be used to measure movement in parts of the body, for example, the mandible (Suzuki et al. 2009; Naito et al. 2010; Iwata et al. 2015), or aspects of whole body acceleration, for example, longitudinal surge (Sakamoto et al. 2009). Currently, acceleration is most commonly recorded in three axes and, to a lesser degree, in two axes (Brown et al. 2013), to measure locomotion.

Data Processing for Classification

While the observed acceleration data can be used to identify specific movements in animals, HMMs and other machine learning algorithms require more information to accurately classify the unlabelled data. These methods require appropriate features, that is, summary statistics, from a window (or sliding window) of observations. The derived features should be driven by the classes of movements that have been defined and chosen in such a way to accentuate the differences in observed acceleration measurements. There are many commonalities between the features used in applications of classification of acceleration data, though naturally no one optimal set exists (Bao & Intille 2004; Martiskainen et al. 2009; Nathan et al. 2012; Brown et al. 2013). For instance, Nathan et al. (2012) used 38 features in order to distinguish between eating, running, standing, active flight, passive flight, general preening and lying down, for griffon vultures, while Graf et al. (2015) used eight features to distinguish between standing, walking, swimming, feeding, diving and grooming of Eurasian beavers. In each case, means and variances of each of the three axes are used, as well as overall dynamic body acceleration (ODBA), the sum of dynamic body acceleration from the three axes, among others.

Connecting Measures to Behaviours

When the aim is to classify the acceleration data, data processing is driven by identifying a set of features that can be used to distinguish between specific behaviours, even if those features are not themselves interpretable as a specific behaviour when considered on their own. However, there are metrics derived from accelerometer data that, on their own, can be used as proxies for behaviour and as input to an HMM. Repeating patterns in at least one axis tend to arise from behaviours such as stroking (Sakamoto et al. 2009), flapping, running or walking (Shepard et al. 2008), whereas sudden changes, corresponding to bursts of acceleration, are often associated with prey pursuits or capture (Suzuki et al. 2009; Simon, Johnson & Madsen 2012; Watanabe & Takahashi 2013; Heerah et al. 2014; Ydesen et al. 2014), as well as predator avoidance or conflict.

In addition to behaviour, several measures can be used to summarize effort or exertion and relate acceleration to activity levels, such as ODBA (Wilson et al. 2006; Gleiss, Wilson & Shepard 2011; Elliott et al. 2013; Gleiss et al. 2013) and vectorial dynamic body acceleration (Qasem et al. 2012). Minimum specific acceleration (MSA; Simon, Johnson & Madsen 2012) can be used to disentangle the gravitational component of acceleration (static acceleration) from the movement signal or specific acceleration (also dynamic acceleration). One of the simplest and most unambiguous interpretations of static acceleration data is body posture, which in many cases can be directly interpreted as a specific behaviour (Shepard et al. 2008; Wilson, Shepard & Liebsch 2008).

Both ODBA and MSA are used to reduce the dimensionality of three-dimensional acceleration data while retaining important information (Wilson, Shepard & Liebsch 2008; Simon, Johnson & Madsen 2012). They remove the gravitational component from the acceleration signature and produce an overall value of the dynamic acceleration experienced by the animal. ODBA is derived by smoothing over a time period, for example, 1 s, making it useful for continuous data, whereas MSA is calculated pointwise (as the norm of the three vectors minus 1 for the effect of gravity) and is more suited to lower resolution acceleration data.

Analysis of accelerometer data

We first provide a brief overview of the HMM framework (section ‘3.1’). Subsequently, in section ‘3.2’, we review how HMMs can be used for state prediction, that is, classification of animal accelerometer data. In section ‘3.3’, we focus on the implementation of HMMs in a setting where the meaning of the states is driven entirely by the data and the focus lies on general inference rather than classification only.

Hidden Markov Models

An HMM is a stochastic time-series model involving two layers: an observable state-dependent process, denoted by urn:x-wiley:2041210X:media:mee312657:mee312657-math-0001 (in the univariate case), and an unobservable state process, denoted by urn:x-wiley:2041210X:media:mee312657:mee312657-math-0002. The state-dependent process models the observations, while the state process is a latent factor influencing the distribution of the observations. In our case, the observations are the accelerometer metrics considered, and the latent states are closely related to the animal's behavioural state. More specifically, the state process {Ct} takes on a finite number of possible values, 1,…,M, and its value at time t, ct, selects which of M possible component distributions generates observation yt. The Markov property is assumed for {Ct}, that is, the (behavioural) state at time t only depends on the (behavioural) state at time t − 1, such that evolution of the process over time is completely characterized by the one-step state transition probabilities. These models are natural and intuitive candidates for modelling animal accelerometer data, for two reasons: (i) they directly account for the fact that any corresponding observation will be driven by the underlying behavioural state, or general activity level, of the animal, and (ii) they accommodate serial correlation in the time series by allowing states to be persistent. HMMs seek to capture the strong autocorrelation in accelerometer data in a mechanistic way, rather than either neglecting this feature completely or only including it in a nuisance error term. HMMs can therefore be used for inference on complex temporal patterns, including the behavioural state-switching dynamics and how these are driven by environmental variables (Patterson et al. 2009; McKellar et al. 2015).

