The diver s guide for the German Bight distribution of divers. (Gavia sp.) in the light of progressing offshore wind farm

Transcription

1 The diver s guide for the German Bight distribution of divers (Gavia sp.) in the light of progressing offshore wind farm construction 5 6 For submission to J Environmental Management Bettina Mendel 1, Philipp Schwemmer *1, Verena Peschko 1, Sabine Müller 1, Henriette Schwemmer 1 Moritz Mercker², Stefan Garthe Research and Technology Centre (FTZ), University of Kiel, Hafentörn 1, Büsum, Germany ²BIONUM - Büro für Biostatistik, Finkenwerder Norderdeich 15 A, Hamburg, Germany * corresponding author. Tel.: addresses: mendel@ftz-west.uni-kiel.de (B. Mendel), schwemmer@ftz- west.uni-kiel.de (P. Schwemmer * ), info@bionum.de (M. Mercker), garthe@ftz- west.uni-kiel.de (S. Garthe) 23

2 Abstract. Seabirds select for suitable habitats at sea, however, such habitats are currently strongly impacted by marine spatial planning in many sea areas. One major issue is the construction of Offshore Windfarms (OWFs) and associated ship traffic. Divers (Gavia spec.) are particularly vulnerable with respect to anthropogenic activities while at the same time they have a high conservation status and are thus most relevant for planning processes. We investigated effects of OWF construction and ship traffic on diver distribution in the German North Sea on a large-spatial scale using a BACI ( before after control impact analysis) approach and a longterm data set. Our study shows that many OWFs have been built in or close to core areas of diver distribution. Divers showed significant shifts in their distribution in the after period and aggregate now between two OWF clusters, indicating a remaining suitable habitat. The distance at which the decrease of divers to the closest OWF becomes significant is around 16 km. The impact of ship traffic on divers is significantly negative, thus OWFs scare divers by the joint effect of ships and wind turbines. Our study is the first analyses of extensive effects of OWPs and ships on a large spatial scale. The results presented are an essential baseline for future marine spatial planning processes in the German North Sea and elsewhere Key Words: offshore wind farm, ship traffic, environmental impact, Marine Special Protection Area, Gavia spec. habitat loss, habitat model, management, BACI 46 1

3 INTRODUCTION Shallow shelf sea areas have been used by humans since a very long time. The North Sea belongs to the sea areas with the most intensive human activities worldwide that include fishing, transports, oil and gas drilling as well as gravel extractions (Halpern et al. 2008; Emeis et al. 2015). A relatively new human use that demands a considerable degree of attention in the marine planning process is the installation of offshore wind farms (OWFs) which are currently being built in many sea areas throughout Europe and elsewhere. In order to meet their climate goals, many European governments have started to install and plan further OWFs within comparably large sea areas (e.g. Breton & Moe 2009; Langston 2010), Germany being one of the countries with the most extensive plans for OWF installation (Beiersdorf & Radecke 2014). Germany is intending to extend the offshore power to 6,500 MW by 2020 and to 15,000 MW by 2030 which will lead to a strong increase in number of OWF sites mainly in the German North Sea. Currently (2018), 17 OWFs are in operation, five further ones are under construction and several more are being approved in German sea areas (BSH 2017). Within the process of marine spatial planning, these permanent installations at sea are profoundly adding to other types of marine human activities and compete with sea areas assigned for nature conservation (Moksness et al. 2009; Nolte 2010; Emeis et al. 2015;) and may overlap with areas used by resting and foraging seabirds. Previous studies point out contrasting effects (negative or positive) of OWFs on seabirds that strongly vary between areas and species, respectively (Garthe & Hüppop 2004; Fox & Petersen 2006; Drewitt & Langston 2006; Masden et al. 2009; Furness et al. 2013; Dierschke et al. 2016). The construction and 2

4 maintenance of OWFs is furthermore associated with a strong increase in shipping activities in and around OWFs (Exo et al. 2003). OWFs may cause direct effects, such as collision of individuals with wind turbines that may impact directly the whole population (Goodale & Milman 2014, Fox et al. 2006, Masden et al. 2009). Furthermore, indirect effects such as habitat loss and reduced food availability (Drewitt & Langston 2006, Fox et al. 2006, Stienen et al. 2007) may impact the energy budget and condition of individual birds. Long-term effects of such indirect effects on population level are hard to estimate (Fox et al. 2006, Goodale & Milman 2014). However, birds have been shown to lose suitable resting and foraging habitats or to select for less suitable sea areas (Stienen et al. 2007). Furthermore, they may need to enhance their flight time by circling around OWFs on their way to suitable foraging sites (Drewitt & Langston 2006, Masden et al. 2009). This study aims at quantifying the levels of indirect effects (i.e. habitat loss by OWFs and associated ship traffic) on divers (Gavia sp.) to build a base line for future studies that might address population consequences. Divers belong to the most sensitive species group with respect to avoidance of OWFs, which has already been shown for single OWF sites in the North Sea (e.g. Petersen et al. 2006; Leopold et al. 2010, Dierschke et al. 2012; 2016, Mendel et al. 2014, Welcker & Nehls 2016). Furthermore, red-throated divers are also very sensitive with respect to ship traffic: It has been shown that this species shows high flush distances in front of approaching vessels (Bellebaum et al. 2006) and exhibits significantly lower densities in areas with permanently higher ship traffic (Hüppop et al. 1994; Schwemmer et al. 2011). Due to their sensitive nature and due to a significant proportion of the biogeographic population resting in European waters, divers are listed on the Annex I of the EU Birds Directive and are rated particularly 3

