^{1}

^{2}

^{3}

^{*}

^{4}

^{5}

^{6}

^{1}

^{2}

^{5}

^{3}

The authors have declared that no competing interests exist.

Conceived and designed the experiments: MAS IDJ DJFR MFB. Performed the experiments: MAS DJFR. Analyzed the data: MAS. Contributed reagents/materials/analysis tools: MAS IDJ DJFR DT. Wrote the paper: MAS. Reviewed the manuscript and approved the final version: IDJ DJFR RP MFB.

Argos recently implemented a new algorithm to calculate locations of satellite-tracked animals that uses a Kalman filter (KF). The KF algorithm is reported to increase the number and accuracy of estimated positions over the traditional Least Squares (LS) algorithm, with potential advantages to the application of state-space methods to model animal movement data. We tested the performance of two Bayesian state-space models (SSMs) fitted to satellite tracking data processed with KF algorithm. Tracks from 7 harbour seals (

The collection of individual animal movement data has become widely utilized by ecologists in the last decade due to the improvement of the underlying technologies and reduction of operational costs involved in animal telemetry. Of the several technologies available, one of the most popular is that based on satellite tags (platform transmitter terminals, PTTs) using the Argos system

Varying accuracy and precision, and unevenness in space and time of telemetry data can affect the determination of distribution, habitat use and behavioural patterns of animals and severely bias the calculation of movement metrics

Jonsen et al.

Geolocation of animals tracked with Argos systems is based on the Doppler shift of the tag’s fixed transmission frequency; i.e. the frequency shift of the tag’s signal received at the orbiting satellite as it approaches and moves away from the tag

In May 2011, Argos implemented a new algorithm that accounts for movement dynamics and uses a Kalman filter (KF) to estimate positions

Although the new processing algorithm may bring significant advantages, it may also introduce changes in the autocorrelation structure of the Argos satellite data. Given that many published SSM applications for animal tracking data do not currently account for the potential autocorrelation in location errors introduced by the new KF algorithm, models fit to datasets with differing degrees of autocorrelated errors could lead to biased estimates of movement parameters, behavioural states, and their uncertainties. Several studies have examined the validity of SSMs applied to data obtained with the LS positioning algorithm and quantified the precision of predicted locations (e.g.

Our aim is to assess the performance of Bayesian SSMs fit to satellite tracking data processed with the new KF positioning algorithm introduced by Argos. We use two real datasets from marine taxa that differ greatly in their movement ranges– harbour seals (

All seal handling and tagging procedures were carried out under license number 60/4009 issued by the UK Home Office under the Animals (Scientific Procedures) Act 1986.Whale tagging was approved by the Regional Directorate for Sea Affairs, Autonomous Region of the Azores under research permits 20/2009/DRA and 16/2010/DRA. All procedures in whales followed the guidelines of the American Society of Mammalogists

In the interest of clarity we’ll use the following terminology throughout the paper: i) LS locations/data and KF locations/data refer to the locations/data provided by Argos that were derived from the application of the LS and KF algorithms, respectively; and ii) LS or KF model refer to the state-space models fit to data derived from the application of either the LS or KF algorithm. As explained below, the same models were fit to LS and KF datasets.

GPS/Argos tags were deployed on harbour seals in the Eden Estuary, south-east Scotland and around Eday, Orkney between May and July 2012. Animals were caught on or close to haul-out sites using hand, seine or tangle nets and subsequently anesthetised with Zoletil as detailed in Sharples et al.

The Fastloc GPS data used in this study were transmitted via the Argos system, providing high resolution at sea locations. The Argos transmissions also generated a concurrent series of standard Argos locations. At our request, messages from the satellite transmitters were processed by the Argos service provider (CLS, Ramonville Saint-Agne, France) using both the LS and KF processing algorithms.

