Predicting the distribution of deep-sea vulnerable marine ecosystems using high-resolution data: Considerations and novel approaches

Deep-Sea Research I 93 (2014) 72–82

Contents lists available at ScienceDirect

Deep-Sea Research I journal homepage: www.elsevier.com/locate/dsri

Predicting the distribution of deep-sea vulnerable marine ecosystems using high-resolution data: Considerations and novel approaches Anna M. Rengstorf a,n, Christian Mohn b, Colin Brown c, Mary S. Wisz d, Anthony J. Grehan a a

Earth and Ocean Sciences, School of Natural Science, NUI Galway, Galway, Ireland Department of Bioscience, Aarhus University, Roskilde, Denmark c Ryan Institute for Environment, Marine and Energy Research, NUI Galway, Galway, Ireland d Department of Ecology and Environment, DHI Water and Environment, Hørsholm, Denmark b

art ic l e i nf o

a b s t r a c t

Article history: Received 18 March 2014 Received in revised form 7 July 2014 Accepted 16 July 2014 Available online 2 August 2014

Little is known about species distribution patterns in deep-sea environments, primarily because sampling surveys in the high seas are expensive and time consuming. The increasing need to manage and protect vulnerable marine ecosystems, such as cold-water corals, has motivated the use of predictive modelling tools, which produce continuous maps of potential species or habitat distribution from limited point observations and full coverage environmental data. Rapid advances in acoustic remote sensing, oceanographic modelling and sampling technology now provide high quality datasets, facilitating model development with high spatial detail. This paper provides a short overview of existing methodologies for predicting deep-sea benthic species distribution, and illustrates emerging issues related to spatial and thematic data resolution, and the use of transect-derived species distribution data. In order to enhance the ecological relevance and reliability of deep-sea species distribution models, novel techniques are presented based on a case study predicting the distribution of the cold-water coral Lophelia pertusa in three carbonate mound provinces in Irish waters. Specifically, the study evaluates (1) the capacity of newly developed high-resolution (250 m grid cell size) hydrodynamic variables to explain local scale cold-water coral distribution patterns, (2) the potential value of species occurrence proportion data to maintain semi-quantitative information of coral prevalence (i.e. coverage) and sampling effort per grid cell within the response variable, and (3) mixed effect modelling to deal with spatially grouped transect data. The study shows that predictive models using vertical and horizontal flow parameters perform significantly better than models based on terrain parameters only. Semiquantitative proportion data may decrease model uncertainty and increase model reliability, and provide a fruitful avenue of research for analysing large quantities of video data in a detailed yet time-efficient manner. The study concludes with an outlook of how species distribution models could improve our understanding of vulnerable marine ecosystem functioning and processes in the deep sea. & 2014 Elsevier Ltd. All rights reserved.

Keywords: Cold-water corals Generalized linear model Hydrodynamic modelling Lophelia pertusa Species distribution modelling

1. Introduction Due to the great cost and time required for its exploration, the deep sea remains one of nature’s great scientific frontiers. Merely 5% of the deep-sea floor has been explored remotely and less than 0.01% has been sampled (Ramirez-Llodra et al., 2010). Given the deep sea’s value in providing important goods and services, such as commercially important fish, chemical compounds for pharmaceuticals, oil, gas and minerals, amongst others (Armstrong et al., 2012), and its vulnerability to increasing anthropogenic impacts (Davies et al., 2007), there is an urgent need to deepen our knowledge of deep-sea ecosystems, to map their spatial distribution, and to

n

Corresponding author. Tel.: þ 44 759 777 5122. E-mail address: [email protected] (A.M. Rengstorf).

http://dx.doi.org/10.1016/j.dsr.2014.07.007 0967-0637/& 2014 Elsevier Ltd. All rights reserved.

ensure their sustainable management and conservation. This has motivated the use of predictive modelling tools that relate species occurrence data with environmental predictor variables to estimate full-coverage species distribution in geographic space (Elith and Leathwick, 2009; Guisan and Zimmermann, 2000). Species distribution models (SDMs, also known as habitat suitability models, environmental niche models or resource selection functions) have been used for a variety of applications (Elith and Leathwick, 2009), including the prediction of the potential distribution of invasive species (Tyberghein et al., 2012), the estimation of the impact of a changing climate on future distributions (Tittensor et al., 2010), and conservation planning (Carroll, 2010). The effort to develop SDMs in the deep sea has mainly focused on cold-water corals (CWCs), driven by national and international obligations to develop conservation measures for these vulnerable marine ecosystems (VMEs, Rogers et al., 2008). Vierod et al. (2013) have recently reviewed the

A.M. Rengstorf et al. / Deep-Sea Research I 93 (2014) 72–82

application of SDMs to those deep-sea benthic species that typify VMEs. The authors draw attention to general modelling issues related to spatial scale, sampling bias (clustering of data in certain areas) and spatial pseudo-replication, amongst others, and particularly highlight the limitations of currently available deep-sea data. The present study addresses some of these data limitations by exploring the use of full coverage, high-resolution ( 250 m) hydrodynamic variables derived from a newly developed oceanographic model, to explain and predict CWC distribution. It further proposes a novel modelling approach using high-quality, semi-quantitative occurrence proportion data derived from standardized video transects, ultimately aiming to enhance the ecological relevance and reliability of deep-sea SDMs. Conventionally, SDMs are fit with information on both the presence and absence of a species (Guisan and Zimmermann, 2000). The difficulty to establish reliable absence data in deep-sea environments, however, has prompted most studies targeting deep-sea VMEs to make use of presence-only or presencebackground modelling techniques (Vierod et al., 2013). Species presence data can originate from a variety of sources including fisheries bycatch, trawling and dredging surveys, boxcores and grab samples, video and photographic surveys. Such multi-source data vary significantly in quality, with potential issues including poor spatial precision, taxonomic errors and lack of metadata (e.g. information on the organism’s life stage), amongst others. Quality control of readily available distribution data is therefore essential to ensure realistic representation of the target species (Ross et al., 2012; Rengstorf et al., 2013). Important environmental drivers affecting the spatial distribution of deep-sea benthic species include primary productivity, near-seabed water chemistry and hydrodynamics, terrain morphology and seabed substrate (e.g. Davies and Guinotte, 2011; Howell et al., 2011; Rengstorf et al., 2013; Yesson et al., 2012). Fullcoverage environmental data can be derived from ship- or satellite borne remote sensing (such as sea surface temperature and primary productivity), point-interpolations from in-situ measurements, ocean circulation models, and hybrids of these. A wide range of global scale data layers are readily available and have facilitated spatial predictions of cold-water sclaractinians and octocorals on a global scale (Tittensor et al., 2009; Davies and Guinotte, 2011; Yesson et al., 2012). While the resulting maps provide valuable biogeographic information and guidance for future survey efforts, their reliability and application to marine management is limited by the relative coarseness of the underlying environmental data (Vierod et al., 2013). An important consideration when predicting the distribution of sessile species, such as CWCs, is that all habitat requirements must coincide at exactly the same spatial location (Guisan and Thuiller, 2005). Thus, high-resolution data are required to minimize spatial mismatch between distribution records and corresponding environmental conditions, and between the different environmental conditions. Davies et al. (2008), for example, showed that a 11 by 11 temperature grid was too coarse to resolve changes in water temperature, leading to a mismatch between CWC occurrences and temperature values beyond the species’ thermal tolerance limit. Further spatial scale issues may arise when the environmental variables used in the predictions do not resolve, and thus do not reflect, the environmental factors or processes actually shaping the species distribution. Rengstorf et al. (2012) found that bathymetric data of 1 km resolution are too coarse to resolve the often small carbonate mounds that support CWC growth in Irish waters, and stressed the need for high-resolution bathymetry data to avoid over-estimation of the predicted coral habitat. Further, several authors (e.g. Frederiksen et al., 1992; White et al., 2005) have demonstrated, using in-situ current measurements, the preference of benthic suspension feeders for enhanced current

