Scottish Marine and Freshwater Science Volume 5 Number 16:The Avoidance Rates of Collision Between Birds and Offshore Turbines

2. Objectives

2.1 Produce definitions for the types and scales of avoidance rates that will be used throughout the review document

It is important to make a distinction between avoidance rates, as used in collision risk models, and avoidance behaviour. Avoidance behaviour refers specifically to the behavioural response of birds to wind turbines. However, at present, in addition to accounting for avoidance behaviour, avoidance rates are often used as a 'fudge-factor' to account for error in the model itself and in its input parameters (see May et al. 2010, Douglas et al. 2012). Whilst SOSS guidance (Band 2012) sets out how these uncertainties should be accounted for in the collision risk modelling process, in practice, this is rarely done. The purpose of this review is to identify suitable avoidance rates for use in collision risk models; these rates will be informed, where appropriate, by recorded estimates of avoidance behaviour.

A lack of clear, working definitions for different avoidance rates has hampered attempts to come up with standardised measures. Present definitions of avoidance rates rely on an ability to collect empirical data with which to compare predicted and observed collision rates ( SNH 2010). As this is impractical for the offshore environment, Band (2012) proposes combining estimates of micro- (or near-field) avoidance, where a bird takes action to avoid collision at a point close to the turbine, and macro- (or far-field) avoidance, where a bird takes action to avoid collision at a point distant from the turbine, to generate an estimate of total avoidance. However, the empirical data underpinning such definitions is currently inconsistent and difficult to interpret.

A key problem is often the lack of detail over what spatial scale data have been collected at. For example, radar monitoring has shown that birds may take action to avoid entering a windfarm at distances of up to 6 km (Christensen et al. 2004), far further than could be observed by eye. As a result, by relying on visual observations, avoidance rates may be under-estimated as a significant proportion of birds will have taken action to avoid the windfarm before they are visible. Similarly, at present, it is not possible to identify birds to species level on the basis of radar echoes; consequently, by relying on radar, it will not be possible to derive species-specific avoidance rates. This is further complicated by evidence that avoidance can occur in a three-dimensional space, with horizontal avoidance, where a bird alters its heading to avoid collision, and vertical avoidance, where a bird alters its altitude to avoid collision (Krijgsveld et al. 2011, Plonczkier & Simms 2012). Such alterations may be relatively subtle and difficult to detect by eye. Where radar is utilised to monitor movements in response to turbines, it requires the use of both horizontal and vertical radar. Evidence describing three-dimensional avoidance behaviour, if it exists, is likely to be extremely limited. In defining different avoidance behaviours, the review therefore gives careful consideration to the methodologies used to collect the necessary data.

Wind turbines are most typically in the order of seven rotor diameters apart (Meyers & Meneveau 2012), based on existing turbine designs, this may vary from 480 m to 1.5 km, depending on the capacity used. Given the variable distances between turbines and the difficulties in obtaining consistent estimates of avoidance behaviour over the relevant spatial scales, the review considers whether it is possible to define micro-and macro-avoidance with reference to distance to turbines, or whether a more pragmatic approach, basing definitions on whether a bird is inside or outside a windfarm would be more appropriate. The review considers whether these definitions are appropriate to all species and groups, or whether a more flexible approach is necessary. This may depend on what evidence is available for different species. For example, avoidance rates for terns have often been derived from observed collision rates (Everaert 2008), whilst for other species, such as northern gannets, avoidance rates may be more reliant on radar data (Krijgsveld et al. 2011). The review then considers evidence for avoidance behaviour occurring over horizontal and vertical planes.

The review provides clear and concise definitions for micro-horizontal avoidance, micro-vertical avoidance, macro-horizontal avoidance and macro-vertical avoidance. Definitions are produced based on the behaviour of the birds as opposed to the requirements of a model and offer guidance about how final values can be adapted for use in different models.

Defining the different forms of avoidance behaviour represents a major step forward in collision risk modelling. These definitions are central to the rest of the project, and, as such, have been agreed through discussion with the project steering group of key stakeholders and experts.