To complete the basic HMM formulation, we first summarize the probabilities of transitions between the different states in the M × M transition probability matrix (t.p.m.) urn:x-wiley:2041210X:media:mee312657:mee312657-math-0003, where urn:x-wiley:2041210X:media:mee312657:mee312657-math-0004 (for any t), urn:x-wiley:2041210X:media:mee312657:mee312657-math-0005. Note that here we are assuming that the state transition probabilities are constant over time; this assumption is relaxed in section ‘3.3’. The initial state probabilities are summarized in the row vector δ, where urn:x-wiley:2041210X:media:mee312657:mee312657-math-0006, urn:x-wiley:2041210X:media:mee312657:mee312657-math-0007.

Secondly, we need to specify state-dependent distributions (sometimes called emission distributions), urn:x-wiley:2041210X:media:mee312657:mee312657-math-0008, or more succinctly urn:x-wiley:2041210X:media:mee312657:mee312657-math-0009, for urn:x-wiley:2041210X:media:mee312657:mee312657-math-0010. These distributions can be discrete or continuous, and possibly also multivariate (in which case we write urn:x-wiley:2041210X:media:mee312657:mee312657-math-0011). Usually, the same parametric distribution is assigned to all M states, such that each state differs in terms of its associated values of the parameters. Selection is driven by the data itself, for example, count data or continuous observations.

State Prediction

Hidden Markov models provide a solid framework for the classification of data with strong serial dependence, such as sequential acceleration data, which are often processed to represent movements over a few seconds, or less, at a time (Ward et al. 2006; He, Li & Tan 2007; Mannini & Sabatini 2010). In this section, we cover the implementation and testing of an HMM when the focus of the analysis is state prediction. A full example and r code implementing this approach is provided in the Appendix S1 (Supporting Information).

State prediction can be accomplished in three manners, commonly referred to as supervised, semi-supervised or unsupervised learning. We discuss the implementation of an HMM in the supervised learning case, such that each state corresponds to one behaviour of interest, and briefly comment on the other two cases at the end of the section. Hastie, Tibshirani & Friedman (2001) detail how to split the labelled time series into training, validation and testing data, in order to estimate the prediction error. Other approaches to estimating prediction error, such as a leave-one-out cross-validation (here treating a time series as an observation), are also provided in detail.

Since the states are known, the maximum likelihood estimates (MLEs) of the HMM parameters are obtained by maximizing the complete-data likelihood, which conveniently splits into several independent parts, each of which is fairly straightforward to maximize (details provided in the Appendix S3). First, the mth entry of urn:x-wiley:2041210X:media:mee312657:mee312657-math-0012 is simply the proportion of the time series that start in state m. Secondly, the entries of the t.p.m., are estimated by
for urn:x-wiley:2041210X:media:mee312657:mee312657-math-0014. (Note this is the MLE conditional on the initial state, c1.) Finally, for each urn:x-wiley:2041210X:media:mee312657:mee312657-math-0015, the parameters of the state-dependent distribution given state m are estimated using only the observations allocated to state m. As a multivariate normal distribution (MVN) is a common choice in these cases, we cover the steps to fit the HMM with MVN state-dependent distributions in the Appendix S3 and Supporting Information. Given a fitted HMM, we can use the Viterbi algorithm to decode the most likely state sequence, thereby assigning each observation to a state, at low computational effort. Full details for state decoding are provided in Zucchini, MacDonald & Langrock (2016). The state predictions can be compared to the known states, and the proportion of correctly decoded states serves as an estimate of the prediction accuracy.

As mentioned previously, there are two other approaches to state prediction: semi-supervised and unsupervised learning. In a semi-supervised approach, classes are pre-defined, as in the supervised learning context, but there is additional flexibility provided in that the data do not have to be assigned to one of the pre-defined classes. Instead, multiple additional states can be estimated from the data. In an unsupervised learning approach, classes are not pre-defined in any manner. In these two cases, one objective can be to identify the number of distinct movement patterns exhibited by the animal, with the resulting estimated HMM states depending on the features selected for interpretation. However, as multiple movement modes can correspond to the same behaviour (e.g. foraging), interpretation of the estimated states should be made with caution. In the next section, we detail the implementation of the unsupervised learning approach where the focus is to construct biologically relevant classes of animal behaviour in order to make inferential statements.