73

flows. However, broad-scale SDMs predicting CWC distribution have failed to identify this relationship, a fact attributed to the coarse resolution of the hydrographic data (SODA, Carton et al., 2005), resulting in a lack of information on topographically influenced and locally varying factors (Davies and Guinotte 2011; Yesson et al., 2012). Ocean circulation models of high spatial resolution can simulate ecologically relevant processes, such as internal wave dynamics and resonant tidal amplification (Mohn et al., 2014). Integration of such hydrodynamic data is highly desirable, as it could add to the SDMs ecological relevance and provide valuable insight in critical processes such as larval dispersal, sediment settling rates, and – most importantly – food supply. The present study investigates the value of such high resolution hydrodynamic variables for predicting deep-sea benthic species distributions. The spatial scale issues discussed above, coupled with the need for more detailed maps relevant for marine spatial planning, have motivated an increasing number of SDMs at finer spatial scales (Vierod et al., 2013). In contrast to global, explorative SDMs, these regional or local scale models are usually targeted towards specific areas of interest, where VMEs are expected to occur. Due to the lack of high-resolution, near-seabed oceanographic data, most 'fine scale’ models to date are exclusively based on bathymetric data, which can be derived at high resolution from multi-beam echosounders (Brown et al., 2011). Bathymetry-derived parameters such as seabed slope and bathymetric position index (Weiss, 2001) act as proxies for seabed substratum (Dunn and Halpin, 2009) and current flow (Genin et al., 1986), and have been used to predict distribution patterns of a variety of benthic biota (e.g. Kostylev et al., 2001; Holmes et al., 2008), including CWCs (e.g. Dolan et al., 2008; Guinan et al., 2009a; Woodby et al., 2009). Video and photographic surveys have evolved to be the standard method for such regional habitat mapping and modelling efforts (Brown et al., 2011), as they are non-destructive, provide information on seabed substrata and biological assemblages, and allow for precise spatial matching with high-resolution environmental data (e.g. Dolan et al., 2008). In a few fine-scale studies, standardized video transects could even provide (relatively reliable) coral absence data, enabling the use of conventional presence–absence modelling techniques, such as generalized linear models (GLM, Guisan and Zimmermann, 2000), to predict CWC distribution (Woodby et al., 2009; Marshall, 2012). Grid cells containing at least one occurrence observation are coded as “present” (1), whereas grid cells where the species was not observed are coded as “absent” (0), assuming that the limited area surveyed by the camera’s footprint is representative of the entire area encompassed by the grid cell. Fig. 1 illustrates two main shortcomings of this simplification; it eliminates information on the species’ prevalence (i.e. abundance or coverage) within the grid cell, and it eliminates information on how much of the grid cell has been actually surveyed (i.e. sampling effort). In this paper we explore the

Fig. 1. Scheme showing observations of species presence (filled circle) and species absence (open circle) along a video transect in context with the environmental data grid cells (large squares). The corresponding response codes specified for a presence–absence model (PA) and a proportion model (PR) are given for each cell. The use of proportion data retains information on both prevalence (n presences/n absences) and sampling effort (n presencesþ n absences) within each grid cell.

74

A.M. Rengstorf et al. / Deep-Sea Research I 93 (2014) 72–82

effectiveness of a simple modification of the response variable to proportion data (as opposed to presence–absence data), which retains information on both prevalence (n presences/n absences) and sampling effort (n presencesþn absences) within each grid cell. While such proportion data are commonly used in, for example, medical studies (Crawley, 2007), the approach has not yet been applied in deep-sea SDM. The aim of this study is to extend the limits of existing methodologies for SDM of VMEs in the deep sea, by using appropriately scaled, ecologically relevant predictor variables, and by maximising the information content of the species distribution data. On the basis of a case study targeting the framework building CWC Lophelia pertusa, the objectives of this paper are to explore:

the open slope (i.e. likely non-coral areas). ROV transects were acquired at an altitude between 1.5 and 2.5 m above seabed and at a constant speed of  0.6 knot ( 0.3 m/s). The camera’s field of view was roughly 2.5  2.5 m depending on seabed topography. The presence and absence of L. pertusa framework was annotated at intervals of  5 m and records were geo-referenced using the ROV navigation data. This resulted in a total of 900 data points, including 25 presences and 249 absences within the Arc mound province, 98 presences and 184 absences within the Belgica province, and 117 presences and 227 absences within the Logachev province.

1) the applicability of high-resolution (250 m), simulated hydrodynamic parameters to explain and predict local-scale distribution patterns of vulnerable marine ecosystems; 2) the effectiveness of proportion data as a response variable; 3) the applicability of generalized linear mixed models (GLMM), as an alternative to the traditional GLM, in order to deal with spatially grouped (i.e. pseudo-replicated) transect data.

A large range of bathymetric (n ¼6) and hydrodynamic (n ¼17) variables were generated and tested for their potential in explaining coral distribution (Table 1). Within-parameter variations were moderate, and no transformations were performed prior modelling.

L. pertusa was selected as a target species for this case study, as it forms an emblematic VME and its habitat requirements are now largely understood (Roberts et al., 2009). Transect-based distribution data have been collected in a standardized manner, which enabled the generation of semi-quantitative proportion data and justified the use of mixed modelling techniques. Three study areas were chosen to assess species-habitat relationships in different environmental settings and to further allow for investigations of model transferability.

2. Materials and methods 2.1. Study areas Three study areas comprising CWC carbonate mounds off the West of Ireland were investigated (Fig. 2). The Logachev mound province (Fig. 2a) on the southern margin of the Rockall Bank is a belt of giant carbonate mounds with heights 4 300 m and widths of several kilometres, which are predominantly arranged in downslope oriented clusters (Kenyon et al., 2003; Mienis et al., 2006). The Arc mound province (Fig. 2b) on the south-western Porcupine Bank is made of much smaller mound features  50–100 m high and base lengths seldom exceeding 500 m (Rengstorf et al., 2012). The Belgica province (Fig. 2c) on the eastern Porcupine Seabight is roughly arranged in north-south trending ridges. Mounds are up to 100 m high and may reach base lengths up to 2000 m (Wheeler et al., 2007). Video footage has revealed thriving L. pertusa reefs on summits and terraced flanks of the Logachev mounds (Olu-Le Roy et al., 2002), and on the summits of the Arc mounds (Grehan et al., 2009). In the Belgica mound province corals can be found only on the deeper, western alignment of the mounds, whereas the eastern mounds are characterized by asymmetric drift accumulations, sediment clogged dead corals and coral rubble (Foubert et al., 2005).

