Marine Biodiversity

, Volume 39, Issue 4, pp 291–302

Mapping the benthos: spatial patterns of seabed-dwelling megafauna in a Swedish Fjord, as derived from opportunistic video data


    • Department of Marine Ecology, TjärnöUniversity of Gothenburg
  • Per Bergström
    • Department of Marine Ecology, TjärnöUniversity of Gothenburg
  • Tomas Lundälv
    • Sven Lovén Centre for Marine Sciences, TjärnöUniversity of Gothenburg
  • Lisbeth Jonsson
    • Department of Marine Ecology, TjärnöUniversity of Gothenburg
  • Mats Lindegarth
    • Department of Marine Ecology, TjärnöUniversity of Gothenburg
Original Paper

DOI: 10.1007/s12526-009-0028-1

Cite this article as:
Gonzalez-Mirelis, G., Bergström, P., Lundälv, T. et al. Mar Biodiv (2009) 39: 291. doi:10.1007/s12526-009-0028-1


It is widely acknowledged that mapping of benthic diversity is needed to aid in the management and conservation of marine ecosystems, but the choice of scale is contingent upon the patterns of spatial structure inherent to the benthos, which are often unknown. In this paper, spatial autocorrelation analysis is used to detect and describe fine-scale patterns of spatial structure in assemblages of epibenthic megafauna of the seabed below 20 m depth at the Koster fjord/archipelago area (Sweden). Presence/absence of benthic organisms was obtained from video images, which had been collected by means of a remotely operated vehicle. For sample sizes (grain) of <1 m2, and maximum between-sample distance (lag) of 200 m, rank-correlograms revealed the presence of patches in all 12 sites surveyed, and faunal homogeneity (positive spatial autocorrelation) was always detected within distances <20 m, though there was variation across sites in the sizes of patches. These findings were further used to resample the data at coarser grain sizes, to enable exploring of the faunal composition of the patches. We conclude that spatial autocorrelation analysis can greatly improve the design of sampling schemes by ensuring parsimoniousness, and maximizing the chances of detecting patterns. The procedure shown here is especially well suited to carry out subsequent mapping because it can readily discriminate between types of biotopes.


Spatial structureSpatial autocorrelation analysisEpibenthic megafaunaUnderwater videoMapping


One of the issues ecologists are concerned with is elucidating patterns of distribution of species and assemblages, which is done in the interest of science as much as for developing managerial applications. Regrettably, we are forced to draw our conclusions regarding such big questions from data compiled from relatively small numbers of scattered, limited observations of the world. Conducting these observations at the appropriate spatial scales, that is, at the scales at which organisms respond to their environment and interact with one another, is therefore of utmost importance (Andrew and Mapstone 1987; Underwood and Chapman 1996). Not only does sampling with scaling issues in mind boost our chances of detecting the patterns that we set out to investigate, but it also provides the necessary knowledge to design sampling schemes that are time and cost efficient.

This is not at all a new insight: plant ecologists have long been aware of the effects that sampling design specifications can have on the data gathered and on the conclusions reached through analyzing those data (e.g., Greig-Smith 1961; Kershaw 1957). Building from sampling theory, researchers have considered three spatial elements of sampling design (Legendre and Legendre 1998): grain, defined as the size of the elementary sampling unit and expressed as the surface area supporting the measurement (see also Wiens 1989; He et al. 1994); sampling interval, defined as the average distance between neighboring sampling units, also referred to as sampling step, or lag; and extent, or the total area included in the study. In marine benthic environments, marine ecologists are asking the same fundamental questions, while bearing the burden of even more costly gathering of data. Exploring spatial scales of variation in the benthic landscape, in advance of any attempts to disclose patterns of distribution of benthic biota, is, then, all the more important.