So far, we have mostly focused on the case where there is a training sample, that is, acceleration data together with the associated behavioural states. Corresponding analyses involve training the HMM based on such labelled data and then using that HMM to categorize incoming new, unlabelled data. While certainly of interest in some settings, in practice, more often than not, labelled data will not be available but only the accelerometer data. In such unsupervised learning settings, the HMM framework can be equally useful, but is typically applied for different purposes than in classification. More specifically, the meaning of the states in such cases is often not of interest per se. Instead, an HMM is used simply as an approximate representation of the real data-generating process, and this may or may not entail that the nominal HMM states are biologically meaningful. (However, metrics derived from the accelerometer data, as described in section ‘2’, have been shown to provide insight into activity levels or correspond to classes of behaviours, such that when used as response variables in the HMM, these can lead to biologically interpretable states.) Unsupervised learning of HMMs for accelerometer data has the advantage that the states are estimated in a data-driven manner. In particular, for many of the metrics described in section ‘2’ that are connected to behaviours, assignment of classes is difficult, to say the least, especially for animals where behaviours are not well defined. These include animals which cannot be directly observed for long periods such as aquatic organisms.

There are three different possible purposes of having an approximate representation of the real process: (i) a mathematical description of the dynamics of the system (e.g. in order to have a concise description of how accelerometer measurements evolve over time, in terms of a small number of interpretable parameters and associated stochastic distributions); (ii) extraction of information (e.g. a hypothesis test on whether or not some environmental covariate increases the probability of an animal switching to a particular behavioural state); and (iii) prediction of future or missing values (e.g. behavioural state prediction given accelerometer data; Konishi & Kitagawa 2008). In the ecological literature on animal movement modelling, HMMs are used primarily to address (i) and (ii), the former in the sense that concise descriptions of movement patterns are sought and the latter in the sense that inference on the interaction of animals with their environment is drawn. In general, the ability to make inferential statements provides an avenue to answer questions about the behavioural processes, movement patterns and transitions between behaviours under different conditions or in relation to other covariates.

Addressing a research question related to aim (ii) usually involves the incorporation of covariates into the statistical model. In the HMM setting, this is commonly done at the level of the hidden states. For the general case of time-varying covariates, we define the corresponding time-dependent t.p.m. urn:x-wiley:2041210X:media:mee312657:mee312657-math-0016, where urn:x-wiley:2041210X:media:mee312657:mee312657-math-0017. The transition probabilities at time t, urn:x-wiley:2041210X:media:mee312657:mee312657-math-0018, can then be related to a vector of environmental (or other) covariates, urn:x-wiley:2041210X:media:mee312657:mee312657-math-0019, via the multinomial logit link:

Essentially, there is one multinomial logit link specification for each row of the matrix urn:x-wiley:2041210X:media:mee312657:mee312657-math-0021, and the entries on the diagonal of the matrix serve as reference categories.

While with labelled data the likelihood of interest is the complete-data likelihood, for unlabelled data the likelihood of interest is the density of the observations only, urn:x-wiley:2041210X:media:mee312657:mee312657-math-0022, the evaluation of which requires the consideration of all possible state sequences that might have given rise to these data. The powerful forward algorithm, detailed in the Appendix S3, can be applied to accomplish this, opening up a straightforward and usually feasible avenue to MLEs, namely direct numerical maximization of the likelihood. In practice, one needs to consider multiple starting values in order to make sure to have found the global maximum. The expectation–maximization algorithm provides a popular alternative route to MLEs, despite being much more technically involved and having no clear practical advantages (MacDonald 2014). Since it is our view that users are better off focusing on the simpler direct maximization approach, it is only this approach that we present in detail in the Appendix S3 and Supporting Information (for a more comprehensive introduction to maximum likelihood estimation for HMMs, see Zucchini, MacDonald & Langrock 2016).

