Harbour and grey seals: distribution maps

4. Discussion

4.1. Summary of Differences to Previous Work

4.1.1. Overview

4.1.2. Tracking Data

The distribution estimates provided here benefit from an enhanced tracking dataset compared to that used in Carter et al. (2022). In addition to the GPS tracking data from SMRU, UCC and University of Aberdeen used in Carter et al. (2022), the dataset here was augmented with data from grey seals tagged in France by Université de La Rochelle. Furthermore, the deployment by SMRU of tags on harbour and grey seals in Shetland in 2022 has provided a valuable resource to improve distribution estimates for that area. The distribution estimates provided here therefore benefit from a larger sample size of tracking data with better spatial coverage than those presented in Carter et al. (2022). As such, the analysis should give a better approximation of the population-level mean habitat preference relationships, and thus more robust at-sea distributions.

Tracking data were assigned to different habitat preference regions based on the location of the haulout sites used before and after a trip (Fig. 4). Trips starting and ending in different regions were excluded. The region designations followed those of Carter et al. (2022), with two exceptions: (i) Shetland was modelled separately from Orkney and the north coast of mainland Scotland for both species, and (ii) the Western Isles and West Scotland regions were combined for harbour seals in this work because previous work showed little difference in habitat preference between the two (Carter et al., 2022), and combining them allowed for a greater sample size and increased predictive power.

Seal location data were modelled alongside an availability sample; control points randomly spaced within an area deemed accessible to each seal on each foraging trip. In previous work, the accessible area was determined per species based on the maximum swimming distance (i.e., avoiding land) to haulout recorded for any seal in the dataset (Carter et al., 2022). However, this approach does not account for regional differences in scale of movement, which can range from tens to hundreds of kilometres from the haulout (Carter et al., 2022). Such differences may be related to complex regional drivers such as genetics (Carroll et al., 2020), or fear of predation (Moxley et al., 2020) (as is likely relevant for harbour seals in Shetland; see below). With insight from the new Shetland tracking data, it was deemed more appropriate to define the accessible area radius on a species-region basis for this work. Therefore, for each seal trip, control points were placed within an area out to the maximum swimming distance recorded by any seal of that species in that region. The implications of this are that fine-scale distribution patterns are likely to be more accurately represented in regions where seals do not travel far from the haulout site (e.g., Shetland).

The distribution estimates for harbour seals in Shetland reflect a tight coastal distribution compared to those of Carter et al. (2022) (Fig. 7). This provides a better alignment with patterns seen in tracking data collected under this project; seal tracks often traced the coastline of Shetland, and the vast majority of locations (95%) were within 3 km of the nearest land (Figs. 3). This coastal distribution is potentially due to a landscape of fear effect attributable to the presence of killer whales (Orcinus orca), with killer whales observed predating seals year-round in Shetland. This further demonstrates the importance of using region-specific tracking data and habitat preference relationships for distribution estimates, since such behaviour was not as frequently observed in tracking data from harbour seals in Orkney and the north coast of mainland Scotland, where predation pressure by killer whales may be less persistent. The impact of killer whale predation on seals is the focus of ongoing PhD research in the Ecological Consequences of Orca Predation on Seals (ECOPredS) project.

4.1.3. Haulout Count Data

Haulout counts used to scale estimates of at-sea distribution have been updated since Carter et al. (2022) due to the availability of more recent survey data. Given that seal population trajectories can vary regionally through time (SCOS, 2022), updating the count data is important to give a more accurate representation of current seal distribution. A summary of these updates is shown in Table 2 below. A map of survey coverage by year used in the current analysis is shown in Appendices Section 8.3 Figure A4. As in Carter et al. (2022), no August survey data were available for the archipelago of St Kilda, thus predictions do not include seals hauled out there.

4.1.4. Environmental Covariates

In Carter et al. (2022), a range of dynamic (temporally varying) and static environmental variables were used as potential explanatory covariates in the habitat preference models. Dynamic covariates comprised seasonal means of sea surface temperature (SST), water column stratification and frontal intensity. Environmental data were extracted for the years coinciding with the tracking data, and predictions were made for a focal year of 2018, corresponding to the most recent available count data (Carter et al., 2022). One disadvantage of this approach is that predictions are always only relevant to the focal year and may be strongly influenced by variation in these dynamic covariates.

To address this issue, the current approach fits the models with only static covariates, including static representations of dynamic processes. Thus, covariates included distance to haulout, distance to coast, seabed substrate type, seabed geomorphology and, for grey seals, summer mean potential energy anomaly (PEA; a metric of water column stratification). The PEA data were static in that they represent stratification conditions in a “typical” year (Jones, 2024). For more information on these covariates, see Appendices Section 8.2.2. This approach eliminates the possibility that predictions will be made outside of the covariate space in which the models were fitted. Predictions can also be more easily updated in the future when new count data become available. However, it is important to note that the resulting predictions therefore represent seal distribution in a “typical” year. For some applications where the influence of dynamic processes is of particular interest (e.g., understanding the influence of temporal variation in a particular dynamic oceanographic feature on seal distribution), it will be necessary to use dynamic covariates and generate multiple predictions corresponding to different conditions.