In Sweden, an agreement was reached recently to establish the first Marine National Park at the Koster Fjord and archipelago area, off the West Swedish coast (see Fig. 1). The “Koster Sea National Park Project” has prompted much research in recent years so as to lay the basis for management and monitoring plans, as well as to provide baseline data against which to measure the impact of the park. The present paper is a pilot study, in which we report on the spatial patterns of heterogeneity of benthic biota, with the aim of providing the knowledge to design adequate sampling programs through which data can be collected to answer questions concerning the geographic distribution of benthic diversity in the area. Specifically, we address the question: at what spatial scales can heterogeneity be detected in the composition of benthic assemblages? The issue that we tackle more directly is that of the optimal size of the sampling unit, which we will refer to as grain, following the tradition of spatial ecology (O’Niell et al. 1986; Wiens 1989; Turner 1989), from the standpoint of how sampling resolution may influence the classification and subsequent mapping of biotopes.
Fig. 1

The location of the Koster Fjord and Archipelago, as well as the bathymetry. Also shown are the remotely operated vehicle (ROV) survey sites whose data were used in this study (solid circles)

At most scales, heterogeneity is readily detected in the distribution of organisms across space, meaning that the probability of occurrence of the organism under study is not evenly distributed throughout the area. Spatial structure in species’ distributions can be caused by contagious biotic processes (e.g., predators occurring near their prey), the spatial variation of environmental drivers such as habitat, temperature, etc., and historical factors (Segurado et al. 2006; Legendre and Fortin 1989). Depending on which set of processes has the strongest combined effect, and on how biota respond to those, different types of spatial structure may appear. Ecological data are in fact a composite of spatial structures, with trends at macroscales, and patches and gradients at meso- and local scales (Fortin et al. 2002). It is nevertheless possible to find homogeneous distributions as well (Andrew and Mapstone 1987). Random patterns are generally found at local and microscales, or simply below a critical size of grain, where the probability of finding a species becomes evenly distributed throughout the sample. The theory applies equally, by hierarchical extension, to the multivariate case of species assemblages, where the covariance of the multiple variables substitutes the variance in defining heterogeneity (Kenkel et al. 1989).

Spatial structure can be detected and measured by means of spatial autocorrelation analysis. This type of analysis is grounded on the fact that, given a combination of grain, lag, and extent, the value of a variable is not independent from the value of the same variable in nearby samples, thereby potentially biasing the results of standard statistical tests. This is an issue that has been extensively covered in the literature (e.g., Dormann et al. 2007; Segurado et al. 2006), but which is not our concern here. When series of geographically referenced samples are available, the relationship between spatial autocorrelation and distance can be evaluated, and the scales at which patterns emerge revealed. This relationship can, in turn, be used to determine the amount of spatial detail that can safely be ignored. Accounting for this fact gives the investigator the possibility of attaining the most parsimonious sampling scheme to explore an ecological phenomenon.

Few studies have dealt with local scale patchiness of biological assemblages in subtidal benthic environments (e.g., Parry et al. 2003), presumably due to the practical difficulties of gathering adequate data, but it has long been hinted at (Angel and Angel 1967; Butman 1987; Eastwood et al. 2006). Rather, meso-scale patterns dominate in the literature. These have been inferred from grab samplers (Eastwood et al. 2006; Brown et al. 2002,2004; Somerfield and Gage 2000), photographs (Kostylev et al. 2001), beam trawls (Jennings et al. 1999), and visual censuses through SCUBA diving or snorkelling (García-Charton and Pérez-Ruzafa 2001; Galzin and Legendre 1987), and scales considered range from tens of meters to kilometres. Underwater video sampling is well suited for capturing sufficiently small samples, sufficiently close to one another, so as to study small scale patterns of similarity of deep benthic communities. By use of the repository of video footage accrued over the last few years under the auspices of the National Park project, this paper explores fine-scale patterns of spatial structure of epibenthic megafaunal assemblages, including attached fauna, free-living fauna, and demersal fish of the Koster Fjord greater area. We focus on scales from one to hundreds of meters, while the term microscale will apply to one to tens of meters, and local scale to ten to hundreds of meters. The purpose of this paper is twofold: (1) to describe spatial patterns of benthic biota using existing material from the area, and (2) to draw conclusions useful for the optimization of future studies on patterns of distribution and abundance of benthic fauna.


Data collection

The video-material used in this study is labelled “opportunistic” because it was not collected for the purpose of studying spatial patterns of variability of the benthos. Rather, the surveys from which the data were obtained were of a reconnaissance nature, specifically targeted at (1) compiling an overall list of deep benthic species, (2) finding species-rich areas, and (3) capturing good photos and attractive footage to be used in outreach and public awareness activities. This setting has not surprisingly resulted in a number of challenges regarding the use of video images for obtaining quantitative data suitable for statistical analyses. Problems included variable distance to the bottom, variable flying speed, and winding surveying tracks. However, the wealth of data as well as their immediate accessibility outweighed the above inconveniences, and ultimately, the nature of the data proved to be of much value. By virtue of the state-of-the-art technological equipment in place and the computational power available, it was possible to acquire, out of the archived underwater footage, semi-quantitative data on epibenthic fauna that were geographically referenced and highly spatially resolved.