Fastloc GPS positions are more accurate and precise than Argos locations and in the present study were assumed to represent the seals’ “true” position. However, GPS accuracy is known to decrease when Fastloc calculations are based on fewer satellites

As central-place foragers harbour seals haul-out on land between foraging trips. Thus, we needed to remove haul-out locations from the data before fitting any models. Although the GPS/Argos tags have a wet/dry sensor which records haul-out events, only a subset of these records are received via the Argos system. These animals often range in near shore waters and the large measurement error in Argos observations means such observations could not be used to define whether a location fell on land. Thus, we used the Fastloc GPS positions to define the precise time seals departed and returned to land. Positions within 200 m from all shorelines were also considered as haul-out to buffer against errors in GPS positions and because harbour seals haul-out on intertidal sandbanks. This procedure may have excluded valid parts of a few foraging trips but this shouldn’t affect algorithm comparison in anyway. Consecutive at sea locations between haul-out events thus formed an individual foraging trip. We defined a series of trips within each seal GPS track and, for each trip, we selected all LS and KF locations obtained between 5 minutes prior to and 5 minutes after the trip. Only trips with ≥30 LS and KF locations were subsequently used for model fitting. The seal dataset analysed in the next sections consisted of 1174 GPS, 1339 Argos LS and 2083 Argos KF positions obtained during 31 foraging trips of 7 different seals (

The data consisted of Argos-derived surface positions obtained from PTTs (model SPOT5-implantable, Wildlife Computers, Redmond, Washington, USA) attached to the flanks of 6 fin whales. Whales were tagged off Faial and Pico islands (38°N 28°W), Archipelago of the Azores (Portugal), in September 2009, April and May 2010. All tags were programmed to transmit on a daily basis, every hour of the day up to a maximum of 500 messages per day. Details about the tagging methodology, movements and inferred behaviours of these whales are described in Silva et al.

The KF algorithm consistently yielded more positions per individual whale than the LS algorithm (

State-space models couple two stochastic models: a process model (transition equation) that estimates the current state (e.g. location and behavioural state) of an animal given its previous state, and an observation model that relates the unobserved location states estimated by the process model to the observed data (locations obtained from Argos).

The SSM described in Jonsen et al.

We initially attempted to fit a SSSM to the harbour seal data but encountered the same problems noted by Breed et al.

We therefore chose to fit a SSM

By letting _{t}_{-1} is the displacement between unobserved locations _{t}_{-1} and _{t}_{-2,} and _{t}_{t}_{t}_{-1}. _{t}_{-1} to _{t}_{2} is a bivariate Gaussian distribution with covariance matrix ∑ and represents the randomness in animal movement.

The observation equation accounts for the irregularity and variable errors in the observed Argos locations. Errors in latitude and longitude are modelled with a t-distribution using independent parameter estimates derived for each Argos location class

We fitted the Bayesian switching state-space model (SSSM) described in Jonsen et al.

The observation equation used to model the irregularly observed LS and KF fin whale locations was that same used for the SSM.

Models were fit using R (R Development Core Team 2012) code provided in the supplement to Jonsen et al.

The hSSM was fitted separately to the harbour seals’ location data (excluding Z class locations) obtained from each algorithm using a time step of 2 hours, corresponding to the average temporal resolution of the LS data. For the hSSM fit to the KF and LS satellite datasets, we ran two MCMC chains for 60000 iterations, dropping the first 50000 samples as a burn-in and retaining every 10^{th} sample from the remaining 10000 assumed post-converge samples from each chain to reduced sample autocorrelation. Thus, model parameters and estimates of seals’ locations were calculated using a total of 2000 MCMC samples.

The SSSM was fitted separately to the fin whales’ data obtained from each algorithm (after removing Z class positions from both datasets) using a time step of 3 hours, corresponding to the average temporal resolution of the LS data. For each SSSM we ran two MCMC chains for 45000 iterations, discarding the first 40000 samples and retaining every 5^{th} from the remaining 5000 samples from each chain. A total of 2000 MCMC samples were used to calculate model parameters and estimates of whales’ locations and behaviours.

hSSM and SSSM convergence and sample autocorrelation were assessed by visually inspecting trace and autocorrelation plots and using the Gelman and Rubin scale reduction factor (R-hat) diagnostic available in R package boa.

The Argos locations per seal trip greatly exceeded those of Fastloc GPS, and the latter were also more irregular in time (

To investigate if and how the quality of Argos telemetry data affects spatial accuracy of LS and KF models, we compared location errors from seal trips with different temporal resolutions, spatial precisions and frequency of observations. We used linear mixed-effects models with seal and individual trip as random effects to account for behavioural differences among seals and unequal sample sizes across trips. Errors were log transformed to ensure linearity with continuous predictors. Algorithm (LS vs. KF) was included in the model as a categorical predictor and continuous predictors were number of Argos locations used to fit the model, average length of time between locations (hereafter time step), and proportion of positions of LC 0, A and B (hereafter LC 0-B). Values of these continuous predictors for each seal trip are given in

In the case of the SSSM fit to the whale data, we could only determine how well the KF models performed in relation to models fit to the LS algorithm. For each whale, we compared the medians, inter-quartiles and 95% credible limits (95% CL) of parameter estimates of LS and KF models. We also calculated the longitudinal and latitudinal differences between pairs of location estimates from the LS and KF models for each whale. For each location predicted by the LS model we estimated a probability ellipse determined by the 95% CL obtained from the model. We then calculated the proportion of location estimates from the reduced KF model that fell within the 95% probability ellipse of the corresponding LS position.