Model selection techniques, in particular information criteria, can be used to choose an adequate family of state-dependent distributions, to select an appropriate number of states or to determine whether or not a covariate should be included in the model. However, users should not blindly follow such information criteria, especially with regard to the selection of the number of states. For animal behaviour data, it is our experience that such formal model selection approaches tend to favour models with more states than would be expected based on biological intuition, often to an extent such that selected models become near-impossible to interpret and very difficult to work with in practice (Langrock et al. 2015). One explanation for this is that often additional states are included to compensate for a model formulation that ignores some pattern in the data. These patterns can be due to the influence of an unobserved covariate, within-day variation or individual heterogeneity which is not accounted for, a violation of the Markov assumption or outliers – which usually cannot be avoided in data structures as complex as those studied here, and which may not be pertinent to the ultimate aim of the study. Further, accelerometer data are directly connected to the movement of an animal, such that an HMM with a large number of states may reflect multiple movement modes, or general classes of movement, connected to the same behavioural class, for example, foraging or active behaviour. In such cases, a healthy dose of pragmatism is required. If the choice of the number of states turns out to be difficult, then it is often useful to carefully examine all plausible models (with lower and higher numbers of states), for example, using model checking tools, in order to understand what exactly it is that the more complex models capture that is not already captured by the simpler models. Langrock et al. (2015) discuss this issue in detail, demonstrating many of the points made above in a real data example.

The HMM framework encompasses various other useful tools for drawing inference. In particular, incorporating random effects into the model formulation is crucial when there is substantial heterogeneity across multiple individuals observed. There are various ways in which this can be accomplished within the class of HMMs – see McKellar et al. (2015) and Chapter 13 in Zucchini, MacDonald & Langrock (2016) for comprehensive overviews, including discussions on the importance of acknowledging any potential heterogeneity. Furthermore, the dependence structure can be modified in various ways, for example, allowing for more complex memory in the state process without losing the ability to efficiently calculate the likelihood using the forward algorithm (Langrock et al. 2012). Assessment of the model adequacy, that is, model checking, is commonly done using (pseudo)residuals, which can reveal any notable lack of fit (Zucchini, MacDonald & Langrock 2016).

Real data examples

Modelling Activity in a Soaring Raptor

Large soaring birds, like raptors, depend on favourable meteorological conditions, as well as the underlying topography, for generation of updrafts required for low-energy flight (Pennycuick 2008). Lift availability is known to be driven largely by wind speed and temperature, as well as their interaction with the underlying topography, though other factors also contribute. Lift adequate for soaring flight is generated by two mechanisms: (i) by upward thermal convection of air warmed by solar radiation (Ákos et al. 2010) (thermal soaring) and (ii) by the movement of air over slopes and ridges in the landscape (orographic or ridge soaring).

Recently, empirical studies relating bird activity patterns to weather conditions have become possible due to advances in biologging technology that allows for collection of high-resolution movement (e.g. acceleration) data. In particular, acceleration data can be used to distinguish between different movement modes or, more simply, as a proxy of overall activity level, even if they do not correspond clearly to different behaviours (Williams et al. 2015).

An adult Verreaux's eagle Aquila verreauxii was instrumented with a remotely downloadable multisensor data-logger (UvA-BiTS, University of Amsterdam, the Netherlands; Bouten et al. 2013) in the Western Cape, South Africa, in 2013. The data-logger recorded three-dimensional acceleration (at 20 Hz) for 1 s directly after recording GPS location. The GPS location sampling rate depended on the solar-powered battery charge and thus was higher during the midday. Data were collected over nine consecutive days, with a variable amount of acceleration data sampled each day and none collected overnight.

We were primarily interested in identifying potential drivers of activity level. As such, we extracted the MSA, which serves as an index of activity, over each 1-second sample of acceleration data recorded. On average, each day produced approximately 135 observations (s.e., 23·32). Before fitting an HMM to the time series of MSA values, we first needed to resolve the irregular sampling of the acceleration data, as this is a clear violation of the HMM assumptions. The time series of MSA across days were taken to be independent and, within a day, the acceleration data were subsampled to produce one value of MSA every 112 s (Fig. 1). Only one consecutive missing value was allowed before splitting the daily MSA time series into two or more segments.

Details are in the caption following the image
Minimum specific acceleration values derived from three-axis acceleration data from a Verreaux's eagle collected over 9 days, 16–24 July 2013 (top). Minimum specific acceleration values from the 21st of April, corresponding to the shaded area in the upper plot (bottom).

The histogram of MSA values revealed two peaks close to zero, which may reflect general low-active behaviours such as roosting and preening. As we did not wish to discriminate between these two general types of behaviours, we fit a two-state HMM with state 1 represented by a mixture of gamma distributions and a gamma distribution for state 2. The fitted state-dependent densities are shown in Fig. 2, which we post hoc interpreted as low-activity and high-activity behaviour. Although we do not connect state 2 to a specific flight behaviour, such as orographic soaring, we expect that behaviours requiring more energy are reflected by larger MSA values.