The video data had been collected by means of a video camera mounted on a Sperre Subfighter 7500DC Remotely Operated Vehicle (ROV). The ROV, deployed from R.V. "Lophelia", is fitted with a transponder, which communicates with the vessel’s positioning system and sends back a stream of positional data. Survey tracks can thus be reconstructed in a Geographic Information System through straightforward plotting of the (smoothed) pings.

In the laboratory, video files were viewed and manually chopped into fragments that could be georeferenced (i.e., the timecodes had been recorded) whose field of view, as detected by sight, did not overlap with one another. These units are referred to as “video frames” and they constitute the elementary sampling units of this study. Technically, they were in fact short video sequences, such that the field of view of the video frames making up the sequence overlapped at least in part. This approach ensured that sampling units were as small as possible, and of equal size. Grain size is therefore defined by the swath of the camera as <1 m2. Conveniently, in areas of highly diverse fauna, ROV flying speed was low, occasionally stationing at certain spots. These samples have a real-time duration of up to several minutes, which allowed identification of a great variety of taxa.

The georeferencing of the video frames, i.e., the assignation of geographic coordinates to the samples, was done through linear interpolation with the navigation data, where timecodes were used to calculate the expected distance to the previous ping and the next. In this process, samples were considered dimensionless (i.e., points). Throughout this study, we used only projected units (meters), from UTM zone 32N.

Twelve video surveys were chosen for this paper (Fig. 1), with linear distances ranging between 100 and 250 m, and depths between 20 and 180 m (Table 1). We will refer to these as tracks. This selection was made based on whether the material was available in digital format, and in an effort to obtain data from as broad an area as possible. For each one of a total of 1658 video frames, all discernible organisms were identified to the species level whenever possible, and noted as present. When organisms could not be assigned to a genus, they were noted as “unknown” and were excluded from the present analyses.
Table 1

Summary of remotely operated vehicle (ROV)-surveyed sites



Date (dd/mm/yyyy)a

No. of frames checkedb

Max. dist (m)c

Min. depth (m)d

Max. depth (m)e


O Nordkoster







NO Nordkoster







O Saltholmen







O Ramskär







S Grisbådarne







V Svartskären







S Stenbrottet







SV Syd Hällsö







SV Ursholmen







V Norr Hällsö







O Kalkgrund







V Nordkoster






aDate of ROV survey

bTotal number of video frames along surveyed track suitable for statistical analyses

cMaximum between-frame distance

dMinimum depth registered along surveyed track

eMaximum depth registered along surveyed track

All video-related data, pertaining to both identified taxa and ROV dive information, was stored in a geodatabase. The structure of this geodatabase allowed for a great deal of flexibility in the production of customized datasets. In some of the analyses performed, samples were aggregated at different levels, representing different scales of analysis (increasing grain size). Details are provided at the relevant points.

Data analyses

To test for the presence of spatial structure in the data, a number of Mantel tests were performed (Mantel 1967; Legendre and Fortin 1989; Legendre and Legendre 1998). The non-parametric Mantel test, based on rank-correlations, (hereafter referred to as Mantel test) is a means to correlate the corresponding elements of two distance matrices, in the present case one of biological dissimilarities and another one of geographic distances, to describe the relationship between objects (samples) in relation to their species composition and separation distance (Legendre and Fortin 1989; Somerfield and Gage 2000).

The Mantel statistic used was the rank-based Spearman's correlation coefficient (Rho, ρ) between biological dissimilarity and geographic distance. The null hypothesis of this test states that there is no relationship between biological resemblance and geographic distance between samples, this being the outcome of a lack of spatial structure. Significance was tested through a permutation procedure. Thus, the rows in one distance matrix were randomly shuffled and a new correlation coefficient calculated, iteratively, for 9,999 runs. The original correlation coefficient was checked against the distribution thus generated, and significance assessed by means of a two-tailed test.

To set the context of the local scale analyses, it was firstly determined whether at the regional scale (1–10 km) there was any spatial structure. For this test, species presences and absences in all video frames were aggregated to yield an overall species composition for the whole track. Biological similarity between tracks was computed based on presence/absence-transformed data using the Bray-Curtis measure, and then subtracted from 1 in order to obtain a measure of biological distance. A correlation coefficient was then calculated between dissimilarities, and geographic distances between the center points of the tracks.