To understand if the KF algorithm introduced significant changes in the ability of the SSSM to resolve behavioural state, we calculated percentage of agreement in behavioural classification between the LS and KF models. Whale behaviour at each 3-h location was inferred from the output of the SSSM. Because behaviour is treated as a binary variable, MCMC samples can only assume the values 1 (inferred as transiting) or 2 (inferred as ARS),

Finally, we investigated how the whale tracks from a model fitted to the full KF dataset compared to those from the models applied to LS data. We fitted the SSSM to the full KF data using the same time step as above. For each whale we calculated the distance (in km) from locations estimated by the full KF model to the track estimated by the LS model. We compared only data from days when both methods delivered satellite locations.

Means are presented ± standard deviation (SD) throughout. All distances were calculated using a great-circle route. Statistical analyses were performed in R software using packages nlme and MASS.

The KF algorithm provided 2083 locations, 1.5 times more than the LS algorithm and 1.8 times more than the GPS transmitted via Argos (

Seal | Trip | LS model | KF model | Variationin mean errors (%) |
||||

N |
Mean error (km) | Error range (km) | N |
Mean error (km) | Error range (km) | |||

1545 | 11 | 8 | 2.6±1.7 | 0.4–5.1 | 8 | 1.7±1.4 | 0.1–3.9 | −34 |

21 | 9 | 1.9±2.0 | 0.1–5.6 | 8 | 0.8±0.4 | 0.4–1.5 | −57 | |

23 | 5 | 2.2±1.3 | 1.2–4.3 | 6 | 0.9±0.3 | 0.6–1.4 | −56 | |

27 | 14 | 1.9±2.3 | 0.1–8.6 | 15 | 2.2±2.4 | 0.1–7.9 | 16 | |

28 | 13 | 1.5±2.1 | 0.1–7.6 | 13 | 1.5±1.5 | 0.1–5.1 | −2 | |

28503 | 11 | 12 | 2.2±1.9 | 0.2–6.9 | 14 | 2.0±2.3 | 0.2–8.5 | −10 |

18 | 12 | 3.9±2.4 | 1.7–9.4 | 11 | 3.1±2.7 | 0.3–9.3 | −23 | |

19 | 14 | 4.5±3.1 | 0.2–10.5 | 14 | 4.4±3.4 | 0.8–13.9 | −3 | |

23 | 8 | 4.8±1.8 | 1.6–6.8 | 8 | 4.2±2.9 | 0.9–8.9 | −13 | |

42 | 5 | 2.9±1.7 | 1.0–5.5 | 6 | 2.4±1.1 | 1.1–3.8 | −19 | |

43844 | 4 | 15 | 2.5±2.6 | 0.1–10.0 | 14 | 1.6±1.6 | 0.1–5.7 | −36 |

8 | 8 | 1.7±1.5 | 0.2–4.6 | 8 | 1.2±0.6 | 0.2–1.9 | −29 | |

14 | 10 | 1.7±2.4 | 0.1–8.0 | 12 | 1.4±1.7 | 0.1–6.4 | −23 | |

16 | 13 | 1.5±1.5 | 0.2–6.3 | 11 | 1.0±0.7 | 0.2–2.3 | −32 | |

22 | 14 | 4.9±3.7 | 0.9–13.6 | 14 | 2.5±2.7 | 0.2–9.0 | −48 | |

43871 | 7 | 24 | 3.6±3.5 | 0.3–12.7 | 23 | 3.4±3.4 | 0.3–11.7 | −5 |

8 | 17 | 5.9±3.6 | 0.8–11.8 | 17 | 4.3±3.2 | 0.7–10.1 | −27 | |

13 | 21 | 4.9±3.7 | 0.3–10.8 | 20 | 3.0±2.5 | 0.1–9.4 | −39 | |

19 | 20 | 5.9±2.9 | 0.6–11.5 | 20 | 4.5±2.8 | 0.2–10.6 | −24 | |

120346 | 24 | 13 | 1.9±1.4 | 0.5–5.6 | 13 | 1.2±0.7 | 0.3–2.6 | −37 |

25 | 15 | 2.2±1.7 | 0.4–5.4 | 15 | 1.6±1.3 | 0.2–4.1 | −31 | |

26 | 5 | 4.3±2.9 | 1.3–8.4 | 5 | 3.0±3.2 | 0.9–8.7 | −30 | |

30 | 13 | 2.1±1.4 | 0.4–4.9 | 13 | 1.1±1.0 | 0.1–3.2 | −48 | |

32 | 7 | 1.7±1.4 | 0.6–4.1 | 7 | 1.7±1.9 | 0.2–5.9 | −1 | |

120349 | 3 | 8 | 4.2±2.3 | 1.2–7.1 | 8 | 4.5±3.5 | 0.6–10.8 | 6 |

4 | 8 | 3.5±2.8 | 0.4–7.9 | 8 | 4.2±3.9 | 0.9–12.2 | 20 | |

5 | 5 | 3.7±1.7 | 2.0–6.1 | 5 | 1.9±1.7 | 0.8–5.0 | −48 | |

6 | 10 | 7.7±3.5 | 1.1–11.1 | 10 | 5.0±3.0 | 0.5–8.8 | −35 | |