5 threatened with respect to human activities (e.g. Garthe & Hüppop 2004; Furness et al. 2013). As negative effects on an individual and population level of divers by avoidance of OWFs can currently not be ruled out (Dierschke et al. 2016; 2017), divers have currently been rated as a species group that needs to be particularly considered with respect to marine spatial planning in Germany and UK (Busch et al. 2013). In the North Sea the group of divers mainly consist of red-throated divers (G. stellata; 90 %) and to a minor proportion of black-throated divers (G. arctica; 10 %) (Garthe et al. 2007; Dierschke et al. 2012). One of the most important resting sites for divers is located in German North Sea waters, where internationally important numbers are reached mainly during spring migration (Skov et al. 1995; Garthe et al. 2007; Mendel et al. 2008; Garthe et al. 2015). Around 20,200 divers use the German waters during this period (Garthe et al. 2015). Due to the high importance of this resting site and due to the high sensitivity of divers with respect to human disturbances, the Special Protection Area (SPA) Eastern German Bight has been established (Fig. 1). However, within this SPA there is a potentially high conflict with an OWF ( Butendiek ) that had been approved before and installed after the establishment of the SPA (Garthe et al. 2012). Further OWFs ( Helgoland Cluster ) are located just south of the border of the SPA (Fig. 1)

6 Fig. 1: Location of the study site within central Europe (inserted map in Fig. 1a) and in the eastern German Bight (North Sea) with location of the different OWFs 5

7 and the area surveyed for diver abundance (yellow to red squares) across the Special Protection Area (SPA) Eastern German Bight for the before (a) and after period (b). Start of construction: Wind farm Nordsee Ost during summer 2012; end of all construction works: Wind farm Butendiek during summer So far, there is only limited information on long-term and large-scale effects of OWFs on divers and a long-term comparison of their distribution before and after the installation of OWFs is missing. Furthermore, the effect of increasing construction and maintenance ship traffic has rarely been considered (Christensen et al. 2003, Boon et al. 2010). Thus, we firstly hypothesized that divers will avoid the OWF areas and show different distribution patterns as compared to the time before the installation of OWFs. Secondly, we hypothesized that ship traffic associated with the OWF sites will lead to avoidance reactions of divers. Against this background, this study was set up to shed light on the five subsequent topics: (1) We are able to make use of a long-term data set, covering the situation before the installation of OWFs (hereafter named before ) for 14 years. This study, thus, aims at directly comparing the distribution of divers after the installation of OWFs (hereafter named after ) in contrast to before this time on a long-term perspective. This could not be achieved in most of the studies published so far. (2) Virtually all of the previous studies investigating potential effects of OWFs on divers focussed on effects of single OWF sites and their direct vicinity (see Dierschke et al. 2016). This only enabled to study the reactions of divers on a rather small spatial scale only showing that diver numbers were impacted within the respective site but not able to show where they moved (Rexstad & Buckland 2012). In contrast, 6

8 this study aims at analysing the large scale effects of multiple OWFs on diver distribution considering also potential shifts from the before to the after period. (3) There is currently still a big need to disentangle the potential effects of OWFs from the effects of natural habitat characteristics that determine the distribution of divers (Garthe 1997; Winiarski et al. 2014). We set up a model including stable natural parameters such as water depth and distance to land as well as anthropogenic predictors like distance to closest OWF and shipping traffic. (4) As the installation and maintenance of OWFs is associated with strong increases in ship traffic, the effects of shipping needs to be quantified and disentangled from the effects of the OWFs themselves. So far this only has been analysed based on general ship densities (e.g. Leopold et al. 2014; APEM 2013, 2016). But OWF ships are a dynamic source of disturbance for divers. Hence, this study aims at relating diver and ship-distribution at a very high spatial and temporal scale, i.e. relating ship distribution from the Automatic Identification System (AIS) with diver abundance assessed during aerial surveys. (5) Given a negative effect of OWFs on divers, it will be necessary to quantify the avoidance distance to wind farms in order to draw conclusions on the degree of resulting (permanent) habitat loss. Thus, in this study we chose two different approaches for analysing different aspects of the OWF effects on divers: First, we used before -data to include the importance of the OWF areas before construction. Second, we focused on the simultaneous effect of OWF and ships associated with OWF after construction of the OWF. The combined interpretation of both approaches allows comprehensively evaluating the OWF effects on divers