For each one of the 12 tracks, an equivalent test was subsequently carried out, where the video frames are now the objects of comparison. Distances between objects included from 0 (adjacent samples) to hundreds of meters. Bray-Curtis (dis)similarity between geolocated video frames was also used here. The distance between samples was calculated in meters, and it was simple Euclidean distance measured in Cartesian space.

Spatial patterns were also quantified by means of multivariate rank-correlograms. This type of plot (hereafter referred to simply as “correlogram”) was originally developed by Sokal and Oden (1978) and Oden and Sokal (1986). It enables the comparison of the average biotic similarity of samples within a preset distance class, with the average similarity of all other pairs of samples, by plotting correlation coefficients, which are now standardized, against distance class. Briefly, correlation coefficients are computed sequentially between a group of cells in the similarity matrix (those corresponding to the samples separated by the distances in the class at hand) and everything else, one class at a time. The purpose is to approximate a continuous function of the autocorrelation with geographic distance, partitioned in intervals. In such plots, significant autocorrelation minima represent the distance between samples that are least similar biotically. Significant autocorrelation maxima that occur in distance classes greater than an autocorrelation minimum indicate the distance between the centres of successive patches of biotic similarity (Parry et al. 2003; Legendre and Fortin 1989). A spatially homogeneous faunal assemblage will be represented by a rank-correlogram in which no Spearman’s ρ values are significantly different from zero, or in other words, by a more or less flat line levelled around the origin axis.

For the construction of the correlograms, classes were made on the basis of numbers of video frames (rather than equal distance intervals), to enable a direct comparison between classes across the entire range of spatial scales. Thus, the number of two-way connections in all classes of all correlograms is approximately 1,000, the number of classes varying greatly from track to track, from as few as 2 classes (KH33) to as many as 13 (KH21). The significance of the Mantel statistic was tested for each distance class using 10,000 permutations with replacement for a two-tailed test, in a way analogous to that used for the overall Mantel tests (see above). Global significance was then assessed by checking that at least one class was significant at the Bonferroni-corrected level, based on the number of distance classes in the relevant correlogram.

Because we were interested in overall patterns, regardless of potential differences in driving processes, the tests were done using data from all taxa, both attached and free-living biota, together. This approach also served to reduce the impact of stochastic variation. Keeping in line with the aims of the study, no distinction was made between substrate types or depth classes.

The most interesting output from a rank-correlogram is insight into the scale at which heterogeneity can be detected, which is given by the intercept of the autocorrelation function with the x-axis. This value of distance, in meters, was used to re-sample the data at coarser grain sizes. Species presences and absences were lumped together for the whole set of contiguous video frames that fit inside a circular area, dubbed “bubble” for simplicity, whose diameter was the distance determined by the autocorrelation intercept (Fig. 2). We wanted to ensure that sharp corners in the tracks would not end up as two consecutive sections, and thus overlapping in the area that they represented, but rather, in one single sample. This method guaranteed the presence of a neighbourhood around the data points, of which the data are assumed to be representative. The samples thus obtained, hitherto referred to as “bubbles”, represent variability at the local scale, and were used to carry out further multivariate analyses.
Fig. 2

The scales of analysis used. The solid line depicts the survey track, as interpolated from the navigation data. Solid circles represent georeferenced video frames. The procedure for aggregating frame-by-frame data into large grain size samples using the “bubble” method is illustrated

The multivariate, local scale analysis remains largely exploratory, mainly because there was variation in the size of the sampling units (the sizes match local patterns of heterogeneity, which vary across sites). A hierarchical cluster analysis combined with a SIMPROF test was performed (see Clarke et al. 2008 for details on how this test works), using group average as clustering method. This final analysis enabled us to describe the patterns of association among taxa which emerge at the scale of analysis chosen. Furthermore, it allowed us to describe the composition of the faunal patches.

The Mantel tests were done by means of a script written for MATLAB. All other analyses were carried out in PRIMER 6, where the RELATE function was used, as well as the built-in suite of multivariate analysis tools.