120350 | 3 | 13 | 3.7±3.1 | 0.5–9.4 | 13 | 3.1±2.3 | 0.1–3.9 | −16 |

4 | 14 | 3.2±3.1 | 0.3–12.5 | 13 | 5.0±4.5 | 0.4–1.5 | 56 | |

5 | 15 | 4.2±2.4 | 0.7–7.5 | 23 | 4.9±4.3 | 0.6–1.4 | 17 | |

Total | 368 | 3.5±3.0 | 375 | 2.9±2.9 |

N: Number of locations used to calculate errors in locations estimated from LS and KF models.

Errors in locations estimated from LS and KF models showed the same elliptical distribution in relation to interpolated GPS positions, with a clear directional bias in the longitudinal error component (

Errors in harbour seal locations estimated from state-space models fit to Least Squares (LS) (black) and Kalman filtered (KF) (red) data are plotted as offsets from “true” GPS positions. Standard ellipses were fitted to 95% of LS (black line) and KF (red line) error points.

Two representative tracks of foraging trips reconstructed using GPS positions, and LS and KF modelled locations are shown in

Estimated locations (circles) and tracks (lines) of harbour seals obtained from fitting state-space models to Least Squares (LS) (black) and Kalman filtered (KF) (red) data, in relation to the “true” GPS positions and track (yellow). A. Example of a trip with higher quality of Argos data: trip 7 of harbour seal #43871. B. Example of a trip with lower quality of Argos data: trip 22 of harbour seal #43844.

Uncertainty in KF model estimates, as indicated by the width of the 95% CL (measured in km), was significantly lower than that of LS model estimates (KF model: 5.6±5.6 km; LS model: 11.6±8.4 km; t-test = −11.41, df = 741,

Observation frequency, temporal resolution and spatial precision of Argos data used to fit the SSMs varied among seals and trips and between the LS and KF models (

Relationship between mean errors (±SD shown as vertical bars) in locations estimated from state-space models fit to Least Squares (LS) (black) and Kalman filtered (KF) (red) data per harbour seal trip and quality of Argos telemetry data used to fit the models: A–B. Number of locations. C–D. Time step (h) between locations. E–F. Proportion of locations of LC 0-B. Different trips from the same seal have the same symbol.

Mean errors (±SD) of LS and KF modelled trips were plotted in relation to the Argos quality parameters described above (

We fitted a linear mixed-effects model to examine the effects of type of algorithm and of Argos quality parameters (spatial precision, observation frequency and time step) on estimated errors of modelled locations. The interactions between algorithm and the continuous predictors were the first to be dropped from the linear mixed-effects model based on AIC results, suggesting that quality of Argos data influenced the accuracy of LS and KF models in a similar way. The best fitting model indicated that observation frequency and time step of Argos data had no effect on the errors of locations estimated from the models, and only algorithm and proportion of locations LC 0-B were significant (^{−7},

Predicted error in harbour seal locations according to the best fitting linear mixed-effects model for A. State-space models fit to Kalman filtered (KF) data. B. State-space models fit to Least Squares (LS) data.