Details are in the caption following the image
Histogram of minimum specific acceleration (MSA) from a Verreaux's eagle, truncated at MSA = 2, with marginal density (the distribution of observations not conditional on process history) and state-dependent densities weighted according to the proportion of observations assigned to each state (left). Unweighted state-dependent densities (top right) and close-up of the tail behaviour of the densities (bottom right). A square root coordinate transformation for the x-axis was used in all plots and for the y-axis only for the tail behaviour plot.

In order to examine the effect of wind speed and temperature on the state-switching dynamics between the two activity levels, we obtained hourly observations from the South African Weather Services (Lambert's Bay Station). The station is approximately 30 km from the general area in which the eagle was tracked, which lead to a slight spatial and temporal mismatch between the available weather data and the conditions actually experienced by the eagle. The range of temperatures and wind speeds experienced by the eagle during the study period was between 12·3 and 31·5 °C and 0 and 7·4 m/s, respectively. We allowed the entries of the t.p.m. to be a function of wind speed, temperature and their interaction. The wind-only model is written as urn:x-wiley:2041210X:media:mee312657:mee312657-math-0023, for urn:x-wiley:2041210X:media:mee312657:mee312657-math-0024, urn:x-wiley:2041210X:media:mee312657:mee312657-math-0025, urn:x-wiley:2041210X:media:mee312657:mee312657-math-0026, with the intercept term urn:x-wiley:2041210X:media:mee312657:mee312657-math-0027 reflecting the t.p.m., when wind speed is at 0 m/s. The model including wind speed alone was favoured by the Bayesian information criterion (BIC) and the full model, with temperature and the interaction term, favoured by the Akaike information criterion (AIC) (Table 1). After examination of the (pseudo)residuals of the models selected by AIC and BIC, we selected the model favoured by BIC as there was a similar lack-of-fit evident in both models. Further, we may not have captured a large enough range of temperatures in order to make general inferences about its effect on the activity levels of the eagle and as such were cautious of overfitting or overinterpreting the model results. We present confidence intervals and a plot of the (pseudo)residuals for assessment of goodness-of-fit in the Appendix S3 for the model with only wind speed included. r code to simulate MSA data and fit a two-state HMM with the t.p.m., entries as functions of wind speed is included in the Supporting Information.

Table 1. Model fitting results for the Verreaux's eagle
Model Log likelihood AIC ∆AIC BIC ∆BIC
No covariates 2000·2 −3980·4 21·6 −3929·4 21·6
Temperature 2001·9 −3979·9 22·1 −3918·6 17·0
Wind speed 2010·4 −3996·9 5·1 −3935·6 0
Wind speed, temperature 2011·6 −3995·2 6·8 −3923·7 11·9
Wind speed, temperature, wind speed × temperature 2017·0 −4002·0 0 −3920·3 15·3
  • Based on the AIC, the model selected is the full model including wind speed, temperature and their interaction. Based on the BIC, the model selected includes only wind speed.
  • AIC, Akaike information criterion; BIC, Bayesian information criterion.

The estimated state transition probabilities suggest that, as wind speed increases, (i) the eagle has a very slightly increased chance of switching to the high-activity state when in the low-activity state and (ii) spends much longer periods of time, on average, in the active state. As a consequence, the equilibrium (stationary) distribution for fixed wind speeds (Patterson et al. 2009) indicates that the eagle spends more time in the active state overall as wind speed increases (Fig. 3). Windier conditions favour orographic soaring, as demonstrated by studies on migrating golden eagles Aquila chrysaetos, which is a more active behaviour (Lanzone et al. 2012). There is also theoretical evidence to suggest that, in general, flying is more energetically demanding in high winds (Pennycuick 1972).

Details are in the caption following the image
For the Verreaux's eagle example, estimated state-dwell probabilities (probability of remaining in a state) as a function of wind speed (left) and estimated equilibrium state probabilities (marginal probability of a state at a fixed value of the covariate) as a function of wind speed (right).

Diel Activity Changes in a Reef-associated Shark

Many species of shark are upper trophic level predators, which may serve an important role in marine ecosystems. However, determining the intensity of their predatory behaviour requires modelling the temporal component as their activity levels are likely to follow a diel and/or tidal cycle (Gleiss et al. 2013; Papastamatiou et al. 2015). Acceleration sensors provide a direct measure of activity; however, many species of shark swim continuously making it difficult to define specific behaviours (e.g. they are never truly at rest), making conventional classification methods problematic. HMMs can identify changes in behavioural states and how these may be related to time of day, tidal state, swimming depth or water temperature. To demonstrate this, we applied HMMs to accelerometry data obtained from a free-ranging blacktip reef shark Carcharhinus melanopterus at Palmyra Atoll in the central Pacific Ocean (data taken from Papastamatiou et al. 2015). A multisensor package was attached to the dorsal fin of a 117-cm female shark. The multisensor data-logger (ORI400-D3GT, Little Leonardo, Tokyo, Japan) recorded three-dimensional acceleration (at 20 Hz), depth and water temperature (at 1 Hz) and was embedded in a foam float which detached from the shark after 4 days (Papastamatiou et al. 2015). The package also contained a VHF transmitter allowing recovery at the surface after detachment.