2.3. Environmental variables

2.3.1. Bathymetric variables Multibeam data covering the study areas were acquired as part of the Irish National Seabed Survey (INSS). The cleannxyz ASCII data were obtained from the Geological Survey of Ireland by the authors and Fledermaus v.7 gridding software was used to produce a digital elevation model (DEM) with a resolution of 0.00051  0.00051 (WGS84). The DEM was imported into ArcGIS v.9.3 and projected onto UTM Zone 28 N (Logachev and Arc mounds) and UTM Zone 29 N (Belgica mounds) with a resolution of 50 m  50 m. Bathymetric variables were derived following Wilson et al. (2007) and included seabed slope, small scale bathymetric position index (BPI3: inner radius¼1, outer radius ¼ 3), large scale bathymetric position index (BPI25: inner radius ¼1, outer radius ¼ 25), aspect (converted to eastness and northness) and rugosity (using the Benthic Terrain Modeller extension to ArcGIS).

2.2. Coral distribution data

2.3.2. Hydrodynamic variables Hydrodynamic variables were obtained from simulations with the Regional Ocean Modelling System (ROMS, http://roms.mpl.ird. fr/) described in Mohn et al. (2014). The model has a spatial resolution of 0.0021 (  250 m) and 32 terrain-following vertical levels from sea surface to bottom layer (thickness  0.01% of water depth). ROMS model output was generated in the form of averages at 6 h intervals for the time period between the 15th of April and the 15th of May 2010 (corresponding to the period of the CE10014 survey). For each variable (except current direction), grids were generated for mean, maximum and minimum values over the one month time period. Model-derived hydrodynamic variables tested for their importance in predicting coral distribution included current speed (m/s; spdmax, spdmean, spdmin), current direction converted to eastward and northward flow (direast, dirnorth), bottom (shear) stress (N/m; botstrmax, botstrmean, botstrmin), vertical flow (m/s; vmax, vmean, vmin), temperature (1C; tmpmax, tmpmean, tmpmin) and salinity (psu; salmax, salmean, salmin) at the near-bottom layer of the three-dimensional model. The resulting grids were projected and interpolated to match the respective 50 m  50 m bathymetry grids.

Video transects were conducted during the RV Celtic Explorer cruise CE10014 in April/May 2010, employing the Irish remotelyoperated vehicle (ROV) Holland I. In each study area, six transects of  2 km length were acquired (Fig. 2), with three transects crossing the summits of carbonate mounds (i.e. possible coral reef areas) and three transects following the general depth-gradient on

2.3.3. A priori variable selection The choice of predictor variables for a particular species is crucial in SDM as the use of too many variables may result in overfitting and multi-collinearity (Guisan and Zimmermann, 2000). The a priori variable selection involved several steps: (1) variables were clustered into ecologically meaningful groups (Table 1); (2)

A.M. Rengstorf et al. / Deep-Sea Research I 93 (2014) 72–82

75

Fig. 2. Maps showing the locations and high resolution multibeam bathymetry (INSS data) of the study areas: Logachev ((a) 61 km  103 km), Arc ((b) 23 km  38 km) and Belgica ((c) 32 km  56 km) mound provinces. ROV-based video observations of coral presences (PR) and absences (AB) are indicated in red and yellow, respectively.

Table 1 Environmental variables developed for this study, clustered into “ecological meaningful” groups. Variables

Native resolution Source (1)

Bathymetric variables Aspect (1)1, slope (1), BPI32, BPI253, rugosity

0.0005

INSS bathymetry

0.002

Mohn et al. (2014)

Vertical flow variables Vertical flow (m/s)4

0.002

Mohn et al. (2014)

Chemistry variables Salinity (ppt)4

0.002

Mohn et al. (2014) Mohn et al. (2014)

Horizontal current variables Bottom stress (N/m2)4, current speed (m/s)4, current direction (1)1

Temperature (1C)

4

0.002

1

Converted to eastness and northness following Wilson et al. (2007). Bathymetric position index computed with an inner radius of 1 and an outer radius of 3 pixels. 3 Bathymetric position index computed with an inner radius of 1 and an outer radius of 25 pixels. 4 Three near-bottom layers computed for maximum, mean and minimum values over a 30 day time period (15 April–15 May 2010). 2

pair-wise Pearson correlation coefficients between environmental variables were calculated (Appendix A); (3) uni-variate GLMs were computed for each environmental variable using presence– absence data (Appendix A); and (4) within each cluster, the variable with highest explanatory power, measured in % deviance

of the uni-variate GLMs, was retained for further analysis. Besides slope, BPI3 and BPI25 were retained as bathymetric variables because of their only moderate correlations and relatively high predictive power in previous coral SDMs (e.g. Howell et al., 2011). The three hydrographic variables selected for further analyses

76

A.M. Rengstorf et al. / Deep-Sea Research I 93 (2014) 72–82

were maximum bottom stress (botstrmax), maximum vertical flow (wmax) and mean temperature (tmpmean) (see Appendix A). 2.4. Statistical modelling All statistical analyses were conducted using the R software package version 2.14.1 for Windows (R Development Core Team, 2008). GLMs were chosen as they are flexible and descriptive modelling tools (Guisan et al., 2002), that can be easily manipulated and customized in R. The software’s built-in glm function was used with binomial error distribution and logit link, which is appropriate for modelling both presence–absence and proportion data (Crawley, 2007). Response variables for both, presence– absence and proportion modelling, were created as demonstrated in Fig. 1. For the proportion models, R uses the number of presences and the number of absences per grid cell bound together as a response variable and then carries out weighted regressions based on the total number of observations per grid cell. Simple percentage data were not used because: (i) the errors are not normally distributed; (ii) the variance is not constant; (iii) by calculating the percentage, information on sampling effort per grid cell is lost and all observations are equally weighted (Crawley, 2007). Most statistical models rely on strong assumptions about the data such as their need to be random and spatially independent from each other (Hirzel et al., 2002). As is the case in the present study, deep-sea data are often derived from continuous transects, which in turn are nested within study areas of scientific interest. Data originating from the same data groups are spatially pseudoreplicated (i.e. individual observations within a mound transect were more likely to be “presences”, whereas observations in an off-mound transect were more likely to be “absences”), and ignoring their correlation may lead to biased parameter estimates and invalid model results (Bolker et al., 2009). Generalized linear mixed models (GLMMs) are an extension to GLMs in which a linear predictor may contain random effects (i.e. data groups) and within-group errors may be correlated (Dormann et al., 2007; Bolker et al., 2009). The “lme4: Linear mixed-effects models using S4 classes” package (http://CRAN.R-project.org/package=lme4) was used to generate GLMMs, which used the same specifications as the GLMs, but additionally to the environmental variables (i.e. fixed effects), “study area” and “transect station” were specified as nested random effects. The Akaike information criterion (AIC) was used for model selection. The AIC quantifies the relative performances between alternative models by comparing their goodness of fit (deviance) and by penalising more complex models based on the number of parameters they use (Burnham and Anderson, 1998). Therefore it provides a basis for model averaging, which can deliver better parameter estimates and confidence intervals that account for model uncertainty (O’Hara and Tittensor, 2010). Following methods described in O’Hara and Tittensor (2010), multiple models were generated covering the entire subspace of all first-order additive combinations of the six a priori environmental variables (26 ¼64 models). The “Model selection and multimodel inference based on (Q)AIC(c)” package (http://www. CRAN.R-project.org/package=AICcmodavg) was used to rank models based on their AICc (a variant of AIC corrected for small sample sizes), to calculate model-averaged parameter estimates and standard errors, and to estimate relative predictor variable importance (Burnham and Anderson, 1998). Distribution data pooled from the three study areas (n¼ 900) were used for generating GLMs and GLMMs using both presence–absence and proportion data. To assess the usefulness of the hydrodynamic parameters in predicting species distribution, performances were compared between models using terrain parameters only (Terrain model), hydrodynamic parameters only (Hydro model), and bathymetric

and hydrodynamic parameters together (Both model). Species distribution maps were generated using the parameter combinations and coefficients of the highest ranked (based on AICc) models. Finally, model transferability was tested in order to assess the capability of a SDM generated in one study area to predict coral distribution in another. Parameter coefficients were determined using locally trained GLMs (i.e. using data from one study area only) and were then projected onto the remaining two study areas. Presence–absence data excluded from model calibration were used for calculating model evaluation indices for the transferred distributions. The “PresenceAbsence” package (Freeman and Moisen, 2008) was employed to derive the area under the curve (AUC) of the receiver operating characteristic (ROC), as well as specificity and sensitivity values (Fielding and Bell, 1997). The results for locally trained models, using the parameter combination of the overall highest ranked GLM model using presence– absence data, are presented.