2.2 Review the current use of avoidance rates

In order to provide context to this work, it is important to consider how avoidance rates are currently used. With this in mind, the review considers published EIAs and identifies what avoidance rates have been used within the collision risk modelling process and what justifications have been put forward for their selection. This will help us determine how consistently existing guidance has been interpreted and applied, and help refine future guidance in order to minimise discrepancies in its application.

2.3 Review and critique existing avoidance behaviour studies and any derived rates

Avoidance rates have been derived from both observed mortality rates and actual observations of birds' behaviour (Cook et al. 2012, Trinder 2012, Moray Offshore Renewables Limited 2012, Smartwind/Forewind 2013, Everaert 2014). In Belgium, at Zeebrugge port breakwater, onshore collision rates in terns and gulls have been used to derive avoidance rates based on recorded movement patterns and assumptions about turbine design (Everaert & Stienen 2007 Moray Offshore Renewables Ltd. 2012, Everaert 2014). However, the difficulties in directly recording collisions in the marine environment mean that studies of avoidance at offshore windfarms have relied on observing behaviour (Desholm et al. 2006, Blew et al. 2008, Krijgsveld et al. 2011). These studies have varied both in the species they have investigated, and also in the potential form of avoidance behaviour reported.

Recognising that appropriate data may be extremely limited, we initially take a broad approach to our review, reviewing evidence for avoidance behaviour in marine birds generally. We demonstrate how this evidence relates to the definitions set out in the previous section of the report. Having done this, we assess whether sufficient evidence exists to draw conclusions about avoidance behaviour in five priority species - northern gannet, black-legged kittiwake, lesser black-backed gull, herring gull and great black-backed gull. If this is not possible, we will consider how to combine evidence within groups of species, on the basis of the ecology of the species concerned. Where this is necessary, we clearly state which species are in each group.

In order to make an assessment of the level of confidence in the reported avoidance rates for each species or species group, we make a detailed qualitative critique of each study. Key questions include:

i. How have avoidance rates been derived?

We consider first whether the avoidance rates reported have been determined from observed mortality rates or actual observations of birds' behaviour. The data collection methods used are summarised, and the limitations of each method discussed. Where avoidance rates have been back-calculated from observed collisions at reference windfarms, they may incorporate error associated with model input parameters including population estimates, flight heights and turbine operational characteristics in addition to the actual avoidance behaviour of the birds. In contrast, direct observations of birds' behaviour in relation to turbines will not incorporate model error. However, these observations may still need careful interpretation given methodological constraints over how data may be collected, for example, the distances over which birds can be observed in comparison to the distances over which they may take avoidance action.

ii. How comparable are the different datasets?

Avoidance rates based on behaviour have typically been derived from a series of visual or radar observations (Desholm & Kahlert 2005, Blew et al. 2008), or through a combination of both (Krijgsveld et al. 2011, Plonczkier & Simms 2012). The range of distances over which data can be collected varies markedly between these platforms (Cook et al. 2012) and it is important to consider whether estimates - particularly of macro-avoidance - are comparable between different studies.

It is also important to consider how and when data have been collected. For example, visual observations from land, or an offshore platform, may differ from those obtained during a boat-based survey, where the movement of the boat may mean that surveyors have a less stable platform or because birds may exhibit a behavioural response to the presence of a boat (although following standard guidance should help to minimise the influence of these factors: Camphuysen et al. 2004). Visibility may also strongly influence results from visual observations. Seasonality may influence the results from both radar and visual observations as foraging birds may respond very differently to migrating birds (Blew et al. 2008, Krijgsveld et al. 2011). This may be particularly important for radar studies, where it is not possible to identify radar echoes to species level and, as a result, it is more difficult to separate observations of migrants from those of local, foraging birds during periods of passage.

iii. Are reported avoidance rates affected by any special factors?

The location of the windfarm may have a strong impact on reported collision rates. If these collision rates are then used to calculate avoidance rates, it may lead to an erroneous assessment of avoidance behaviour. For example, a Belgian study has reported collision rates at a windfarm in Zeebrugge for terns (Everaert & Stienen 2007). The results from this study have been widely used to calculate micro-avoidance rates for terns ( e.g. Whitfield 2008). However, as this windfarm was located on a seawall, next to a breeding tern colony, it is unclear whether behaviour around the turbines would be consistent with that of foraging terns, further out to sea. In addition, the size of turbines planned for offshore windfarms is significantly greater than those installed at many of the sites for which collision data are available. For this reason, we will consider whether there is any evidence for a relationship between turbine size and the avoidance rates derived from mortality data.