General observations

A total of 153 taxa (both genus and species levels) were identified from the ROV footage of the seabed below 20 m, including one seemingly undescribed species of Ascidia. On average, 42 taxa were identified per site, with little variability in taxa richness among sites. The most common taxon was Pandalus sp., of which two species exist in the area, the more common P. borealis (Northern Shrimp), and P. montagui (Pink Shrimp). The images did not always allow this distinction, and the two species were collapsed into one genus in our analyses. This taxon was followed by Sabella pavonina (Peacock Worm), and two sponges: Axinella rugosa and Phakellia ventilabrum. Over 40 taxa were encountered only once, among which were Aphrodita aculeata (Sea Mouse), Astropecten irregularis (Sand Star), Myoxocephalus scorpius (Short-spined Sea Scorpion), Thorogobius ephippiatus (Leopard-spotted Goby), Zeugopterus punctatus (Topknot), and Gobius niger (Black Goby).

Spatial structure

Analyses of spatial autocorrelation revealed homogeneity at the regional scale (ρ = 0.09, p value = 0.6). This means that for a grain size of approximately 2 ha (based on the size of a minimum bounding rectangle fit around the filmed track), and lag distances between 700 m and 23 km, there are no gradients or trends in species composition of benthic assemblages at depths below 20 m throughout the area. Nevertheless, these are very few data points for a very large area (approximately 500 km2), so this finding should be taken with much caution.

At the local scale, spatial structure was found in all sites, and this was always highly significant (Table 2). The weakest signal was found at KH6, according to the intensity of the relationship as is inferred from the value of ρ, along with the significance level of the test. KH9 featured the strongest relationship between biotic dissimilarity and distance, indicating strong spatial structure (or positive spatial autocorrelation) of biological assemblages.
Table 2

Spearman’s correlation coefficients between biological dissimilarity and distance between video frames



























***p value < 0.001, **0.01 > p value > 0.001

Spatial patterns

All rank-correlograms are displayed jointly in Fig. 3, where correlation coefficients have been plotted against the mid-point of their corresponding distance class, so that all autocorrelation profiles can be portrayed together along one continuous abscise axis (the vertical scale is the same, since all correlations have been performed on equal sample sizes).
Fig. 3

Composite multivariate rank-correlogram, for equally-sized distance classes. Spearman’s correlation coefficients (ρ) are plotted against the mid-point of their corresponding distance class. Solid triangles on the x-axis represent the intercepts of the interpolated profiles. Values of ρ significantly different from 0 are represented by solid circles, while non-significant ones are represented by blank circles

All but one (KH6) correlograms were globally significant at the α = 0.05 level (and in fact, not just one, but the majority of the values therein where also significant at the α/v level, where v was the number of classes for that correlogram), indicating that spatial assemblage heterogeneity was indeed captured at the studied range of scales. While for KH6, biotic similarity between samples was no higher or lower in any distance class than in any other, all other sites showed highly positive autocorrelation across the short distance intervals (left-most end of the plot), and highly negative autocorrelation in those distance classes immediately following the first group, occasionally reaching the minimum at this point.

There was a high degree of agreement across sites at short to medium distances, where autocorrelation exhibited a regular monotonic decline with distance at similar rates (with the exception of KH9, whose rate of decrease appeared lower). In other words, homogeneity is found at the micro-scale (one to tens of meters) throughout the study area, and a mosaic of patches can be detected at scales larger than those. Furthermore, the interpolated profiles all cross the zero level within a narrow window of distances at an average of 38 m (Fig. 3, Table 3), indicating the presence of patches of similar dimensions in almost all localities.
Table 3

Estimated values for the intercepts of the autocorrelation profiles


Distance (m)

























When moving into larger distance classes (local scale), the various autocorrelation profiles show some differences. In KH8, similarity between video frames increases, until it becomes higher than expected by chance (significant positive autocorrelation) again, beyond 100 m. This indicates the entry into another patch of the same kind as the previous one (“wide wave” sensu Legendre and Fortin 1989). From this it can be inferred that at this site, the centers of patches are closer together than at all other sites. KH32 reaches a second, lower peak at ≈75 m, possibly suggesting a nested pattern of patches. KH21 seems to stabilize beyond 40 m, similarity remaining lower than expected by chance all the way through the classes studied. This is could be the result of an ample transition zone between patches.

All other sites present a similar type of structure where, for the scales measured, the relationship between similarity and distance continues to decline, indicating that we have left the patch, but not entered a new one (“sharp step” sensu Legendre and Fortin 1989), obscuring what the distance between patches might be, or, indeed, the size of the patch we are in (e.g., KH30, KH31, KH33).