3. Results 3.1. Parameter responses The results of our study suggest that all terrain and hydrodynamic (especially vertical flow and bottom stress) variables are important parameters in explaining coral distribution. Fig. 3 shows the uni-variate presence–absence and proportion response curves for the six parameters remaining after the a priori selection. In accordance with previous studies, the probability of coral presence rapidly increased with terrain complexity (slope, BPI3 and BPI25), reflecting the corals’ association with elevated topography. The strong positive response curves of the horizontal and vertical flow parameters (botstrmax and wmax) capture the corals’ preference for locally enhanced current flow and downwelling, preventing smothering by sediment accumulation and ensuring plentiful food supply. The slightly negative relationship with temperature (tmpmean) most likely reflects the temperature range sampled, as all study areas lie well within the temperature range thought to be suitable for the species (  4–13 1C, Freiwald et al., 2004).

3.2. Hydrodynamic vs. terrain-based variables The addition of high resolution hydrodynamic parameters to the modelling framework significantly improved model fit of both GLMs and GLMMs using both presence–absence and proportion data (GLMPA, GLMPR, GLMMPA, GLMMPR). Table 2 shows the top five ranked AICc scores and parameter combinations derived from the GLMPA, as well as the highest ranked models that used only hydrodynamic (rank 43) and terrain (rank 49) variables. Thus, the best 48 out of 64 parameter combinations included at least one of the three hydrodynamic variables available, highlighting their importance in explaining local scale coral distribution. Similar results were found for GLMPR, GLMMPA and GLMMPR (not presented). Table 3 summarizes parameter estimates, standard errors and relative parameter importance averaged over the 64 model runs conducted for the GLMPA, GLMPR, GLMMPA and GLMMPR trials, as well as the percentages of deviance explained by the highest ranked models when using terrain-only (Terrain), hydrodynamiconly (Hydro) or combined (Both) environmental predictor variables. The percentage of deviance explained was consistently higher for the Both models (mean 7standard deviation: 0.49 7 0.04), than for the Terrain (0.39 70.04) and Hydro (0.37 70.03) models. The model averaged relative variable importance revealed

1.0

0.8

0.8

0.8

0.6 0.4

Probability

1.0

0.6 0.4

0.6 0.4

0.2

0.2

0.2

0.0

0.0

0.0

0

5

10

15 20 slope

25

−10

0

5 10 bpi3

20

−50

1.0

1.0

0.8

0.8

0.8

0.6 0.4

Probability

1.0

Probability

Probability

77

1.0

Probability

Probability

A.M. Rengstorf et al. / Deep-Sea Research I 93 (2014) 72–82

0.6 0.4

0.2

0.0

0.0

0.0

0.15 0.20 botstrmax

0.25

0.00

0.04

0.08 wmax

0.12

100

0.4

0.2

0.10

50 bpi25

0.6

0.2

0.05

0

7.5

8.0

8.5 9.0 tmpmean

9.5

Fig. 3. Uni-variate GLM response curves for the selected environmental predictor variables based on data pooled from all study areas (n¼ 900). The solid line shows the presence–absence response curves, the dashed line shows the proportion response curves.

Table 2 A selection of the 64 GLMs generated based on presence–absence data and ranked by the AICc value. The best 42 models use a combination of terrain and hydrodynamic parameters. The best model using only terrain parameters combines slope and BPI3 and was ranked no. 49. Model rank

AICc

Terrain parameters slope (1)

1 2 3 4 5 … 43 … 49 … 64

571 572 573 574 577 … 640 … 677 … 1045

● ● ● ● ● …

Hydrodynamic parameters

bpi3 bpi25 wmax (m/s)

● ● ● …

… … ● ● … … (Null model)

● ● ● ● ● …

● ● ● ●

…

… ● …

…

…

botstrmax (N/m2)

tmpmean. (1C)

● ● ● ● ● …

● ● …

…

…

…

…

high predictive power for slope, BPI25, wmax and botstrmax (Z 0.94), while BPI3 and tmpmean were less important (r0.54). Fig. 4 shows CWC distribution maps generated by the highest ranked GLMPA: Both (slope, BPI25, wmax and botstrmax), Hydro (wmax) and Terrain (slope and BPI3). The best Hydro model generated almost identical results to the Both model, revealing the strong dependency of coral distribution on areas of vertical flow (i.e. local downwelling), as well as the interplay between local current regimes and morphological features. Interestingly, maps including wmax and botstrmax variables showed generally higher probability values in the Logachev mound province (Fig. 4). Inspection of summary statistics of the hydrodynamic variables revealed significantly stronger current flows in the Logachev mound province compared to the other two areas (see Mohn et al., 2014). As data pooled from all study areas were used for model calibration, the generally lower flow speeds in the Arc and

Belgica mound provinces resulted in lower probability scores in these areas. 3.3. Proportion data vs. presence–absence data Based on the percentage of deviance explained, GLMs and GLMMs using proportion data performed on average slightly better (0.4370.08) than models using presence–absence data (0.4070.04). Consistently lower standard errors (SE, Table 3) suggest a decrease in coefficient uncertainty and thus an increase in model reliability. Model averaged parameter estimates (Table 3) and uni-variate response curves (Fig. 3) were comparable between response variables; however, presence–absence curves were slightly steeper than the equivalent proportion curves. This was to be expected, as even small proportions (e.g. 1 presence record out of 10 observations: 10%) would be coded as 1 (100%) in the presence–absence response. Preliminary analyses using more complex, data-driven modelling techniques (generalized additive models, GAMs) resulted in even more pronounced discrepancies between the two response curves. 3.4. GLMMs vs. GLMs Uni-variate response curves created with the GLMMs (not presented) were significantly flatter and strongly skewed towards the x axis. Model averaged parameter coefficients differed considerably from the ones generated by GLM (Table 3). GLMMs essentially compute individual responses for each within-group dataset (e.g. 6 response curves for 6 individual transects), and then generate a model explaining the maximum variance between these individual response curves. Therefore, while GLMs used individual observations (n¼ 900) as independent records, the sample size in GLMMs was based on the number of groups (e.g. transects). In the present study, a relatively small number of groups were available (six per study area). Model specifications using more than 3 or 4 environmental predictors frequently resulted in numerical problems and some models failed to converge. Further, three out of six transects