2.4 Provide summary avoidance rates and a total avoidance rate for each priority species/species group based on the evidence available at present

Based on the information compiled from the above review, we derive avoidance rates based on published evidence for each of the five priority species - northern gannet, black-legged kittiwake, lesser black-backed gull, herring gull and great black-backed gull, and other species as relevant. Where necessary, this involved going back to the source material of the studies concerned and back-calculating avoidance rates following the methodology set out by Band (2000). Where insufficient data were available to make recommendations for individual species, we combine estimates within species groups, based on the ecologies of the species concerned. Based on our critique of the studies from our review we then indicate where our confidence in each reported value is affected by the quality of the data it is based on.

Where possible, we combine avoidance rates collected at different scales, in order to calculate a total avoidance rate for each species. Estimates of micro-avoidance and macro-response can be combined to give an overall avoidance rate following equation 1, if sufficient data are available, we will extend this equation to include horizontal and vertical avoidance, as detailed in equations 2 and 3. Given the limited evidence available, it may be necessary to draw in data from closely related species and derive avoidance rates based on a group, rather than species-specific basis. Where this is necessary, we will clearly state what we have done and indicate our confidence in the derived rate accordingly.

A _rate = 1 - [(1 - A _micro) X (1 - A _macro)] [Eq. 1]

A _micro- = 1 - [(1 - Mi _horiz) X (1 - Mi _vert)] [Eq. 2]

A _macro- = 1 - [(1 - Ma _horiz) X (1 - Ma _vert)] [Eq. 3]

Where A _rate is the total avoidance rate, A _micro-is the micro-avoidance rate, A _macro-is the macro-avoidance rate, Mi _horiz is the micro-horizontal avoidance rate, Mi _vert is the micro-vertical avoidance rate, Ma _horiz is the macro-horizontal avoidance rate and Ma _vert is the macro-vertical avoidance rate. Note that the ability to combine horizontal and vertical movements in this way will depend on how data have been collected. It is likely that some birds will make horizontal and vertical movements concurrently, and therefore, it would not be appropriate to combine data in this way.

This summary is used as the basis for a gap analysis based on our earlier definitions of avoidance behaviour. In combination with the above critique of avoidance rate studies, this gap analysis will help provide a target and possible methodologies for future research on avoidance behaviour of birds in relation to offshore windfarms, for example the Offshore Renewables Joint Industry Project ( ORJIP), due to get underway in summer 2014 (Davies et al. 2013).

2.5 Undertake an assessment of the sensitivity of the conclusions reached to inputs and conditions under which they were collected

The final avoidance rates are likely to be sensitive to both factors which are directly parameterised within the collision risk model, such as species' flight heights, turbines' operational time and rotation speed, those parameterised in collecting collision data such as corpse collection, and also those which are not directly parameterised, such as seasonality, weather conditions and whether data have been collected during the day or night. Whether estimates of avoidance behaviour have been derived from behavioural observations or recorded collision rates, they are likely to be influenced by the factors which are not directly parameterised. For this reason, we assess how such variables are likely to have influenced the final avoidance rate in each study. For example, avoidance rates based on data only collected during conditions with better than average visibility may be expected to differ from those based on data collected during periods of poor visibility, a potential source of model error. Where avoidance rates have been derived from collision data, there is the also potential for the model input parameters to influence the final values.

These methodologies have typically been restricted to turbines at onshore locations (Everaert & Stienen 2007), where corpse collection is practical. There are concerns that this may lead to an over-estimate of the avoidance rate as some corpses go undetected and correction factors to account for this (Winkelmann 1992, Bernardino et al. 2013) may not be correctly applied. With this in mind, we focus on the best quality studies, but also consider how undetected corpses may influence the avoidance rate we derive.