9 METHODS Study area The study was conducted within the eastern part of the Exclusive Economic Zone (EEZ) of the German North Sea, south of N, north of 54.18, east of 6.51 E and west of 8.16 E (Fig. 1a). The study site was located within an area between 8 km and 100 km off the Wadden Sea islands of northern Germany. Water depths ranged from 10 m to 40 m. Diver distribution was recorded within the SPA Eastern German Bight and beyond and therefore the study site covered the core area of highest diver densities within German waters (Garthe et al. 2015). The OWF Butendiek is located in the core area of the SPA, while OWFs of the Helgoland Cluster are located at the border of the SPA and south of the core diver distribution (Fig. 1a) Recording of diver distribution and processing of data Distribution data for the period before the installation of OWFs are taken from the data base as used in Garthe et al. (2015). These data cover the months of spring migration (i.e. March and April) of the years , the period of the highest concentration in German North Sea waters. The records originated from environmental impact assessment studies (EIA) issued by the German Federal Maritime and Hydrographic Agency as well as from seabird monitoring and research programmes (for details see Garthe et al. 2015; Fig. 1a). The overall data-set was combined from visual aerial and ship-based surveys. Briefly, divers were counted along transects of a known area which enabled to compute densities (see Diederichs et al. 2002; Garthe et al for a full description of both recording methods). 8

10 Distribution data for the period after the installation of OWFs originate from the compulsory environmental impact studies issued by the German Federal Maritime and Hydrographic Agency during operation of the turbines as well as from the research project Helbird, funded by the German Federal Ministry for Economic Affairs and Energy. Data from the period after the OWF installations consists of 10 digital aerial surveys in (five surveys conducted during research and EIA, respectively). Similar as to the before scenario those flights covered the main spring migration of divers, which included the last week of February and the first week of May to allow for a higher sample size of survey days (Fig. 1b). Those data were obtained by video-based digital recordings instead of visual observations. Briefly, an aircraft sampled a transect of a known area by a video camera and all seabirds found were recorded to compute overall densities (for a detailed description of the method see Thaxter & Burton 2009; Buckland et al. 2012). A change from visual to digital survey methods was mandatory for safety reasons as during the construction and operational phase of the turbines the flight altitude needs to be higher (550 ft instead of 300 ft in visual observations) which does not allow for visual recordings. Data of the main construction phase were not considered in our study. As disturbance during the construction of the OWF is temporary and mainly associated to construction ships, its contribution to the overall OWF effect on the population is assumed to be of low importance regarding the expected OWF lifetime (Christensen et al. 2003). We therefore only focused on the operational phase during which OWFs can affect individuals as a visual barrier or supply vessels working in the area can disturb seabirds in the vicinity (Exo et al. 2003, Drewitt & Langston 2006). Visual observations of seabird distribution are known to underestimate birds detected in more distant spots of the transect band from the observer (Buckland et 9

11 al. 2001; 2015). Therefore, a species-specific correction factor was applied for aerial and ship-based observations, respectively (see Garthe et al for details). No distance correction was necessary for the video-based digital surveys, as the probability to detect a bird is equal across the whole transect. All three recording methods rely on the same principle, i.e. transect sampling of birds to eventually compute densities. However, we did not compare absolute density values between the before and after period as it is not yet sure whether the visual and digital methods lead to the same absolute values (Buckland et al. 2012, Skov et al. 2016). This can only be tested by performing both methods at the same time. Such a data set is currently not available. Thus, for comparing both periods the relative deviance from the maximum density in each period in % was computed and visualized to compare the distribution and location of high density areas of divers between the two periods. For each of these three above mentioned methods separately (visual aerial and shipbased surveys, video-based digital recordings), data have been spatially pooled in a grid with cells of 2.5 x 2.5 km for the before and after period respectively. For this, bird numbers and monitored area have each been summed up per grid cell, which eventually can be used to compute mean densities for each period, whereas geographical coordinates have been averaged for each cell Integrating covariates for the BACI-control-impact approach We related the averaged distribution data of divers with environmental variables using ArcGIS (version 10.3; Environmental System Research Institute 2016). These included (1) dist_coast = minimal distance to the mainland and larger islands (without Helgoland), (2) dist_helgoland = minimal distance to the island of Helgoland, (3) 10