78

A.M. Rengstorf et al. / Deep-Sea Research I 93 (2014) 72–82

Table 3 Model-averaged parameter estimates, standard errors (SE) and relative parameter importance for GLMs and GLMMs using presence–absence and proportion data. The parameters used by the best Both (B), Hydro (H) and Terrain (T) models are indicated as superscripts beside the variables, and their respective percentage deviance explained is given. Presence–absence data

GLM

Variable

Variable importance

Estimate

Standard error

Variable

Variable importance

Estimate

Standard error

Intercept slopeB, T bpi3T bpi25B wmaxB, H botstrmaxB tmpmean

1.00 0.97 0.43 0.98 0.94 0.99 0.29 Both 0.46

 4.66 0.10 0.05 0.03 29.67 19.11 0.04 Hydro 0.39

1.51 0.03 0.04 0.01 11.83 5.81 0.24 Terrain 0.36

Intercept slopeB, T bpi3B, T bpi25B, T wmaxB, H botstrmaxB, tmpmeanH

1.00 1.00 0.54 1.00 1.00 1.00 0.39 Both 0.53

 4.65 0.06  0.02 0.05 13.28 20.37  0.09 Hydro 0.34

0.75 0.01 0.01 0.00 3.46 2.23 0.08 Terrain 0.43

1.00 0.68 0.66 0.85 1.00 0.44 0.58 Both 0.44

5.73 0.09 0.12 0.04 83.81  27.75  2.02 Hydro 0.39

13.78 0.05 0.07 0.02 23.16 26.97 1.31 Terrain 0.36

1.00 1.00 1.00 1.00 0.37 1.00 Both 0.52

13.17 0.08 0.09 0.03 47.81  7.15  2.32 Hydro 0.34

5.46 0.01 0.02 0.00 5.44 7.37 0.59 Terrain 0.43

% Dev. expl. GLMM

Proportion data

Intercept slopeB, T bpi3B, T bpi25B, T wmaxB, H botstrmaxH tmpmean % Dev. expl.

H

% Dev. expl. Intercept slopeB, T bpi3B, T bpi25B, T wmaxB, H botstrmaxH tmpmeanB, H % Dev. expl.

Fig. 4. Maps showing probability of coral presence in the three study areas (Logachev, Arc and Belgica mound provinces, from left to right) as generated by the highest ranked models using presence–absence data: Both (variables used: slope, BPI25, vertical flow and bottom stress), Hydro (vertical flow) and Terrain (slope and BPI3). Probability of presence values ranged from low (0, blue) to high (1, red). Maximum and minimum probability values are indicated in each map.

were conducted in off-mound areas and hence included few if any coral presence records. These within-group response curves resembled straight horizontal lines of zero probability of presence, severely skewing the averaged GLMM curves towards lower values. Preliminary GLMMs, including only mound-transects, produced responses more comparable to the GLMs, but here calculations were hampered due to data reduction by 50%.

3.5. Model transferability Fig. 5 shows coral distribution maps generated by the GLMPA, calibrated within one study area only and then projected onto the remaining study areas. Predictions of the GLMPR (not shown) were comparable but with slightly lower probability scores. Model evaluation indices are presented in Table 4. Model transferability

A.M. Rengstorf et al. / Deep-Sea Research I 93 (2014) 72–82

79

Fig. 5. Maps displaying spatial transferability of the highest ranked model (GLMPA using the variables slope, BPI25, wmax and botstrmax). The maps show probability of presence generated by models trained using data from one study area (Logachev, Arc, and Belgica data, from top to bottom), and then projected onto the remaining study areas (Logachev, Arc and Belgica mound provinces, from left to right). The maps highlighted by a black border (top left, middle, and bottom right) show maps where models were trained and applied in the same area. The six remaining maps display transferability of models from one study area into another. Probability of presence values ranged from low (0, blue) to high (1, red). Maximum and minimum probability values are indicated in each map.

Table 4 Model performance indices calculated for the locally trained and transferred models, based on the best Both model (slope, BPI25, vertical flow and bottom stress) as analyzed by the GLM model using presence–absence data. Underlined values indicate those models that were trained and projected in the same area. The “PresenceAbsence” package’s default threshold value of 0.5 was used to calculate sensitivity and specificity values; the AUC score is threshold independent (Freeman and Moisen, 2008). Calibration data

Projection area

Sensitivity

Specificity

AUC

Logachev data

Logachev Arc Belgica Logachev Arc Belgica Logachev Arc Belgica

0.96 0.12 0.24 0.97 0.8 0.36 1 0.04 0.66

0.97 1 0.93 0.95 0.99 0.88 0.83 0.96 0.92

1 0.98 0.73 0.99 0.99 0.71 0.95 0.64 0.9

Arc data

Belgica data

was limited by the fact that hydrodynamic conditions varied considerably between study areas. Models performed well in areas where they had been calibrated (bold AUC values 40.9, Table 4), but poorly to moderately well in the other study areas. For example, when calibrating the SDMs using data from the Logachev area, where current speeds are generally higher, projection onto the Arc and Belgica mound provinces, where current speeds are relatively low, resulted in severe under-prediction of coral presence. This is evident from the distribution maps, as well as from the low model sensitivity values of 0.12 and 0.24 for Arc and Belgica provinces, respectively (Table 4). Accordingly, the model

trained with data from the Belgica mound province severely overpredicted coral presence in the Logachev area (Fig. 5). The reduction in model calibration data probably further decreased model performance. We hence recommend building more general models, using data pooled from several regions, when the aim is to predict species’ distributions in independent regions. We acknowledge that the reported model evaluation indices are highly influenced by data availability and distribution, and are therefore to be interpreted with caution.

4. Discussion While the use of depth-related proxy variables for hydrodynamic regimes is common practice in benthic SDM, more proximal and functionally relevant predictors (e.g. vertical current flow and bottom shear stress) are highly desirable for ecological explanation and insight. The present study revealed a strong relationship between CWC distribution and areas of local down-welling, highlighting the dependence of suspending-feeding corals on an effective transport mechanism delivering food from the euphotic zone to CWC depths (Mohn et al., 2014). CWCs were further associated with areas of locally enhanced currents and bottom shear stress, most likely resulting from the presence of resonantly amplified and rectified diurnal tidal currents (Mohn et al., 2014). The presence of internal waves or bottom intensified diurnal waves has previously been observed to shape coral distribution patterns along the Irish continental margin (White, 2007).