Medians and 95% CL of estimated model parameters of the reduced dataset were similar across whales and between the LS and KF algorithms. Both the LS and KF models distinguished well between the two behavioural modes (transiting and ARS), as indicated by the parameter estimates that aggregated into two non-overlapping groups.

The estimated locations inferred from the KF model applied to the reduced dataset differed little from the locations output by the LS model. Differences in latitude and longitude between paired KF-LS locations were centred around zero but the latter showed a wider range of values (range for latitude: −1.1–0.7°; range for longitude: −1.2–2.0°) (

Differences in locations estimated from switching state-space models fit to Kalman filtered (KF) (red dots) data are plotted as offsets from locations calculated from the same models fit to Least Squares (LS) data. Standard ellipses were fitted to 95% of KF data points. A. Fin whales #80702 (red), #80704 (blue) and #80707 (green). B. Fin whales #80713 (black), #89969 (orange). C. Fin whale #80716 (pink).

The proportion of estimated locations from the SSSM applied to the reduced KF data that fell within the 95% probability ellipse of locations inferred by the LS model varied between whales but was very high, ranging from 69 to 100% (mean = 88%). We also compared differences in the width (measured in km) of the 95% CL between pairs of locations estimated from the model fit to the reduced KF data and the LS data. For five whales, the reduced KF model resulted in lower average widths of 95% CL (paired t-test:

In 94% of the cases, the behavioural mode inferred by the KF model matched the classification from the model fit to the LS data (

KF model | ||||

Transit | ARS |
uncertain | ||

LS model | Transit | 353 | 0 | 6 |

ARS | 0 | 524 | 40 | |

uncertain | 5 | 6 | 83 |

As expected, the KF processing algorithm yielded more positions and improved the temporal resolution of the 6 whale tracks. The increase in number of locations per track ranged from 18 to 272% with an average of 75%. The average number of daily locations per whale track varied between 6.0–38.6 for the full KF data, compared to 1.6–30.8 for the LS data (

The width of the 95% CL of locations estimated by the full KF model (47.3±76.9 km) was significantly lower than the width of 95% CL of locations estimated from the LS model (57.2±113.0 km) (t = 2.38,

Estimated locations (circles) and tracks (lines) of fin whale #89969 obtained from fitting a switching state-space model to Least Squares (LS) (black) and the full Kalman filtered (KF) (red) data. The 95% probability ellipses of locations derived from the LS-based model are shown in green. A. Complete tracks showing the increase in track length resulting from the application of the KF algorithm (red). B, C, D. Detail of the tracks showing the majority of KF locations within the 95% probability ellipses of LS locations.

Estimated locations (circles) and tracks (lines) of fin whale #89969 obtained from fitting a switching state-space model to Least Squares (LS) (black) and full Kalman filtered (KF) (red) data. The 95% probability ellipses of locations derived from the LS-based model are shown in green. A. Complete tracks showing the increase in track length resulting from the application of the KF algorithm (red). B, C. Detail of the tracks showing the majority of KF locations within the 95% probability ellipses of LS locations.

Since the recent introduction of the Kalman filtering (KF) algorithm for the processing of satellite tracking data by the Argos system, the service providers have made this the default processing method for new transmitters (PTTs), giving the user the option to choose the Least Squares (LS) algorithm in alternative. The data processing of old PTTs that were already being processed with the LS algorithm remains unchanged, unless KF processing is requested, and stored data from 2008 onwards can be reprocessed using either method (albeit with additional processing costs). Processing of data with the new KF algorithm is bound to become more common as old PTTs end their life, and data processed with this algorithm will soon become the standard for Argos-based tracking.

State-space modelling approaches provide the statistical rigor needed in analysing animal movement data, but SSMs are not simple and require considerable care in their use

Our study shows that Kalman filtering consistently provided more estimated locations per animal track than the LS algorithm, supporting previous claims by the Argos service

Like Boyd and Brightsmith

Our results demonstrate that the Jonsen et al.