12 dist_owf = minimal distance to the border of the wind farms, and (4) mean_depth = mean water depth. This first model, hereafter named BACI (BACI control-impact approach) did not consider the effect of ships as ship data in a sufficiently high spatio-temporal resolution were only available for the after period. Therefore, to distinguish between the effect of the OWFs and the effect of ship traffic on divers, we have set up a second model using only the data from the after period (hereafter named shipmodel; see below). To merge the environmental variables in an optimal way with the bird count data, we firstly also pooled the covariates to a spatial grid of 2.5 x 2.5 km, and secondly fitted each covariate with a GAM using the function gam() from the R-package mgcv (Wood 2006; R Core Team 2017; R version 3.4.2). Here, we used only latitude and longitude as a smooth 2D-predictor based on cubic splines with the maximal degree of freedom, so that the result represents a cubing interpolation on the given (possibly irregular) grid. Thirdly, we used the predict() function to predict the values straight to the coordinates as given in the pooled bird count data. Finally, the additional categorical indicator variable owf_zone for inside OWF affected area vs. outside OWF affected area has been defined for two different zones: 1) inside: 3 km vs. outside: > 3 km (as measured from the nearest turbine), as OWF associated ships mainly operated in a radius of up to 3 km around the OWF and this distance class has been used already in previous studies on the impact of single OWFs (Vanermen et al. 2015; Welcker & Nehls 2016) and 2) inside: 10 km vs. outside: > 10 km as a first analysis showed the strongest decrease in diver densities to a distance of 10 km

13 Set up and validation of regression models for the BACI-approach The BACI-approach is based on surveying a potentially impacted situation and a control situation before the impact (variable period ), and relative comparisons of spatial and temporal differences can then be used to extract the unbiased impact (Smith 2002; Schwarz 2014). We formulate the BACI-approach within the framework of generalised additive mixed models (GAMMs) which are well known to appropriately describe biological count data (Zuur et al. 2007; 2009; 2012). We use a continuous linear or smooth predictor measuring the distance to the border of the next OWF. This allows estimating how abundance of divers changes depending on the distance to OWFs and furthermore estimating avoidance distances. Especially, a variable for the observation method ( visual ship based surveys vs. visual aerial surveys vs. digital aerial surveys ) has been introduced as a random intercept in order to account for differences in detection between these methods. We are aware that this variable is partially collinear with the variable period, since in the after period only digital aerial surveys have been used and in the before period, only visual surveys have been performed. Importantly, the estimation of the interaction term period x wind_farm (see below) representing the BACI-approach of the wind farm (which is in the main focus of this study) is not influenced by this, since here, only relative differences in diver densities are evaluated. This results in the following full model for the BACI-approach (not yet thinned regarding its predictors; see below): log(y ij ) = β 0 + u i + f(mean_depth j ) + f(dist_coast j ) + f(dist_helgoland j ) + s(latitude, longitude) + [wind_farm j ] + period j + [wind_farm j ] x period j + offset(log(area j )) + ε ij (1)

14 with ε ij ~ N(0, ϭ 2 ) and u i ~ N(0, ϭ 2 u) independent and identically distributed. Here, y ij is the vector of bird numbers, where the index j refers to the observation number and i is related to the method-id. f() depicts either a linear term or a cubic regression spline s() (which have been tested both during predictor selection), where in the case of a spline the optimal number on knots has been estimated via cross-validation. The variable [wind_farm j ] was either considered as a linear term, dist_owf j measuring the distance to the next wind turbine, or as an additive smoother, s(dist_owf j ), or as a bivariate variable, owf_zone j, the latter distinguishing between inside OWF affected area and outside OWF affected area. For each model, an appropriate probability distribution has been selected for y ij via AIC analysis (see below). In order to validate the optimal GAMM-model, we modified the common selection and validation strategies (Field et al. 2012; Zuur 2012; Zuur et al. 2009; 2010; 2012; Korner-Nievergelt et al. 2015) by the following steps: (1) Based on the entire model (1) we have selected an appropriate probability distribution / stochastic part of the model using the Akaike Information Criterion (AIC; Aikaike 1973). Namely, we compared a poisson-, negative binomial-, Tweedie-, a zero-inflated Poisson-distribution, and an observationlevel random intercept Poisson model. All five probability distributions are known to describe the stochastic part in regression models of (overdispersed) count data reasonably well (Kokonendji et al. 2004; Linden & Maentyniemi 2011; Zuur et al. 2012; Korner-Nievergelt et al. 2015). (2) The optimal model regarding the set of fixed effect predictors was selected from the full model by comparing 16 different models. 13

15 (3) Model validation was carried out by a visual inspection of the residual plots to assess all required model assumptions (Zuur et al. 2010). Corresponding auto-correlation structures were added to the model if required. According to AIC the negative-binomial distribution was favoured, and subsequent predictor selection revealed the following final model: log(y ij ) = β 0 + u i + β 1 dist_coast j + s(latitude, longitude) + [wind_farm j ] + period j + [wind_farm j ] x period j + offset(log(area j )) + ε ij (2) The residual analysis revealed no violation of linearity, homogeneity, independence or normality of the random intercept Integrating covariates for the ship-model As ship traffic has been shown to significantly affect distribution of divers (Bellebaum et al. 2006; Schwemmer et al. 2011), and ship traffic has strongly increased due to construction and maintenance of OWFs in the study area, it is important to disentangle the effects of those two sources of anthropogenic activities on divers. As ship traffic varied throughout the day with increased traffic in the morning and evening hours, this temporal inhomogeneity has to be taken into account. Therefore it was necessary to consider the data as spatio-temporally instead of pure spatially as with the BACI-approach. Here only the data of five different digital survey flights from the after period were used as no real-time ship data were available for the before period or for any other different survey days in the after period. Bird data were spatially assigned to an optimal grid of 2.5 x 2.5 km for every survey day separately and treated as described above. To consider the time we calculated for 14