80

A.M. Rengstorf et al. / Deep-Sea Research I 93 (2014) 72–82

It is worth noting that the present study, generating static predictions of present-day conditions, does not yet exploit the full potential of these 3D hydrodynamic models. Coupled with virtual particle tracer studies, hydrodynamic models could provide valuable information on deep-sea dynamic processes that influence spawning patterns and larval dispersal (Levin, 2006). Once the relationships between hydrodynamic processes and species distributions are sufficiently established, hind-casts of hydrodynamic settings can be linked to fossil species records (e.g. dated remains of corals, bivalve or brachiopods) to estimate past species distributions and to assess niche stability over geological time-scales (Stigall, 2009). Model forecasts can provide insight into the effects of regional climate change and possibly identify new recruitment areas for vulnerable marine species (Tittensor et al., 2010). Future efforts could also concentrate on the development of processbased models combining spatial predictions from SDMs with 3D physical and biological mechanisms to resolve the dynamical controls of bentho-pelagic coupling (Soetart, pers. comm.). Given the longevity of vulnerable marine ecosystems such as CWC reefs, it might have been ecologically more relevant to integrate data derived from a hydrodynamic model generated over multiple years or decades, rather than using the 6-hourly intervals during a month-long sampling cruise. While global hydrodynamic data are available over longer time periods (e.g. Carton et al., 2005), long term simulations of high-resolution hydrodynamic models are limited by computing time and computer power. While seasonal differences certainly affect the strength of the signals found within the oceanographic key variables, their relative spatial distribution is likely to be maintained, as near bottom tidal currents are relatively stable over long time periods. The present study is the first to explore the application of proportion data to estimate the distribution of deep-sea VMEs. Few deep-sea benthic investigations relate environmental variables to quantitative species data in the form of abundance (Dolan et al., 2008; Howell et al., 2010), density (Orejas et al., 2009) or percentage coverage (Guinan et al., 2009b; Vertino et al., 2010). The latter is usually derived during time-consuming post-processing, which involves calibrating images and overlaying sampling quadrates (Vertino et al., 2010) or random sampling points (Guinan et al., 2009b). While proportion data are commonly used in, for example, medical studies (Crawley, 2007), the approach has not yet been applied to SDM to the best of our knowledge. The results in the present study suggest that the use of proportion models increases model fit and reliability. By taking into account sampling effort per grid cell, the effects of false absences, generated by extrapolating video-based observations onto entire grid cells, is greatly reduced. With converging trends in autonomous surveying (Williams et al., 2010), automated image recognition techniques (Purser et al., 2009; Schoening et al., 2012), and increasingly available high definition video data, the proportion modelling approach described in this paper suggests a fruitful avenue of research for analysing large quantities of video data in a detailed yet time-efficient manner. SDMs are highly dependent on data quality and quantity. In the deep sea, commonly applied survey strategies such as random or random-stratified sampling (Hirzel et al., 2002) are currently difficult to implement. Continuous transect and spatially clustered data require the use of statistical methods that take into account spatial pseudo-replication and autocorrelation. These methods include mixed effect models, generalized estimation equations (GEEs), autocovariate models or spatial eigenvector mapping, amongst others (Dormann et al., 2007). In the present study, the limited amount of data available within each region precluded a meaningful investigation of issues concerning spatial correlation. Accordingly, the GLMM approach tested was severely hampered by the small number of transects available (see Section 3.4). Woodby et al. (2009) compared

traditional GLMs to GEEs in order to extrapolate deep-sea coral distribution from 27 transects distributed among 9 study sites. Even with a dataset larger than in the present study, the authors found that not enough data clusters (i.e. transects) were available to generate reliable GEEs. The compromise between the demands of statistical sampling and the logistics of continuous data acquisition suggests that the most effective sampling will be achieved with a large number of short, randomly distributed transects. Future mapping activities will probably use a combination of AUVs, ROVs and drop-frame cameras in order to maximize the data information content and spatial coverage. ROVs are particularly useful for the in situ sampling of voucher specimen, often required for identification and genetic bar-coding of poorly known deep-sea species. Multibeam backscatter data, commonly used for the discrimination of seabed substrate and habitat mapping in shallow waters (Brown et al., 2011), would significantly improve deep-sea SDMs (Davies and Guinotte, 2011; Woodby et al., 2009). In the high seas, ship-borne multibeam echosounder systems (MBES) result in large seafloor footprints and a reduced ability to discriminate smallscale features, especially in poor weather conditions (Anderson et al., 2008). While some studies have suggested the possibility of using shipborne MBES for distinguishing substrate in water depths 4200 m (e.g. Howell et al., 2011), issues of spatial resolution and data quality mean that MBES backscatter data will be acquired increasingly using ROVs or AUVs that fly closer to the seabed. With technological and hydrodynamic modelling developments, process-driven habitat characterisation, pioneered for shallow waters, will become increasingly feasible in the deep sea. For example, Kostylev and Hannah (2007) computed characteristic friction velocities and critical shear stress to develop habitat maps based on ecological relevant variables such as seafloor disturbance and scope for growth. Models quantifying disturbance of deep-sea benthic habitats will be valuable for assessing habitat resilience and vulnerability to anthropogenic adverse impacts such as sediment smothering due to deep-sea bottom trawling (Althaus et al., 2009). 5. Conclusions and future perspectives The spatial prediction of deep-sea VMEs requires high-resolution data to ensure precise spatial matching between model components and reliable model outcomes. High-resolution hydrodynamic variables (e.g. current speed, vertical flow, temperature, etc.) significantly improve model fit compared to purely terrain-based models and are ecologically more relevant than commonly used proxy variables. However, care must be taken when hydrodynamic parameter values in the projected regions are outside the range sampled within the calibration area, as model transferability might be compromised. Semi-quantitative proportion data can be easily generated and may decrease model uncertainty by providing more detailed information on species prevalence within grid cells and by placing less weight on unreliable observations. Issues related to spatial pseudo-replication should be considered when using transect-based distribution data. A large number of short transects rather than a small number of long transects should be conducted for efficient survey design. We anticipate future developments in SDM in including larval stages, hind- and forecasting, and integrating aspects of 3D hydrodynamic processes, bentho-pelagic coupling models and ecological modelling to enhance our understanding of environmental processes limiting species distributions.

Acknowledgements The research leading to these results has received funding from the European Community’s Seventh Framework Programme

A.M. Rengstorf et al. / Deep-Sea Research I 93 (2014) 72–82

(FP7/2007–2013) under grant agreement no. 213144 and by the Department of Communications, Energy and Natural Resources under the National Geoscience Programme 2007–2013. The views and recommendations contained in this study reflect the views of the authors and do not necessarily reflect the views and opinions of the Irish Minister for Communications, Energy and Natural Resources. The EC is not liable for any use that may be made of the information contained in this paper. The authors wish to thank the Geological Survey of Ireland for access to the Irish National Seabed Survey data. We also thank two anonymous reviewers for comments that have enhanced the paper.