The geographic distance where the autocorrelation becomes not significant measures the distance at which samples become structurally independent. Another way of interpreting this distance is as the average “zone of influence” (Legendre and Fortin 1989; Wilkinson and Edds 2001) of a given combination of biophysical processes, as is manifest in the presence of a particular benthic assemblage (i.e. the diameter of the “core” area of the patch). The intercepts are summarized in Table 3, as well as depicted in Fig. 3. The mean zone of influence of local processes across the greater Koster area was 38.6 m (95% CI: 27.5–49.8). If the value for KH9 is removed, the mean becomes 34.4 m (95% CI: 27.6–41.2).

Megafaunal assemblages

The hierarchical cluster analysis was based on standardized frequencies (sum of presences divided by number of video frames contained in each “bubble”). Between 3 and 8 “bubbles” were obtained for each track. Site KH6 was excluded from this analysis, owing to the fact that no patch structure was found there at the measured scales. When the dendrogram was cut off at the (arbitrarily chosen) 20% similarity level, six clusters of >2 members could be identified (see Fig. 4). We interpret these groupings as biological assemblages, and describe them below according to their typical compositions, i.e., species that were encountered most frequently, and species that were very abundant.
Fig. 4

Hierarchical, group average, cluster classification. Samples (“bubble” level) are labeled with the track site name, followed by a numerical indicating the sample number for that site. Gray branches indicate the partitionings where the SIMPROF test was not significant. Clusters that are discussed in the text are also labeled, using numbers from 1 to 6

Assemblage 1 was largely dominated by Pandalus sp., which co-occurred with Sabella pavonina, Parastichopus tremulus, Axinella rugosa, Mesothuria intestinalis, and Pachycerianthus multiplicatus among other less frequent species. The most abundant taxa, occurring in much lower numbers than Pandalus sp., were M. intestinalis, Kophobelemnon stelliferum, and P. tremulus. This cluster also included the only occurrences of the fish Myxine glutinosa.

The most frequent, as well as most abundant, taxa of Assemblage 2 included, in decreasing order of abundance, S. pavonina, Phakellia ventilabrum, Ascidia mentula, A. rugosa, Terebratulina retusa, and Novocrania anomala.

Assemblage 3 featured the following taxa as most frequently encountered: A. mentula, P. ventilabrum, A. rugosa, Porania pulvillus, Caryophyllia smithii, Ctenolabrus rupestris, Asterias rubens, and Labrus mixtus. The most abundant species were C. rupestris, and Alcyonium digitatum.

Assemblage 4 was characterized by the presence of Gadus morhua (Cod) and Suberites virgultosus which were both the most frequent and most abundant, alongside of A. mentula, S. pavonina, N. anomala, Lumpenus lampretaeformis, and Tetilla cranium.

The most abundant, as well as frequently-encountered species of Assemblage 5 were M. intestinalis, V. mirabilis, K. stelliferum, P. pulvillus, Funiculina quadrangularis, and Astropecten irregularis. Present in large numbers were P. pulvillus, and F. quadrangularis. There are only three members of this assemblage, and they are all at the same site (KH32).

Assemblage 6 was distinguished by the following widely occurring taxa: M. intestinalis, G. morhua, A. rubens, Hyas sp., and Pennatula phosphorea. The following species appeared also in large numbers: A. rubens, P. phosphorea, M. intestinalis, Liocarcinus sp., Melanogrammus aeglefinus, and Taurulus bubalis. There are only two members of this assemblage, and they are both at the same site (KH33), which is, additionally, in the neighborhood of KH32.

A SIMPROF test carried out in combination with the clustering revealed that clusters 1, 3, 4, 5, and 6 are significant groups and that therefore, further subdivisions of them would be as good as arbitrary. Conversely, cluster 2 could be further split into three meaningful groups (of three or more members each), which we regard as different “phases” of Assemblage 2. Once completed, this procedure rendered a total of eight classes.

Assemblage 2.1 was defined by the presence of Geodia barretti, N. anomala, Pandalus sp., and S. pavonina, while the latter was dominant. This assemblage was present at five sites, all situated on the edges of the Fjord (at the top of the slope). This was the cluster with most cluster-specific taxa, including Polycarpa pomaria, Ascidia callosa, Bonellia viridis, Clathrina sp., Dendrodoa grossularia, Geryon trispinosus, Guancha sp., Lebbeus polaris, and Macandrevia canium.