16 each grid cell additionally the mean time at which the diver observations have been recorded. Data on ship distribution have been recorded parallel to the digital survey flights to record diver distribution using an AIS spotter ( As ship data consist of position data irregular in time and space, they had been linearly interpolated to obtain positions at least every minute. In order to merge ship data with diver distribution, for each time point t and each pair of spatial coordinates (x, y) it has been assumed that all ships within the time interval [t δ t, t] and within a circle around (x, y) with radius r may influence bird density. Since the optimal values δ t and r are not known a priori, we tested all existing combinations between δ t ε {2, 60, 120, 180, 250, 300, 350, 400, 600, } sec and r ε {1, 2, 3, 4, 5, 6, 7, 8, 9, 10} km, and created for each of the 100 combinations a separate variable counting all ships within the given time- and space interval. Here, δ t = depicts the case where all available ship data have been only spatially correlated to bird count data without considering temporal distance to the observations. Subsequently, we compared 100 resulting regression models (see below) to find the optimal values δ t and r. The AIC value, however, for this regression model is not an appropriate measure to select the optimal model, since ship densities and OWF-related variables are collinear, and thus the model with only one of both variable types will be favored due to the parsimony of the AIC-based selection. In contrast, we aimed to explicitly consider both (collinear) variables to explicitly distinguish between the unique effects of ships and those of the wind turbines on diver abundance. Thus, an appropriate measure should relate the effect size of the ship-dependent variable with its reliability. Hence, we selected the model with the highest β / SE β value, where β is the ship-related regression coefficient and SE β is its standard error. 15

17 Set up and validation of regression models for the ship-model The GAMMs were set-up as described above for the BACI-approach. Especially, the ID of the digital survey flight has been introduced as a random intercept to incorporate different amounts of birds or different monitoring conditions between different surveys. This results in the following GAMM structure of the ship-model (not yet thinned regarding its predictors): log(y ij ) = β 0 + u i + f(mean_depth j ) + f(dist_coast j ) + f(dist_helgoland j ) + [wind_farm j ] + [ship_number j ] + offset(log(area j )) + ε ij (3) with ε ij ~ N(0, ϭ 2 ) and u i ~ N(0, ϭ 2 u) independent and identically distributed. Here, y ij is the vector of bird numbers, where the index j refers to the observation number and i is related to the survey flight ID. f() depicts a linear or smooth predictor (which has been both tested during AIC-based predictor selection). The variable [wind_farm j ] was either considered as a binomial predictor ( inside vs. outside ), a linear term (distance to the OWF border), or a cubic regression spline depending on the latter. The variable [ship_number j ] was considered as the total number of temporally and spatially related ships, additionally depending on the a priori defined parameters δ t and r (see above). In contrast to the BACI-approach, here, we did not consider a spatial smooth since this predictor would interfere the correct estimation of [wind_farm j ]. GAMM model selection and validation strategies have been performed as described for the BACI-approach (see above) including integration of appropriate autoregression structures (if required). 16

18 AIC-based selection of the probability distribution favoured again the negativebinomial distribution. The search for the optimal values δ t and r in order to blend observation with ship data showed that highest (β/se)-values (indicating a high precision for the ship-related regression coefficient) could be found δ t = 5 min and r = 5 km. Subsequent predictor selection revealed the following final model: log(y ij ) = β 0 + u i + f(mean_depth j ) + β 1 dist_coast j + s(dist_helgoland j ) + [wind_farm j ] + [ship_number j ] + offset(log(area j )) + ε ij (4) where s() depicts the cubic regression splines with optimal degrees of freedom estimated via cross-validation. Analysis with different sizes of the underlying spatial grid for spatio-temporal pooling reveals an optimal grid size of 2.5 x 2.5 km leading to a temporal autocorrelation of model residuals of order 2 (in contrast to the model based on raw-data, where the AR-order is > 30). Model validation plots indicate no violation of linearity or homogeneity, spatial residual plots and a semivariogram indicate no violation of spatial independence and a pacf-plot reveals a temporal autocorrelation of approximately order 2 which has been integrated as an AR(2)-structure into the model RESULTS Diver abundance before and after the installation of OWFs The spatial distribution patterns of divers changed profoundly from the before to the after period (Fig. 2): In the before period the core area with the highest densities of divers was clearly overlapping with the area of the planned Butendiek wind farm. 17

