Understanding seabird behaviour at sea part 2: improved estimates of collision risk model parameters

4. Discussion

At present, there are significant uncertainties surrounding input parameters for collision risk models (Masden et al., 2021), meaning that precautionary values must be used in assessments. The uncertainty, and resultant precaution, can be reduced through the collection of better data to parameterize these models. GPS data offer the potential to do this. Our analyses demonstrate how GPS data can be used to generate estimates of nocturnal activity, flight height and flight speed. However, these parameters vary among behavioural states, and can vary among colonies and through time.

4.1. Nocturnal Activity

At present, the recommended values for nocturnal activity are based on the rankings presented in Garthe & Hüppop (2004) with scores of 1-5 reflecting activity levels of 0, 25, 50, 75 and 100% of day time activity (Band, 2012). Based on the rankings from Garthe & Hüppop(2004), nocturnal activity levels for lesser black-backed gull and kittiwake were 50% of daytime activity levels, and for gannet, it was 25% (Table 24).

Previous analysis of data from gannets suggests that actual levels of nocturnal activity may be lower than that recommended in current guidance (Furness et al., 2018). Of the datasets considered in Furness et al. (2018), the 2013 data from Alderney were also considered in our analysis. Our estimate of nocturnal activity from that dataset (24%) was higher than the value presented in Furness et al. (2018) from the same dataset (15.6%). This difference is likely to have arisen due to the fact that the classification in Furness et al. (2018) was based on flight speed only, whilst in our study classifications were based on behaviour inferred from EMbC and likely to be more reflective of activity estimates than those based on point estimates of speed.

Nocturnal activity levels for all three species were found to vary among years and sites (Table 24; Figures 4,6 and 8). Understanding the reasons for these differences in gannets and kittiwakes is complicated by the fact that data were collected using both IGotU and UvA tags. The data from IGotU tags were collected over 2-3 days, typically around the start of chick-rearing, and consequently may not be representative of the breeding season as a whole. The UvA data are collected over a longer time period, so may be more representative of nocturnal activity over the breeding season as a whole. However, kittiwake data were collected from Flamborough Head and Bempton Cliffs using IGotU tags from 2010 – 2015 and using UvA tags in 2017 and 2018. Estimates of nocturnal activity from the IGotU tags were not inconsistent with those obtained using UvA tags (Figure 6). This suggests that the data collected using the IGotU tags may be representative of the wider chick-rearing period. However, such a comparison is only possible for kittiwakes at Flamborough Head and Bempton Cliffs as this was the only species-site combination for which data from both tag types was available.

For all three species, the proportion of time spent active at night was broadly consistent between sites and years, but there was more variation in the proportion of time spent active during the day (Figures 5, 7 and 9). This may reflect differences in foraging areas between years and colonies. For example, kittiwakes at Rathlin Island showed substantial differences in foraging ranges between 2009 and 2010 (Chivers et al., 2012). If foraging activity is concentrated in daylight hours, then in years when birds have to travel further in order to access sufficient resources to successfully fledge young, the proportion of time birds spend active at night relative to the proportion of time spent active during the day is likely to be reduced. This highlights the importance of considering wider environmental conditions when assessing the proportion of time birds spend active at night. Such an understanding is crucial to determining the potential transferability of estimates of nocturnal activity between sites.

4.2. Flight Speed

Flight speed is used twice in the Band (2012) Collision Risk Model, firstly to estimate the flux rate, and secondly to estimate the probability of collision. Consequently, the model is sensitive to estimates used for flight speed, especially in relation to the flux rate (Masden et al., 2021). At present, recommended estimates for flight speed are based on generic values (e.g. Alerstam et al., 2007; Pennycuick, 1997) collected during a limited range of conditions, and often based on limited sample sizes. Our analysis suggests that typical flight speeds may be lower than those reported in these previous studies, which are often collected in areas which may not be representative of conditions experienced offshore (Alerstam et al., 2007; Pennycuick, 1997). Accounting for these differences can result in a substantial reduction in the predicted collision rate (Masden et al., 2021).

For all three species, commuting speeds were faster than foraging speeds. In the context of the Band (2012) Collision Risk Model, this means that the estimated probability of collision will be higher for foraging birds, but the flux rate will be higher for commuting birds. However, whilst the impacts of flight speed on the probability of collision and flux rate act in opposite directions, these do not cancel one another out (Masden et al., 2021). Consequently, collision rates based on commuting speeds will be greater than those based on foraging speeds. At present, average speeds combining both commuting and foraging flight are used in collision modelling, but it may be possible to account for these differences through modifications to the model.

In previous analyses, estimates of foraging and commuting speeds were derived from the step lengths estimated from Hidden Markov Models (Thaxter et al., 2019). Further analysis of data from lesser black-backed gulls demonstrates that there is strong correlation between the trajectory speeds estimated using EMbC and the instantaneous speeds estimated from GPS for both foraging and commuting flight (Figures 15, 18, 22). As might be expected, correlations were strongest for the fastest sampling rates. However, despite these strong correlations, there was a systematic bias, with the estimated trajectory speeds slower than the instantaneous speeds.