Where a collision rate is available for a site, the avoidance rate (A _rate) can be calculated as follows:

C _pred = (Flux rate * P _coll) + error [eq. 4]

A _rate = 1 - (C _obs/C _pred) [eq. 5]

Where C _pred is the predicted number of collisions in the absence of avoidance action, C _obs is the observed number of collisions, flux rate is the total number of birds passing through the rotor swept area and P _coll is the probability of a bird colliding with a turbine. The probability of collision, P _coll can be calculated following the formula set out in Band (2012). However, this highlights a second area where the conclusions about avoidance rates may be sensitive to the inputs as values of P _coll will be specific to the design of turbines (Cook et al. 2011). Consequently, knowledge of rotor speed, radius, chord width and pitch, for the turbine concerned, are required before estimating an avoidance rate from a collision rate. These characteristics can vary considerably, even between turbines of a similar generating capacity (http:// www.4coffshore.com). As a result, error is likely to be introduced into the calculation through inaccuracies in estimates of the flux rate and also through inaccuracies in the estimation of P _coll.

As detailed in Cook et al. (2012), failing to account for turbine design correctly when deriving avoidance rates as described above can lead to erroneous estimates of P _coll and, therefore, the avoidance rate. For this reason, where a study reports a collision rate, rather than an avoidance rate, we have attempted to obtain data on these parameters. Where we are unable to obtain this information, we calculate a value of P _coll based on the parameters from a range of turbines of a similar size. We then consider whether avoidance rates derived from collision estimates are more sensitive to variation in turbine design or to correction factors that account for failure to detect corpses.

2.6 Applicability of avoidance rates to different collision risk models

We finally consider how the total avoidance rate, and its constituent elements, reflect the values necessary for collision risk modelling. At present, the collision risk model formulated by Band (2012) for use in the offshore environment has three different options which can be used to estimate the total number of birds at risk of collision. These options reflect different ways in which estimates of the proportion of birds at collision risk height can be incorporated into the collision risk modelling process. Band model option 1 assumes that birds are distributed evenly within the rotor-swept area of a turbine. It bases estimates of the proportion of birds at risk of collision on data collected during pre-construction surveys of the site in question. Band model option 2 is mathematically identical to the first option, also assuming an even distribution of birds within the rotor-swept area of the turbine. However, the proportion of birds at collision risk height is estimated from continuous distributions derived from data collected across multiple sites (Cook et al. 2012, Johnston et al. 2014a,b). Options 1 and 2 of the Band model are collectively referred to as the basic model. In practice, birds are unlikely to be evenly distributed across the rotor-swept area of a turbine (Johnston et al. 2014a). Band model option 3, often referred to as the extended Band model, accounts for this by using a continuous flight height distribution to estimate collision risk at different points within the turbines rotor-swept area.

As birds are typically clustered to the lower edges of the rotor-swept area (Johnston et al. 2014a), option 3 often results in lower estimates of collision rates. As a consequence, there is intense interest in its use within EIAs for offshore windfarms. However, avoidance rates currently in use that are derived for the onshore environment by combining collision rates with estimates of P _coll from the basic Band model are not suitable for use in the extended model, as accounting for a heterogeneous flight height distribution will result in a lower number of collisions predicted in the absence of avoidance. (Although, note that this difference may be partially offset as avoidance rates derived in this way do not account for changes in flight altitude in response to the presence of a windfarm.) As a result estimates of avoidance behaviour based on the basic model are likely to be higher than is appropriate for the extended model (equations 4 and 5) - this is considered as part of the review.

Where estimates of avoidance rates have been derived from behavioural observations, for example displacement from offshore windfarms, rather than recorded collision rates, the applicability to different models is less clear. We consider how our final avoidance rates have been derived and what implications this has for how they are incorporated in collision risk models.

We also offer guidance not just on the applicability of avoidance rates to the basic and extended Band models, but also their transferability of avoidance rates to alternatives including the Biosis model (Smales et al. 2013).

The data necessary to derive avoidance rates suitable for use with option 3 of the Band model following the formula given by equation 6 are often unavailable. However, a suitable avoidance rate can be derived by estimating the ratio of P _coll from option 2 of the Band model to P _coll from option 3 of the Band model and applying this to the inverse of the avoidance rate used for option 1. For the rationale and a full description of this approach see the supplement to the guidance on 'Using a collision risk model to assess bird collision risks for offshore windfarms' (Band 2012) provided by Bill Band as Annex 1 to this report.