19 ANHANG I 420 Moderately high densities stretched out to the area of the planned Helgoland 421 Cluster. Contrastingly, in the after period there was a clear shift to the area located 422 between these two windfarm sites (Fig. 2). The area of the OWFs themselves as well 423 as the direct vicinity showed extremely low abundance values of divers in the after 424 period. Thus, the core area of divers in the after period was still located in the centre 425 of the SPA but much stronger aggregated within a still undisturbed sea area. 426 Fig. 2: Spatial density plots of the predicted distribution of divers before vs. after 427 the construction of OWFs based on the BACI-GAMM We furthermore introduced the distance to the wind farm as a smooth term, 430 separately estimated for each period. This revealed a striking difference between 431 both periods (Fig. 3): While the plot for the before period suggests that the future 432 wind farm areas were sites with naturally increased diver abundance, the plot for the 433 after period shows a strong decline in abundance due to the OWFs (Fig. 3). The 434 start of this decline is already visible at > 20 km distance to the OWFs (see also 435 dotted black lines in Fig. 2b). To evaluate at which distance to the wind farm the 18

20 decline in abundance is significant, we approximated the first derivative of the corresponding smooth (Fig. 3 after ) by calculating its first finite difference Fig. 3: Change of diver abundance with the distance to the closest OWF site before (left) and after the construction of OWFs. Smoothed curve: predicted number of divers at a given distance to the closest OWF; shaded area: 95% confidence band; small lines on the x-axis: observations of divers at a given distance to the OWF To answer the question at which distance the change of diver density becomes statistically significant, we calculated confidence bands of the first derivatives via bootstrap and subsequently evaluated where the lower band of the confidence interval intersects with zero. This is the case at around 16.5 km distance to the OWFs 19

21 (Fig. 4). Strongest decline in density is, however, visible from 10 km distance to the OWF (Fig. 3, 4). The avoidance of wind farms within this 10 km distance band is also clearly visible in the distribution maps (solid black lines in Fig. 2b) Fig. 4: First finite difference of the smooth depending on the distance to the closest OWF, partially evaluated for the after period. Red line indicates a derivative of zero, blue line indicates at which distance value the derivative becomes statistically significant. Thick black line corresponds to the first derivative, thin black lines depict 95-%-confidence bands Additionally, the binomial wind farm-related variable owf_zone turned out highly significant for both radii (3 or 10 km, respectively). It reveals that the abundance of divers decreased highly significantly by 94.5 % inside the 3-km zone around the OWFs within the study site (interaction term in Table 1a; β = -2.9, p < ). Inside 20

22 the 10 km zone the abundance still decreased by 83.7 %; (Tab. 1b, β = -1.8, p < 0.001). The distance to land (dist_coast) had no significant effect on diver densities (Tab. 1) Table 1: Regression results of the BACI-approach-GAMM using the binomial variable inside wind farm vs. outside wind farm (owf_zone). (a) 3 km, (b) 10 km. (a) Estimate Std.Error z value Pr(> z ) (Intercept) period[after] owf_zone[inside] > dist_coast period[after]xowf_zone[inside] > (b) Estimate Std.Error z value Pr(> z ) (Intercept) period[after] owf_zone[inside] >

23 dist_coast period[after] xowf_zone[inside] > Distinguishing the effect of ships from OWFs If the ships are included in the overall model as a predictor for the after period, diver densities are still reduced as is the case without considering the effect of ships as shown above. When applying a 3 km radius around the wind farms the diver density is reduced by 70.8 % by OWFs alone compared to the sea areas outside the OWFs (p < 0.001; Table 2). If we extend the radius to 10 km distance around the OWFs, the diver density is still reduced by 44.5 % (p < 0.001) by OWFs alone Table 2: Regression results for the ship-owf-approach-gamm distinguishing the (a) effect of ships from the effect of OWFs in the after period. (a) 3 km, (b) 10 km. Estimate Std.Error z value Pr(> z ) (Intercept) owf_zone[inside] > dist_coast n_ships > (b) Estimate Std.Error z value Pr(> z ) (Intercept)

24 owf_zone[inside] < dist_coast n_ships < If ships are removed from the model as single predictor, the effect of OWFs (now joined with the effect of the ships) on divers is estimated to be 84 % in a 3 km radius (p< 0.001). Thus, it can be concluded that ships also have a strong negative effect on diver abundance, here estimated to account for at least 14% of the joint OWFship-effect. Thus, it turns out for the ship-model that the pure effect of OWFs alone is not as strong as estimated by the BACI-approach (i.e. without considering ship traffic; > 94% and > 84%, respectively). There are two reasons for those different estimations: (1) the ship-model was only fitted using data from the after period, as in the project there were no ship data available for the before period. Hence, the estimated reduction effect does not take into account that bird densities within the OWFs showed the highest diver abundances before the construction of the farms (see above) which leads to a strong underestimation of the reduction effect. (2) The shipmodel considered the effect of ships which are, however, at least partially correlated with OWF location (Fig. 5). Thus, the BACI-approach actually estimated the joint reduction effect of OWFs and ships, whereas the ship-model evaluated both impacts separately, which possibly led to a reduction of the OWF effect compared to the BACI-approach. 23