4.1.5. Statistical Modelling

The key aims of the work undertaken by Carter et al. (2022) were to both predict seal distribution, and to understand the key environmental drivers of distribution for grey and harbour seals in different regions. Model selection was undertaken, and non-informative covariates were removed from the model until arriving at a minimal adequate model. In this current work, no model selection was undertaken; non-informative covariates which likely have little to no influence on predicted distributions remained in the model.

Analysis of time-series such as animal tracking data is often affected by the problem of residual serial autocorrelation (Fieberg et al., 2010). If ignored, this can lead to underestimation of model uncertainty (i.e., artificially narrow confidence intervals around the mean) (Fieberg et al., 2010). Having explored the available options for handling this problem, the previous analysis used a time-to-independence approach; effectively thinning the data by removing every n^th observation until residual autocorrelation reached acceptable levels (Carter et al. 2022). Whilst this is a legitimate option, it is not without disadvantages. The key disadvantage is that valid data are discarded, often leading to unnecessarily wide confidence intervals, and ultimately dilution of ecological relationships. In the present analysis, a different approach was applied. A first-order autoregressive correlation structure (AR1) was applied to the models, and calibrated such that correlation between time series observations was modelled, rather than removed. The consequences of this difference in data treatment are that modelled uncertainty is reduced in the current approach, and therefore confidence intervals around the mean prediction are likely to be narrower than those of Carter et al. (2022).

4.2. Considerations and Recommendations

4.2.1. Absolute versus Relative Density

Here we provide distribution estimates as both relative density (percentage of at-sea population per grid cell) and absolute density (number of animals per grid cell). For some applications, absolute density is favourable (e.g., estimating the number of animals within an area of interest). However, the conversion process from relative to absolute density involves use of population scalars derived from telemetry data (see Carter et al. (2022)), and uncertainty in these scalars is not propagated through to the confidence intervals around the mean. Confidence intervals therefore only reflect uncertainty in the modelled habitat preference relationships. Another consideration is that density estimates are scaled using the most recent available count data. While relative density estimates are somewhat robust to changes in abundance (provided the distribution of the population remains the same proportionally among haulouts), the absolute density estimates are not. As such, absolute density estimates provided here reflect an approximation of seal distribution in 2023. Here we show maps of absolute density for ease of interpretation but provide GIS layers for both absolute and relative density. Given the caveats listed above, we recommend that relative density estimates be used wherever possible.

4.2.2. Use of Confidence Intervals

Confidence intervals represent the range of values within which, based on the haulout count data and model used, we would expect the true density of seals to be, and the mean is a measure of the centre of this range. Where possible, mean density estimates should be used in conjunction with the confidence intervals. As in Carter et al. (2022), confidence intervals around the mean prediction are generated on a cell-by-cell basis. Thus, although the mean predictions can be summed across an area (e.g., number of animals present within a wind farm development zone), confidence intervals cannot; doing so would lead to inflated uncertainty. Currently, area-based confidence intervals can be generated on a case-by-case basis, but this requires significant extra work. A priority for future work is to produce a graphical user interface (GUI) where users can specify their area of interest and download area-based mean estimates with associated area-based confidence intervals.

4.2.3. Data Limitations

While the distribution estimates presented here benefit from an improved GPS tracking dataset over those of Carter et al. (2022), there remains a key data gap in Scottish waters. Very little recent high resolution tracking data exist for the east coast of Scotland for either species. Data used to fit models for harbour seals in this region were from tags deployed in 2008 (n = 4), 2011 (n= 5) and 2013 (n= 3). Data from grey seals for this region are from 2005 (n = 2), 2008 (n = 9) and 2013 (n = 2), supplemented by individuals tagged between 2014 - 2018 in Orkney (n = 1), the Moray Firth (n = 3) and Southeast England (n = 5) that hauled out there. Predictions for this region should be treated with caution as they likely contain a high degree of unmodelled uncertainty (i.e., uncertainty not incorporated in the confidence intervals). Since the deployment of tags in this region, harbour seal numbers have continued to decline (Russell et al., 2022; Thompson et al., 2019). The vast majority of data for grey seals in east Scotland are from individuals tagged in the Eden Estuary and Firth of Tay, yet large aggregations are now present ~100 km north in Cruden Bay and the Ythan Estuary. Thus, predictions of distribution for seals hauling out in Cruden Bay and the Ythan Estuary are predominantly based on the habitat preference of seals further south, tagged over 15 years ago. Given the current depleted abundance of the harbour seal in this region, a large-scale tag deployment on harbour seals may not be feasible. However, deployment of GPS tags on grey seals should be considered a priority for future work.

4.3. General Conclusion

In conclusion, the seal distribution maps presented here represent an improvement on those of Carter et al. (2022) due to the contribution of new tracking data from tags deployed on harbour and grey seals in Shetland, updated abundance data, and a number of methodological improvements outlined above. The current estimates should therefore be used in favour of those from Carter et al. (2022) for any applications where the distribution of seals from haulouts in Scotland is required.