Although the overall difference in mean errors between the two algorithms appeared small (mean error in LS models was 3.5±3.0 compared to 2.9±2.9 in KF model) the model fit to KF data improved the accuracy of seal trips by 27% over the LS model. The linear mixed-effects model indicated that, despite significant variations in trip accuracy, errors in locations predicted for LS trips were significantly larger than those predicted for KF trips. For both models the largest deviances from true locations occurred along the east/west axis. This is not unexpected since Argos location errors are strongly biased towards the longitudinal component, regardless of the processing algorithm

Tracks reconstructed from the models applied to KF and LS data provided faithful representations of the true seal trajectories measured with Fastloc GPS. However, the LS track tended to deviate more from the true track when seals were making short displacements and frequently changing direction. This is likely due to the correlated random walk model employed in the KF algorithm which would tend to smooth out uncommonly large changes in direction and/or displacement. As a result, LS locations tended to spread over a wider area compared to the KF. This was a common feature to several LS modelled tracks that can have major implications if these data are used to calculate sizes of home ranges or ARS patches.

The SSMs were fit as hierarchical models to the LS and KF data, meaning that data from all seal trips were combined to estimate model parameters, leading to improved location estimates. We anticipate that larger errors would be obtained if models were fitted separately to each trip. Yet, there is no reason to expect that the hierarchical formulation behaved differently when applied to LS and KF data, so we consider that the comparison between algorithms remains valid.

We fitted the same observation equation to data processed with LS and KF methods, thus assuming that the new algorithm did not change substantially the distribution or magnitude of the errors. A recent study demonstrated that both LS and KF location errors are better described by a long-tailed lognormal distribution

Regardless of which processing method is used, our study showed that accuracy of modelled tracks was sensitive to precision of the raw input data. As the proportion of locations with poor precision increased, the ability of the SSMs to recover accurate locations was significantly worse. This is consistent with findings from other researchers that showed that high measurement error not only impacts accuracy and precision of locations estimated from state-space methods

On the other hand, we found no evidence that observation frequency and temporal resolution of Argos data influenced the magnitude of SSM errors, in contrast to a recent study that suggested that frequency and regularity of raw data may be as important as spatial precision for obtaining accurate estimates of locations from state-space methods

Our results strongly suggest that application of SSSM to the whale tracking data processed with the KF algorithm was appropriate and that models fitted well. Estimated parameters from KF models were very similar across all tracks and to parameters from the LS model despite the fact that models were fitted separately to each whale LS/KF-processed dataset.

Paths inferred from both models were also similar, with most of the locations from the reduced KF model falling within the 95% probability ellipses of locations estimated from the LS model, and the majority of locations from the full KF model being close to the whale tracks inferred by the LS model. Similar to what was observed for the seal data, the longitudinal bias in Argos errors caused the reduced KF locations to differ more from their paired LS positions in the east/west than in the north/south axis.

The estimated precision of locations inferred from the SSSM fit to the reduced KF data was higher for 5 out of 6 whale tracks, as indicated by the lower average width of the credible limits. However, the KF model behaved significantly worse than the LS model in the case of the whale track (#80716) for which less than 2 satellite positions were received per day. This cannot be accounted for by variations in Argos location classes because 28 of 29 positions were assigned the same class in both datasets. A close inspection of the raw KF and LS data indicates that the poorer performance of the reduced KF model was likely associated with the highly tortuous whale path evident in the KF data (and not in the LS data) and caused by the way the data regularization approach used in the SSSM’s observation model dealt with this tortuosity. Because the interval between raw satellite positions was considerably longer than the 3-hourly interval at which the SSSM positions were being estimated, raw positions have more weight on model estimates as the model “forces” derived locations to exactly match raw satellite positions. Such an effect tends to be more pronounced with decreasing linearity of the tracks

It should be stressed that the application of the KF algorithm increased the total number of locations in this whale track from 29 to 108 (see

Estimates of behavioural mode from the KF model agreed well with inferences from the LS model – with 94% of whale locations being assigned the same behavioural category in both models – indicating that the KF algorithm did not introduce appreciable changes in the ability of the SSSM to recover latent behaviours from satellite positions.

These results lead us to conclude that application of widely-used Bayesian state-space models

(DOCX)

(DOCX)

(DOCX)

We are grateful to Cláudia Oliveira, Irma Cascão, João Medeiros, Yves Cuenot, Maria João Cruz and many volunteers at DOP/IMAR that helped with whale fieldwork, and to our skilled skippers and crew Paulo Martins, Norberto Serpa and Vitor Rosa. We thank P. Lovell, P. Hammond and S. Moss and others at SMRU for helpful discussions and fieldwork. We acknowledge Greg Breed and one anonymous reviewer for very useful comments on a previous version of the manuscript. This study is an output of research projects TRACE (PTDC/MAR/74071/2006) and MAPCET (M2.1.2/F/012/2011 - Integrating cetaceans into marine spatial management in the Azores).