25 Fig. 5: Spatial density plot of ship distribution in the after period based on AIS data The ship-model indeed showed a significant negative impact of ships on diver abundance (Table 2). With each additional ship in the spatio-temporal range of divers (i.e. 5 min and 5 km from the diver sighting; see material and methods) the abundance declined highly significantly by 31 % (p< 0.001). This means that every third individual leaves the area if one ship approaches. The spatial component of disturbance by a ship was much stronger compared to the temporal component, i.e. our regression models selecting for the optimal δ t and r revealed that ships in the vicinity of 5 km had a strong impact on diver abundance, whereas the time lag between the diver sighting and the AIS signal of the ship was less relevant (with an 24

26 optimum at approx. 5 min). From this, it can be concluded that ships seem to affect divers most strongly in a distance of 5 km. Same as in the BACI-approach the distance to land had no significant influence on diver abundance (Table 2) DISCUSSION Distribution patterns before and after the OWF installation Our results clearly show that the distribution patterns of divers that were stable over a period of many years (Garthe et al. 2015) were altered substantially on a small and large spatial scale due to the installation of OWFs in the German North Sea. We were able to build our BACI-approach on a very solid data base that included 14 years of large-scale surveys in the before period. To our knowledge all other available studies are based on a maximum of 1-3 years of data prior to the construction of OWFs only and mostly focused on the effect of single OWFs (e.g. Leopold et al. 2013, Petersen et al. 2014). Although it was not possible in our study to compute absolute differences in diver population between the two periods due to a change in survey methods, our results clearly show that there were profound largescale shifts in distribution patterns as well as significant avoidance of the OWF areas. We observed a shift of the diver abundance hotspot to the western-central area of the SPA that was still undisturbed by OWFs in the after period. The hotspot is located in a distance of about 20 km of all surrounding OWFs. Several studies already point out which environmental parameters are most important in determining diver distribution patterns: Besides frontal systems, nearshore and shallow sandy sea areas play a major role (Skov & Prins 2001, O Brien et al 2008, Skov et al. 2016, Heinänen pers. comm.). Particularly frontal systems are expected to increase prey 25

27 availability for divers (Skov & Prins 2001). Our study even suggests that exactly the area of the OWF Butendiek which had been installed in the northern part of the SPA was of particular importance for divers before the construction of this OWF, as diver abundance showed maximum values in this area during the before period. The OWFs of the Helgoland-Cluster are located south-west off the border to the SPA. Our study showed that in contrast to the Butendiek area the abundance values in the before period were significantly lower as compared to within the SPA. However, divers are known to occur here regularly (Garthe et al. 2015). One aim of this study was to disentangle the importance of natural habitat structures and anthropogenic pressures on divers. As shown by our modelling approach the natural habitat predictors such as distance to coast / island of Helgoland and water depth did not play a major role as compared to the effect of OWFs and shipping (see below). Thus, this suggests that anthropogenic pressures are by far the most important factor to drive distribution patterns of divers within their natural hot-spots. Incorporating the distance to the next OWF as a smoothed term in the model, we were able to highlight that divers showed reactions already in a distance of around 20 km to the OWFs with significant changes of densities in distances of 16.5 km and strongest change in abundance in distances of 10 km. These values are by far higher as compared to all other published studies (summarized in Welcker & Nehls 2016; Dierschke et al. 2016). However, previous studies have mostly investigated local avoidance effects only (often only up to 4 km distance; Petersen et al. 2006, Petersen & Fox 2007, Leopold et al. 2013, Welcker & Nehls 2016) and therefore were not able to detect any larger-scale avoidance reactions. This indicates the importance of a sufficiently large-scale approach and the inclusion of multiple OWF sites (Rexstad & 26

28 Buckland 2012) as done in the current study. To highlight the question of scale, we have quantified the effect of OWFs on divers by defining both, a 3 km and a 10 km distance class as affected sea areas. The 3 km distance class was chosen for this study as previous studies on single OWFs have found avoidance distances up to this value (Vanermen et al. 2015; Welcker & Nehls 2016). However, as our results suggest, this distance seems to be far too short in a study where the effect of multiple OWFs are considered on a larger spatial scale. The reason for the rather large-scale effect of OWFs on divers as detected in our study is not completely clear. It is well possible that visual cues alone are not the only reason for a large disturbance distance. Previous studies point out that OWFs do not only directly affect seabirds and other marine wildlife (Lindeboom et al. 2011, Bergström et al. 2014, Goodale & Milman 2014), but they may additionally cause changes in the abiotic environment such as sediment properties and the stratification of the water due to turbulences caused by the piles (Carpenter et al. 2016; Nagel et al. 2018). Carpenter et al. (2016) point out that an individual OWF may lead to an enhanced mixing of the water column and cascading effects on the whole ecosystem (although still little is known on physical-biological interactions in detail) in a surrounding area of km of the OWF. This matches very well with the disturbance distance of divers found in the current study. Petersen et al. (2014) have shown significantly lower diver abundance in areas of up to 13 km from the wind farm. This also matches well our larger-scale approach Distinguishing the effect of ships from OWFs 27