In order to examine active behaviour, we calculated the average ODBA of the shark over 1-second intervals, which resulted in 321 815 observations (after removing the first four hours of data). Figure 4 displays the ODBA time series of 1 day. Compared to metrics such as tail-beat frequency, ODBA has the advantage of measuring change in behaviour in all axes. For example, if the shark is nose down at the seafloor, attempting to capture prey, its tail-beat frequency may be low but it is still active (Watanabe et al. 2012). As we are interested in the times of day the shark was more active, as well as tide effects, we applied a two-state HMM with one state post hoc interpreted as representing less active behaviour and the other more active behaviour.

Details are in the caption following the image
Overall dynamic body acceleration values from a blacktip reef shark, averaged over 1-second intervals, for 13 July 2013.

Although there are clear spikes in ODBA that point to higher energetic activities, various combinations of parametric distributions for state 1 and 2 led to vastly different state-dependent densities. Further, the ODBA values had many extreme values that needed to be accommodated, which further increased the difficulties of selecting appropriate state-dependent distributions. As ODBA is not a metric that can easily be divided into active/inactive behaviours in sharks, we estimated the state-dependent densities nonparametrically, in both states, in order to minimize the bias introduced by assigning inadequate parametric distributions (Fig. 5; Langrock et al. 2015).

Details are in the caption following the image
Histogram of overall dynamic body acceleration (ODBA) from a blacktip reef shark, truncated at ODBA = 2, with marginal density and state-dependent densities weighted according to the proportion of observations assigned to each state. (left). Unweighted state-dependent densities (top right) and close-up of the tail behaviour of the densities (bottom right). A square root coordinate transformation for the x-axis was used in all plots and for the y-axis only for the tail behaviour plot.
To examine potential diel and tide effects on activity levels, we let the entries of the t.p.m., be functions of up to two covariates: time of day and tide level (ebb, flood, low and high). Tide data were obtained from the NOAA tides and currents website for Palmyra Atoll and was processed by denoting high or low tide as ±1 h from reported high or low tide times. Time of day is represented by two trigonometric functions with period 24 h, cos(2πt/86 400) and sin(2πt/86 400) (86 400 is the number of seconds in a day). We use three indicator variables, urn:x-wiley:2041210X:media:mee312657:mee312657-math-0028 and urn:x-wiley:2041210X:media:mee312657:mee312657-math-0029, for tide levels high, flood and ebb, respectively, such that urn:x-wiley:2041210X:media:mee312657:mee312657-math-0030 when tide level is high and urn:x-wiley:2041210X:media:mee312657:mee312657-math-0031 otherwise, and so on, which gives the entries of the t.p.m., the following form
for urn:x-wiley:2041210X:media:mee312657:mee312657-math-0033, urn:x-wiley:2041210X:media:mee312657:mee312657-math-0034, urn:x-wiley:2041210X:media:mee312657:mee312657-math-0035. The intercept term urn:x-wiley:2041210X:media:mee312657:mee312657-math-0036 corresponds to low tide.

Based on the selected model (Table 2), with confidence intervals and (pseudo)residuals provided in the Appendix S3, the shark's activity levels were, on average, lowest from approximately 9:00–13:00 and highest from 21:00–1:00. In Fig. 6, we see that the shark was more active during high tide in general when compared to flood, low or ebb tide. While the equilibrium (or stationary) distribution associated with low and ebb tide overlaps, the state-dwell probabilities, that is, the diagonal entries of the t.p.m. corresponding to the probability of remaining in the same state, are higher during ebb tide than in low tide. Naturally in a short time series, the tide levels are correlated with certain times of the day, but a longer time series or a joint modelling of multiple time series, with tide levels observed during all times of day, can provide robust estimates of the effect of tide on activity level using the HMM formulation provided here.