Assemblage 2.2 was characterized by the extensive presence of S. pavonina, P. ventilabrum, A. mentula, and T. retusa. The most abundant species in this assemblage were P. ventilabrum, T. retusa, and A. mentula. Lastly, Assemblage 2.3 featured S. pavonina and A. mentula, alongside others less abundant such as N. anomala.

Two samples in the overarching Assemblage 2 become singled out by the SIMPROF test as a significant grouping. We did not include it in our description for containing too few samples; however, it is worth pointing out the high number of unique species present there, such as Eudendrium rameus, Galathea sp., Trisopterus luscus, and Zeugopterus punctatus. Because this assemblage appeared at the site farthest away from the fjord trough (KH8), it appears that more exhaustive sampling might reveal a large scale East-to-West trend in species composition.

Epibenthic megafaunal assemblage patchiness and distribution

The site that showed strongest spatial structure was also that which exhibited the most unique autocorrelation profile: KH9. This points to a region of large, well defined patches, where possibly a suite of unique processes operate. Conversely, in one site no evidence of heterogeneity was found (KH6), leading us to believe that patches around that area, the bottom of the fjord, are larger than the scales studied. It is possible, though no evidence is available, that repeated trawling in that area has had an effect in smoothing out any faunal discontinuities.

The results from the cluster analysis provide ancillary evidence to support the claim that faunal patches are distinguishable at the scale ten to hundreds of meters. Spatial structure resulted always in the presence of several significant assemblages. In addition, the autocorrelation profiles predict rather robustly which way the classification of the “bubble” samples will go. For example, in KH8, where the autocorrelation function drops and then starts to ascend again, the same assemblage can be found near both ends of the track (Assemblage 2), while a different one occupies the middle (Assemblage 1). In KH21, where autocorrelation shows a very slight upwards trend, which we interpreted as a stabilizing profile, only one of the “bubbles” along the track turned out to be of a different kind (Assemblage 2.1), while all others were the same (Assemblage 1). Moreover, the sample classified as 2.1 lay 45 m from the end of the track, as expected from the point where the autocorrelation profile reached its lowest value. In KH2 autocorrelation goes up very quickly after the minimum and then starts to drop again immediately. There, a sort of intrusion of Assemblage 2.1 into Assemblage 1 can be detected. All other sites shared a declining profile, and accordingly, featured a series of different assemblages occurring along the track, none of them appearing more than once.

The most widespread assemblage (Assemblage 1) was a Pandalus-dominated one, involving mostly, though not only, organisms associated with soft sediments (e.g., Pachycerianthus multiplicatus, Parastichopus tremulus, Kophobelemnon stelliferum). It was present at five sites, both at the edges of the fjord, and on the site farthest away from it, emphasizing its ubiquitousness. One other assemblage (5) included organisms that are known to be associated with soft sediments (the sea pens Virgularia mirabilis, K. stelliferum, and Funiculina quadrangularis), and in this case Pandalus was entirely absent. This can be differentiated from the singleton KH5_1, already distinguishable at the 20% similarity level (see Fig. 4), where the dominant taxon was Pandalus sp. co-occurring with Bolocera tuediae, and which seemed much more substrate-specialised, including strictly organisms typical from unconsolidated sediments (e.g., Nephrops norvegicus). Thus, the presence of the shrimps, possibly in combination with depth, emerges as one potentially important factor driving the composition of the benthos across areas of similar substrate type.

Conversely, all clusters belonging to the overarching Assemblage 2 appear to include organisms known to occur only in areas of rock outcrops. Notably, the cluster with the most cluster-specific taxa was found exclusively at sites located at the top of the slope on the edges of the fjord. This underlines the high degree of distinctness of this type of habitat, seemingly constrained to this depth range and substrate type.

Ctenolabrus rupestris (Goldsinny-wrasse) was the dominant species of Assemblage 3, together with Alcyonium digitatum (Dead man’s fingers), and was found only at the most shallow areas. The C. rupestris/A. digitatum assemblage exhibited a high number of species that occurred nowhere else in the study area, though this was presumably because all other sites were located at higher depths.

Lastly, there was one more assemblage (4) that featured mostly hard-substrate organisms, as well as many fish species, both from the cod family, and flat fishes. Assemblage 6, present only at one site (KH33) appears to have a mixture of species both from hard substrate habitats, and soft bottoms, and it also seems to be associated with more fish species than other assemblages.