29 The installation of OWFs causes a substantial increase of ship traffic due to maintenance and service in the surrounding area (Exo et al. 2003). Although it is well known that ship traffic may substantially affect distribution patterns of seabirds and particularly of divers (Bellebaum et al. 2006; Schwemmer et al. 2011), the combined effect of OWFs and associated ship traffic has only rarely been previously studied; but the few available studies point out a significant impact of ship traffic on diver distribution (Leopold et al. 2014, APEM 2013, APEM 2016, Skov et al. 2016, S. Heinänen pers. comm.). A behavioural response of divers when confronted with approaching ships has been clearly demonstrated and flight distance of up to 2 km have been documented (Bellebaum et al. 2006; Schwemmer et al. 2011). This corresponds to the results of our study which suggests a significant reduction of diver densities in an area of up to 5 km in the vicinity of ships (and only a little effect of the temporal aspect of ship distribution). When including the abundance of ships in the model, our study shows a reduced density of divers of up to 70% in the 3 km distance zone of the OWFs. This reduction effect can be viewed as the effect exhibited by the OWFs alone. The joint effect of OWFs and ships led to a reduction of 84%. This makes clear that ships have an additionally negative impact on diver densities. The exact reduction in densities by ships alone cannot be computed reliably as both ship traffic and OWFs are strongly collinear. More importantly, due to their mobile nature, ships are a spatially and temporally variable predictor and thus a reliable estimation of their overall effects on birds will always be biased. This will remain hard to solve even in future studies as ships strongly aggregate in the vicinity of OWFs and present no fixed predictor. The strong increase in reduction of diver densities when including ship traffic in the model makes very clear that it is utterly important to review the cumulative impact of 28

30 multiple anthropogenic pressures in the marine environment. Previous studies have focussed on cumulative effects by simply investigating the combined effects of multiple OWFs (Dierschke et al. 2003, 2006, Fox et al. 2006, Desholm 2009, King et al. 2009, Mendel & Garthe 2010; Busch et al. 2013). However, given the strong effect of ships on diver abundance it seems necessary to include also other anthropogenic pressures to estimate their cumulative effect on diver abundance in general Application of the results for management Due to the large-scale avoidance effect of divers with respect to OWFs (and ships), it is very unlikely that divers suffer from enhanced direct mortality, e.g. because of collision (Petersen et al. 2006; Leopold et al. 2010; this study). Furthermore a low flight altitude of only up to 10 m above the sea surface (van Bemmelen et al. 2011) is reducing collision risk. Thus, indirect effects, such as habitat loss, are very likely to play the key role for divers. However, particularly the consequences of such indirect effects e.g. on population level of seabirds and density dependent effects are very hard to assess and proper methods are still broadly missing (Green et al. 2016; Horswill et al. 2017). When judging the consequences of habitat loss due to installation of OWFs and associated enhanced ship traffic it is essential to shed light on the question on alternative sea areas that can be used as resting and foraging ground. In the case of the current study these alternative sites seem to be very limited as the SPA is virtually surrounded by OWFs. This might explain why divers tend to concentrate in the centre of the SPA rather than shift their distribution outside of it. It was not yet possible to compute absolute differences in abundance between the before and after period. Hopefully, this will be solved soon, when there will be enough data available from parallel digital and visual surveys from sea areas where 29

31 visual observations are still allowed. However, the results on relative reduced densities of divers with respect to OWFs and ship traffic as well as avoidance distances as provided in the current study serve as a baseline for further studies. A suitable approach to quantify the overall habitat loss for divers would be to compute the relative proportion of habitat loss within a certain area (e.g. within the SPA). Dierschke et al. (2006) suggest to sum up the total of the wind farm areas and to add an area of a buffer zone to assess the overall habitat loss. When we apply this approach to our study, the minimum habitat loss due to the OWFs in the SPA can be computed: If at least the sea area within a 3 km distance class around the OWFs can be regarded as completely lost for divers (which is strongly supported by the current study) it would mean that 8.8% of the SPA (overall size 3,135 km²) currently is lost for divers. This value should be regarded as an absolute minimum as our results clearly show that the density of divers is still strongly reduced in areas 3 km distance from the next OWF. Although we are not able to compare any absolute density value between the before and after period, our study indicates that divers seem to strongly concentrate in the centre of the SPA which is likely going along with an increase of divers in a much smaller sea area. Given that divers tend to occur in comparatively small flocks only occasionally exceeding 5-10 individuals / km² (O Brien et al. 2012; Garthe et al. 2015), this might promote density dependent effects (Lewis et al. 2001; Blanc et al. 2006; Horswill et al. 2017). A possible shift towards suboptimal habitats may lead to a suboptimal body condition prior to breeding, which may lead to a lower reproductive success and enhanced mortality in adult birds (Coulson et al. 1983; Hüppop 1995). In divers, a slight increase in adult mortality by only 0.3% can lead to significant negative effects on population level (Rebke 2005). 30