Table 2. Model fitting results for a blacktip reef shark
Model Log likelihood AIC ∆AIC BIC ∆BIC
No covariates 639 299·2 −1 278 370 779 −1 277 178 692
Time 639 558·1 −1 278 872 277·2 −1 277 645 225
Time, high 639 657·6 −1 279 063 86·2 −1 277 819 51
Time, high, flood 639 695·2 −1 279 130 19 −1 277 869 1
Time, high, flood, ebb 639 708·7 −1 279 149 0 −1 277 870 0
  • Based on the AIC and BIC, the model selected includes time of day and includes differences in activity levels based on all categories of tide levels.
  • AIC, Akaike information criterion; BIC, Bayesian information criterion.
Details are in the caption following the image
Implied stationary distribution for state 2, the more active state, by time of day and tide level for the blacktip reef shark example. For tide levels, we distinguish between model estimates, such that the corresponding tide level was observed at that time of day, and forecasts, where we did not observe the tide level at that time of day.

Using the Viterbi algorithm, we decoded the optimal state sequence to underlie the ODBA time series. To further understand the effect of vertical habitat on behaviour, we related the decoded state sequence to a grid of depth and temperature values, shown in Fig. 7. The shark spent most of its time over the nearly five-day period in depths of about 3–6 m and between 28 and 29 °C, with some higher counts also in shallower waters, which is reflected in the state 2 counts. However, the percentages of state 2 observations reveal that the shark was generally more active when near the surface in waters of 28–29 °C. There was generally less active behaviour exhibited when the individual was in very shallow warm water (>29 °C).

Details are in the caption following the image
For the blacktip reef shark example, the number of observations in each grid cell that correspond to state 2. Zero counts appear in white (left). Percentage of observations in each cell that correspond to state 2 (right).


We detailed two approaches for analysing animal accelerometer data with HMMs: a supervised learning approach for state prediction, such that classification is of primary interest, and an unsupervised learning approach, where the states reflect biologically meaningful classes of behaviour, in order to infer drivers of animal behaviour. The aim of a study and the type of data available determine which of the two is to be preferred. When the objective is to do classification and there is a set of pre-defined behaviours of interest, then the model's ability to correctly predict and categorize behaviours is of main interest. In this instance, a supervised learning approach may be applied. One of the benefits of such an approach is that the behavioural classes are exactly defined, making interpretation relatively straightforward. Alternatively, if the objective is to infer (or, colloquially speaking, to ‘learn’) new aspects of animal behaviour, then the unsupervised learning approach provides an excellent framework. The latter comes with the implicit caveat that the states do not necessarily map directly to specific animal behaviours. Any post hoc behavioural interpretation of the estimated states is directly connected to the metric(s) used and must draw from background biological knowledge of the species of interest. In many cases, behaviours such as foraging may not be exclusive to one state or another. Nonetheless, if the model is able to identify bouts of behaviour which consistently reappear, then it is often likely that these signify something important in the animal's behavioural repertoire and are worthy of further investigation.

Even when classification is the goal of an analysis, there are certainly practical scenarios which preclude the use of an HMM, for example, if the training data do not reflect the transitions between behaviours or if there are insufficient data. Moreover, multiple studies have shown that other machine learning algorithms, for example, SVMs or random forests, can work well for classification of animal acceleration data (Martiskainen et al. 2009; Nathan et al. 2012; Carroll et al. 2014; Graf et al. 2015). However, disregarding the serial dependence in the acceleration data usually is an unrealistic assumption, which often goes unmentioned or is treated as an afterthought. Adopting the assumption of independence is particularly risky if inferential statistics are applied to the output of a machine learning algorithm. In these cases, secondarily applied statistical tests implicitly assume that the machine learning categorizations contain more information content than is warranted, potentially leading to spurious results. This is not just a statistical nuance and can be a crucial point. Such tests are often applied as decision making tools to sort out ‘what matters’ and setting the direction for much further research effort. Also, in assuming independence, one allows for classifications that may not be biologically realistic or must filter the classifications to properly identify a specific behaviour. For instance, Carroll et al. (2014) used a SVM where one of the primary interests was to identify prey handling/capture for penguins at sea. To confirm a prey capture event, they ruled that if the SVM classified three consecutive observations as prey handling, this counted as a true prey capture. In contrast, an HMM would have bypassed the need to filter through the classification results by accounting for the serial dependence in observations corresponding to prey handling. In general, many behaviours persist over longer stretches of time than those at which the data are processed, also necessitating the use of a model that can account for the serial dependence. It may be difficult for any machine learning algorithm that assumes independence to properly classify a sequence of observations into the same class, unless the boundaries between classes are well defined. In the context of recognition tasks, for example, speech or pattern recognition, HMMs have proven to be extremely successful tools for classification precisely because they do account for the serial dependence in the signal of interest (Rabiner 1989).