4.3. Flight Height

4.3.1. Overall conclusions

For lesser black-backed gull and kittiwake, we found that commuting flight tended to be higher than foraging flight (Table 22). This is consistent with previous findings for lesser black-backed gull (Corman & Garthe, 2014). For gannet, flight heights were similar for both foraging and commuting flights (Table 20). However, the “stop” behaviour may also reflect finer scale foraging as it is characterized by tracks with high tortuosity and low speed. Flight heights for the “stopped” behaviour were at, or close to the sea surface (Figure 34). The nuances of these different foraging behaviours, and their implications for collision risk, would benefit from further investigation.

For lesser black-backed gulls, we found that estimated flight heights were broadly similar among colonies for foraging flight, but that there were greater differences among colonies in relation to commuting flight height (Tables 16 & 17). Birds from Orfordness had lower median flight heights and spent a lower proportion of their time at collision risk height during commuting flights than was the case for birds from Walney or Skokholm. Similarly, gannets considered in our analyses from Flamborough and Bempton Cliffs had lower estimated heights for both foraging and commuting flight than was the case for birds from the Bass Rock (Cleasby et al., 2015). These data highlight the potential for differences in flight heights among sites, especially in relation to commuting flight.

Flight heights have previously been reported for these species measured using approaches including GPS (Corman & Garthe, 2014; Ross-Smith et al., 2016), altimeter (Cleasby et al., 2015), laser rangefinder (Borkenhagen et al., 2018; Harwood et al., 2018) and LiDAR ( Cook et al., 2018). Analyses of these data have highlighted the potential for spatial and temporal variation in species flight heights. Our analyses highlight how some of this variation may be linked to behaviour in any given place at any given time. Where flight heights are estimated from survey data, it may be possible to use the outputs from our analyses to identify areas where birds are likely to be foraging or commuting. However, this must be done with caution. There are differences between flight heights estimated using different methodologies, although, it is unclear the extent to which these reflect genuine differences in flight heights as a result of spatial or temporal patterns, or differences in the bias, error, and precision with which these measurements are made, and further research is needed to properly understand this.

4.3.2. Limitations

We identified negative bias in GPS altitudes for lesser black-backed gulls tagged at North Sea colonies and addressed this by adjusting the data. However, it is not known whether the adjustments made fully resolved the bias. Most importantly, any uncertainty in the mean bias was not carried forward into the remainder of the analysis, and future studies should aim to explicitly model any potential GPS data biases within the observation error model. The modelled commuting heights for lesser black-backed gulls tagged at Orfordness were lower than those for individuals from colonies in the Irish Sea. By exploring precisely how GPS altitudes were calculated, future work could determine whether there is any remaining altitude bias in the data, and whether any remaining apparent altitude bias has an ecological cause.

4.4. Implications for assessment of collision risk

Of the parameters considered here, flight speed appears to be reasonably consistent among colonies, at least in the case of lesser black-backed gulls (Table 15). However, there appears to be more variation in both nocturnal activity and flight height among colonies. Consequently, if GPS data are to be used to assess parameters used in collision risk modelling, this may need to be done on a site-specific basis, particularly in relation to the estimation of commuting flight heights.

Estimates of collision risk based on GPS data are likely to be greater in areas used for commuting than those used for foraging. In part, this reflects the behaviour of the birds, with a higher proportion of commuting birds at collision risk height than is the case for foraging birds. However, it also reflects the model assumptions and is driven by the influence of flight speed in the estimation of the flux rate (Masden et al., 2021). This highlights a key challenge in the use of GPS data in the Band (2012) Collision Risk Model. Avoidance rates are estimated by comparing observed collision rates to those predicted in the absence of any avoidance behaviour. Consequently, these account for both the behaviour of the birds and any errors in how collision rates are estimated. Flight speed strongly influences the estimation of the flux rate, and therefore predicted collision rate. At present, the predicted collision rate in the absence of avoidance behaviour is derived using generic estimates of flight speed (e.g. Alerstam et al., 2007), which are often substantially faster than those estimated using GPS for both foraging and commuting flight, regardless of whether these have been derived from instantaneous or trajectory speeds. This means that flux rates, and therefore collision rates, estimated using generic flight speed estimates will be higher than those estimated using GPS flight speed estimates for both foraging and commuting flights, and that the over-estimation will be greatest in areas used for foraging. As a result, avoidance rates (based on comparisons of predicted and observed collision rates) derived from generic flight speeds would be higher than those based on the GPS speeds. Given the difference between foraging and commuting flight speeds, and the difference in flight heights between the two behaviours, this is also likely to mean that incorporating behaviour into collision risk models may also require the estimation of different avoidance rates.