Appendix A. Supporting information Supplementary data associated with this article can be found in the online version at http://dx.doi.org/10.1016/j.dsr.2014.07.007. References Althaus, F., Williams, A., Schlacher, T., Kloser, R., Green, M.A., Barker, B.A., Bax, N.J., Brodie, P., Schlacher-Hoenlinger, M.A., 2009. Impacts of bottom trawling on deep-coral ecosystems of seamounts are long-lasting. Mar. Ecol. Prog. Ser. 397, 279–294. Anderson, J.T., Van Holliday, D., Kloser, R., Reid, D.G., Simard, Y., 2008. Acoustic seabed classification: current practice and future directions. ICES J. Mar. Sci. 65 (6), 1004–1011. Armstrong, C.W., Foley, N.S., Tinch, R., van den Hove, S., 2012. Services from the deep: steps towards valuation of deep sea goods and services. Ecosyst. Serv. 2, 2–13. Bolker, B.M., Brooks, M.E., Clark, C.J., Geange, S.W., Poulsen, J.R., Stevens, M.H.H., White, J.-S.S., 2009. Generalized linear mixed models: a practical guide for ecology and evolution. Trends Ecol. Evol. 24 (3), 127–135. Brown, C.J., Smith, S.J., Lawton, P., Anderson, J.T., 2011. Benthic habitat mapping: a review of progress towards improved understanding of the spatial ecology of the seafloor using acoustic techniques. Estuarine Coastal Shelf Sci. 92 (3), 502–520. Burnham, K.P., Anderson, D.R., 1998. Model Selection and Multimodel Inference: A Practical Information-Theoretic Approach. Springer-Verlag, New York, NY. Carroll, C., 2010. Role of climatic niche models in focal-species-based conservation planning: assessing potential effects of climate change on Northern Spotted Owl in the Pacific Northwest, USA. Biol. Conserv. 143 (6), 1432–1437. Carton, J.A., Giese, B.S., Grodsky, S.A., 2005. Sea level rise and the warming of the oceans in the Simple Ocean Data Assimilation (SODA) ocean reanalysis. J. Geophys. Res. 110, 1955–1959. Crawley, M., 2007. The R Book. Wiley. Davies, A.J., Roberts, J.M., Hall-Spencer, J., 2007. Preserving deep-sea natural heritage: emerging issues in offshore conservation and management. Biol. Conserv. 138, 299–312. Davies, A.J., Guinotte, J.M., 2011. Global habitat suitability for framework-forming cold-water corals. PLoS One 6 (4), e18483. Davies, A.J., Wisshak, M., Orr, J.C., Roberts, M.J., 2008. Predicting suitable habitat for the cold-water coral Lophelia pertusa (Scleractinia). Deep Sea Res. Part I 55 (8), 1048–1062. Dolan, M.F.J., Grehan, A.J., Guinan, J.C., Brown, C., 2008. Modelling the local distribution of cold-water corals in relation to bathymetric variables: adding spatial context to deep-sea video data. Deep Sea Res. Part I 55 (11), 1564–1579. Dormann, C.F., McPherson, J.M., Araùjo, M.B., Bivand, R., Bolliger, J., Carl, G., Davies, R.G., Hirzel, A., Jetz, W., Kissling, D., Kuehn, I., Ohlemueller, R., Peres-Neto, P.R., Reineking, B., Schroeder, Boris, Schurr, F.M., Wilson, R., 2007. Methods to account for spatial autocorrelation in the analysis of species distributional data: a review. Ecography 30, 609–628. Dunn, D.C., Halpin, P.N., 2009. Rugosity-based regional modeling of hard-bottom habitat. Mar. Ecol. Prog. Ser. 377, 1–11. Elith, J., Leathwick, J.R., 2009. Species distribution models: ecological explanation and prediction across space and time. Annu. Rev. Ecol. Evol. Syst. 40 (1), 677–697. Fielding, A.H., Bell, J.F., 1997. A review of methods for the assessment of prediction errors in conservation presence/absence models. Environ. Conserv. 24, 38–49. Foubert, A., Beck, T., Wheeler, A.J., Opderbecke, J., Grehan, A., Klages, M., Thiede, J., Henriet, J.-P., 2005. New view of the Belgica Mounds, Porcupine Seabight, NE Atlantic: preliminary results from the Polarstern ARK-XIX/3a ROV cruise. In: Friewald, A., Roberts, J.M. (Eds.), Cold-water Corals and Ecosystems. SpringerVerlag, Berlin, Heidelberg, pp. 403–415 (Polarstern ARK-XIX/3a Shipboard Party). Frederiksen, R., Jensen, A., Westerberg, H., 1992. The distribution of the scleractinian coral Lophelia pertusa around the Faroe Islands and the relation to internal tidal mixing. Sarsia 77, 157–171. Freeman, E.A., Moisen, G.G., 2008. PresenceAbsence: an R package for presence absence analysis. J. Stat. Softw. 23, 31.

81

Freiwald, A., Fosså, J.H., Grehan, A., Koslow, T., Roberts, J.M., 2004. Cold-water Coral Reefs. UNEP-WCMC, Cambridge, UK. Genin, A., Dayton, P.K., Lonsdale, P.F., Spiess, F.N., 1986. Corals on seamount peaks provide evidence of current acceleration over deep-sea topography. Nature 322 (6074), 59–61. Grehan, A., Shipboard Scientific Party, 2009. A Multidisciplinary Investigation of Ecosystem Hotspots and Their Importance as Fish Habitat Along the Irish and Biscay Continental Margins. CE0908 Cruise Report. Earth and Ocean Sciences, School of Natural Sciences, NUI, Galway. Guinan, J., Brown, C., Dolan, M.F.J., Grehan, A.J., 2009a. Ecological niche modelling of the distribution of cold-water coral habitat using underwater remote sensing data. Ecol. Inf. 4 (2), 83–92. Guinan, J., Grehan, A., Wilson, M.F.J., Brown, C., 2009b. Quantifying relationships between video observations of cold-water coral and seafloor terrain in Rockall Trough, west of Ireland. Mar. Ecol. Prog. Ser. 375, 125–138. Guisan, A., Edwards, T.C., Hastie, T., 2002. Generalized linear and generalized additive models in studies of species distributions: setting the scene. Ecol. Modell. 157, 89–100. Guisan, A., Thuiller, W., 2005. Predicting species distribution: offering more than simple habitat models. Ecology Letters 8 (9), 993–1009. Guisan, A., Zimmermann, N.E., 2000. Predictive habitat distribution models in ecology. Ecol. Modell. 135 (2-3), 147–186. Hirzel, A.H., Hausser, J., Chessel, D., Perrin, N., 2002. Ecological-niche factor analysis: how to compute habitat-suitability maps without absence data? Ecology 83, 2027–2036. Holmes, K.W., Van Niel, K.P., Radford, B., Kendrick, G.A., Grove, S.L., 2008. Modelling distribution of marine benthos from hydroacoustics and underwater video. Cont. Shelf Res. 28 (14), 1800–1810. Howell, K.L., Davies, J.S., Narayanaswamy, B.E., 2010. Identifying deep-sea megafaunal epibenthic assemblages for use in habitat mapping and marine protected area network design. J. Mar. Biol. Assoc. U.K. 90 (01), 33–68. Howell, K.L., Holt, R., Endrino, I.P., Stewart, H., 2011. When the species is also a habitat: comparing the predictively modelled distributions of Lophelia pertusa and the reef habitat it forms. Biol. Conserv. 144 (11), 2656–2665. Kenyon, N.H., Akhmetzhanov, A.M., Wheeler, A.J., Van Weering, T.C.E., De Haas, H., Ivanov, M.K., 2003. Giant carbonate mud mounds in the southern Rockall Trough. Mar. Geol. 195, 5–30. Kostylev, V.E., Hannah, C.G., 2007. Process-driven characterization and mapping of seabed habitats. In: Todd, B.J., Greene, H.G. (Eds.), Mapping the Seafloor for Habitat Characterization: Geological Association of Canada, Special Paper 47, pp. 171–184. Kostylev, V.E., Todd, B.J., Fader, G.B.J., Courtney, R.C., Cameron, G.D.M., Pickrill, R.A., 2001. Benthic habitat mapping on the Scotian Shelf based on multibeam bathymetry, surficial geology and sea floor photographs. Mar. Ecol. Prog. Ser. 219, 121–137. Levin, L., 2006. Recent progress in understanding larval dispersal: new directions and digressions. Integr. Comp. Biol. 46 (3), 282–297. Marshall, C.E., 2012. Species Distribution Modelling to Support Marine Conservation Planning (Ph.D. Thesis). University of Plymout, UK. Mienis, F., van Weering, T., de Haas, H., de Stigter, H., Huvenne, V., Wheeler, A., 2006. Carbonate mound development at the SW Rockall Trough margin based on high resolution TOBI and seismic recording. Mar. Geol. 233 (1-4), 1–19. Mohn, C., Rengstorf, A., White, M., Duineveld, G., Mienis, F., Soetaert, K., Grehan, A., 2014. Linking benthic hydrodynamics and cold-water coral occurrences: a high-resolution model study at three cold-water coral provinces in the NE Atlantic. Prog. Oceanogr. 122, 92–104. O’Hara, T.D., Tittensor, D.P., 2010. Environmental drivers of ophiuroid species richness on seamounts. Mar. Ecol. 31, 26–38. Olu-Le Roy, K., Caprais, J.-C., Crassous, P., Dejonghe, E., Eardley, D., Freiwald, A., Galeron, J., Grehan, A., Henriet, J.P., Huvenne, V., Lorance, P., Noel, P., Opderbecke, J., Pitout, C., Sibuet, M., Unnithan, V., Vacelet, J., van Weering, T., Wheeler, A., Zibrowius, H., 2002. Caracole Cruise, N/O Atalante and ROV Victor, 1 þ2. Ifremer, Brest. Orejas, C., Gori, A., Lo Iacono, C., Puig, P., Gili, J.M., Dale, M.R.T., 2009. Cold-water corals in the Cap de Creus canyon, northwestern Mediterranean: spatial distribution, density and anthropogenic impact. Mar. Ecol. Prog. Ser. 397, 37–51. Purser, A., Bergmann, M., Lundälv, T., Öntrup, J., Nattkemper, T.W., 2009. Use of machine-learning algorithms for the automated detection of cold-water coral habitats: a pilot study. Mar. Ecol. Prog. Ser. 397, 241–251. Ramirez-Llodra, E., Brandt, A., Danovaro, R., Escobar, E., German, C.R., Levin, L.A., Arbizu, P.M., Menot, L., Buhl-Mortensen, P., Narayanaswamy, B.E., Smith, C.R., Tittensor, D.P., Tyler, P.A., Vanreusel, A., Vecchione, M., 2010. Deep, diverse and definitely different: Unique attributes of the world'slargest ecosystem. Biogeosci. Discuss. 7, 2361–2485. Rengstorf, A., Grehan, A., Yesson, C., Brown, C., 2012. Towards high resolution habitat suitability modelling of vulnerable marine ecosystems in the deep-sea: resolving terrain attribute dependencies. Mar. Geod. 35 (4), 343–361. R Development Core Team, 2008. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria, ISBN 3-900051-07-0, URL: /http://www.R-project.orgS. R Rengstorf, A.M., Yesson, C., Brown, C., Grehan, A.J., 2013. High-resolution habitat suitability modelling can improve conservation of vulnerable marine ecosystems in the deep sea. J. Biogeogr. 40, 1702–1714. Roberts, J.M., Wheeler, A.J., Freiwald, A., Cairns, S., 2009. Cold-Water Corals—The Biology and Geology of Deep-Sea Coral Habitats. Cambridge University Press, New York, NY.