In the literature, inference on behavioural state-switching dynamics has sometimes been made using two-stage (or even three-stage) analyses, where HMMs (or other machine learning algorithms) are used to decode the behaviours underlying given observations, and subsequently a logistic regression is conducted for relating the decoded behaviours to covariates (Hart et al. 2010; Broekhuis et al. 2014). The appeal of such an approach lies in the ease of implementation: fairly basic HMMs, without covariates, are fitted to the accelerometer data and used to decode the states, and, subsequently, standard regression software packages can be used to conduct a regression of the behavioural states on covariates. However, it is our view that such a multistage analysis is less suited to relating accelerometer data to covariates than the joint modelling approach presented in section ‘3.3’, for two reasons: (i) in the multistage analyses, the uncertainty in state estimates is usually not propagated through the different stages of analysis, and (ii) a regression analysis on decoded states needs to take into account the high serial correlation in those states. Rather than ignoring these issues or trying to address them within a multistage analysis (which render such an approach technically challenging), a direct joint modelling approach, where neither of the problems arise, seems preferable.

Using a direct joint modelling approach in section ‘4’, we were able to learn about the effects that atmospheric variables have on activity levels of a soaring raptor, while for the blacktip reef shark we examined temporal and tidal inputs effects on its activity levels. The HMM produced similar temporal patterns of activity to a previous analysis of the blacktip reef shark data set using generalized additive mixed models (Papastamatiou et al. 2015). Both analytical methods revealed crepuscular and/or nocturnal increases in activity with a tidal component, with the shark most active at the high tide or as tide was about to ebb. By incorporating swimming depth and temperature, it was also revealed that highest activity was seen when the shark was at the surface in waters of 28–29 °C. More importantly, the analysis showed that the shark was inactive when in very warm (>29 °C) shallow water or deeper water. These results agree with a previous hypothesis that sharks are ‘hunting warm, and resting warmer’ and use warmer water (>29 °C) to increase the rate of some physiological function such as digestion, and not for foraging (Papastamatiou et al. 2015). The HMM in this case allows us to explain the drivers of activity in the shark and move beyond just describing its movements, but rather explain ‘why’ it may be moving or selecting certain habitats. The HMM also provided a measure of the change in probability of the individual being in active states. Although there was a clear temporal pattern of activity, the HMM identified the shark as 30% more likely to be in an active state during the late evening hours. For the adult black eagle, the HMM provided a direct modelling approach to examine the effect of wind speed and temperature on its activity level. The results suggests that the black eagle spent more time in the relatively active state overall and was more likely active in windier conditions. These results are in line with theoretical (Pennycuick 1972) and empirical (Lanzone et al. 2012) studies.

We have covered the basic HMM framework here, but the popularity of the HMM framework is due in part to its many extensions. In particular, there are two HMM extensions that have been proven useful in classification of human activities: the hidden semi-Markov model (HSMM) (Langrock & Zucchini 2012) and the hierarchical HHMM (Fine, Singer & Tishby 1998). The HSMM models the time spent within a state by some probability distribution with support on the positive real integers, thereby allowing for more complex state-dwell-time distributions than can be provided by an HMM (namely only geometric distributions). For instance, an HMM may not model the time spent in a resting behaviour adequately if the animal is known to rest for long periods of time. The HHMM provides the framework necessary to identify composite behaviours. For instance, lunge feeding in baleen whales is a composite behaviour made up of (i) initial increase in acceleration with (ii) a positive pitch angle, as animals commonly approach prey schools from below, followed by (iii) a rapid deceleration after the whale opens its mouth increasing its drag (Owen et al. 2015). The HHMM models each composite behaviour as its own HMM and models the transitions between composite behaviours, that is, switches between HMMs.


Y.Y.W. was funded by Grants-in-Aid for Scientific Research from the Japan Society for the Promotion of Science (grant reference 25850138). T.P. was supported by a South African National Research Foundation Scarce Skills Postdoctoral Research Fellowship, and T.P. and Y.P.P. received funding from the MASTS pooling initiative (The Marine Alliance for Science and Technology for Scotland) and their support is gratefully acknowledged. MASTS is funded by the Scottish Funding Council (grant reference HR09011) and contributing institutions. T.P. and M.M. gratefully acknowledge the hardware, software, support and expertise contributed by Prof. Willem Bouten and his research group UvA-BiTS (University of Amsterdam Bird Tracking System). Special thanks to J. Caselle for supporting all field work at Palmyra atoll. The authors are grateful to Orr Spiegel, the Associate Editor and an anonymous referee for their useful comments that substantially improved the manuscript.

    Data accessibility

    Data available from the Dryad Digital Repository http://dx.doi.org/10.5061/dryad.6bm2c (Leos-Barajas et al. 2016).