82

A.M. Rengstorf et al. / Deep-Sea Research I 93 (2014) 72–82

Rogers, A.D., Clark, M.R., Hall-Spencer, J.M., Gjerde, K.M. 2008. The Science behind the Guidelines: A Scientific Guide to the FAO Draft International Guidelines (December 2007) for the Management of Deep-Sea Fisheries in the High Seas and Examples of How the Guidelines may be Practically Implemented. IUCN, Switzerland. Ross, S.W., Carlson, M.C.T., Quattrini, A.M., 2012. The utility of museum records for documenting distributions of deep-sea corals off the southeastern United States. Mar. Biol. Res. 2012 (8), 101–114. Schoening, T., Bergmann, M., Ontrup, J.r., Taylor, J., Dannheim, J., Gutt, J., Purser, A., Nattkemper, T.W., 2012. Semi-automated image analysis for the assessment of megafaunal densities at the Arctic deep-sea observatory HAUSGARTEN. PLoS One 7 (6), e38179. Stigall, A.L., 2009. Using ecological niche modelling to evaluate niche stability in deep time. J. Biogeogr. 39 (4), 772–781. Tittensor, D.P., Baco, A.R., Brewin, P.E., Clark, M.R., Consalvey, M., Hall-Spencer, J., Rowden, A.A., Schlacher, T., Stocks, K.I., Rogers, A.D., 2009. Predicting global habitat suitability for stony corals on seamounts. J. Biogeogr. 36, 1111–1128. Tittensor, D.P., Baco, A.R., Hall-Spencer, J.M., Orr, J.C., Rogers, A.D., 2010. Seamounts as refugia from ocean acidification for cold-water stony corals. Mar. Ecol. 31, 212–225. Tyberghein, L., Verbruggen, H., Pauly, K., Troupin, C., Mineur, F., De Clerck, O., 2012. Bio-ORACLE: a global environmental dataset for marine species distribution modelling. Global Ecol. Biogeogr. 21, 272–281. Vertino, A., Savini, A., Rosso, A., Di Geronimo, I., Mastrototaro, F., Sanfilippo, R., Gay, G., Etiope, G., 2010. Benthic habitat characterization and distribution from two representative sites of the deep-water SML Coral Province (Mediterranean). Deep Sea Res. Part II 57 (5-6), 380–396.

Vierod, A.D.T., Guinotte, J.M., Davies, A.J., 2013. Predicting the distribution of vulnerable marine ecosystems in the deep sea using presence-background models. Deep Sea Res. Part II. Weiss, A.D., 2001. Topographic position and landforms analysis. Poster presentation, ESRI User Conference, San Diego, CA. Wheeler, A., Beyer, A., Freiwald, A., de Haas, H., Huvenne, V., Kozachenko, M., Olu-Le Roy, K., Opderbecke, J., 2007. Morphology and environment of coldwater coral carbonate mounds on the NW European margin. Int. J. Earth Sci. 96 (1), 37–56. White, M., Mohn, C., Stigter, H., Mottram, G., 2005. Deep-water coral development as a function of hydrodynamics and surface productivity around the submarine banks of the Rockall Trough, NE Atlantic. Cold-Water Corals Ecosyst., 503–514. White, M., 2007. Benthic dynamics at the carbonate mound regions of the Porcupine Sea Bight continental margin. Int. J. Earth Sci. 96, 1–9. Williams, S.B., Pizarro, O., Webster, J.M., Beaman, R.J., Mahon, I., Johnson-Roberson, M., Bridge, T.C.L., 2010. Autonomous underwater vehicle-assisted surveying of drowned reefs on the shelf edge of the Great Barrier Reef, Australia. J. Field Rob. 27 (5), 675–697. Wilson, M.F.J., O’Connell, B., Brown, C., Guinan, J.C., Grehan, A.J., 2007. Multiscale terrain analysis of multibeam bathymetry data for habitat mapping on the continental slope. Mar. Geod. 30, 3–35. Woodby, D., Carlile, D., Hulbert, L., 2009. Predictive modeling of coral distribution in the Central Aleutian Islands, USA. Mar. Ecol. Prog. Ser. 397, 227–240. Yesson, C., Taylor, M.L., Tittensor, D.P., Davies, A.J., Guinotte, J., Baco, A., Black, J., Hall-Spencer, J.M., Rogers, A.D., 2012. Global habitat suitability of cold-water octocorals. J. Biogeogr. 39 (7), 1278–1292.