At the lowest level of similarity, the cluster analysis split up the data into two main classes, one containing clusters 5 and 6, and another one encompassing everything else. These two assemblages are not only similar in their composition, but they are also very close in space, as they occur in neighbouring sites KH32 and KH33. Furthermore, both span a similar range of depths, from 30 to 60 m. Considered together, sites KH32, KH33, KH2 and KH5, located in the same area northeast of the Koster archipelago and bordering the edge of the fjord, on a platform made up of small mounds, contained four assemblages that do not occur anywhere else in the whole of the study area, namely clusters 4, 5, and 6, as well as sample KH5_1 which was classified as a group on its own, including a number of commercial species (Cod, and Norwegian Lobster).


In the Koster Fjord area, the use of underwater videography in combination with Geographic Information Systems has enabled bridging the gap from a wealth of data of a non-spatial nature, to in-depth knowledge of fine-scale spatial patterns of benthic assemblages. The main merits of this paper are related with improving the allocation of time and money resources in benthic research: On one hand, no additional data were needed to carry out this study; on the other, we hope our findings will help increase the efficiency of further data gathering efforts in the Koster Fjord area. Our findings are also secondarily discussed in the context of theoretical considerations of ecological heterogeneity.

The value of pilot studies in ecological research has been pointed out for decades (e.g., Fortin et al. 1989), but a number of questions remain unsolved (Legendre and Legendre 1998), among which is the size of the sampling unit necessary to give an adequate analysis of an ecological phenomenon. This deficiency is more accentuated in the benthic environment (Rhoads and Germano 1982). Conducting a study using an inadequate size of plot, quadrat, or some other type of sampling framework, may result in one of two consequences: the need to sample an area too intensively (i.e., inefficiently), or the risk of overlooking existing patterns (Wiens 1989). Knowledge of the type and intensity of the spatial structure of the variable of interest can help shed light on this question.

When the objective of the research program is related with the recognition of patterns, rather than its counterpart in the arena of statistical inference, the estimation of parameters, it has been suggested that the size of the sampling unit should be smaller than the structures resulting from the unit process (e.g., a patch) (Kenkel et al. 1989; Dungan et al. 2002). Furthermore, if grain size is made to be equal to, or smaller than the smallest available areal unit with homogeneous community composition, we will obtain the greatest spatial resolution possible.

For Wiens (1989), the whole spectrum of scales could be divided into regions over which, “for a particular phenomenon in a particular ecological system, patterns either do not change or change monotonically with changes in scale” (p. 392). These so-called “domains of scale” are in turn separated by “relatively sharp transitions from dominance by one set of factors to dominance by other sets”. In his review paper, he pleaded for the development of methods to incontrovertibly identify those boundaries of scale domains, as he thought the neglect of these to be a key stumbling stone in the process of understanding ecological phenomena. Spatial correlograms, used in the way shown here, are just one such method.

A multivariate spatial correlogram can be regarded as a summary of megafaunal assemblage variation in space, and it can then be used for delimiting geographic regions that are either homogeneous or patchy (Diniz-Filho and Telles 2002), with respect to their megafaunal composition. By representing patterns of similarity across spatial scales, this plot clearly depicts the threshold at which we would move from a homogeneous environment to a heterogeneous, patchy one. Our analyses lead us to conclude that any area of the seafloor approximately 38 m across is likely to be sufficiently homogeneous so as to consider variation within it largely as noise or unnecessary spatial detail, in the regional context.

The lowest value of the autocorrelation coefficient, when plotted against distance, has been used as a measure of the minimum radius of a faunal patch (e.g., Parry et al. 2003). We did not find any correlation coefficients lower than that found for pairs of video frames 60 m apart (ρ = −0.28, p value = 0), in track KH31 (see Fig. 3). However, it appears that, had data at higher distance classes been available for some of the tracks (e.g., KH31, KH33), still lower values of the Mantel statistic would have been found. Here, we propose a view of the landscape as encompassing peaks of similarity within a continuous gradient of faunal composition and have therefore placed the focus on the scales over which there is no evidence of autocorrelation (i.e., the distance classes at which our autocorrelation profiles cross the x-axis). We regard this as the average linear dimension of the “core” area of a patch of seabed-dwelling megafauna. This is also a conservative approach that we felt necessary given the “opportunistic” nature of the data. The mean intercept distance may then be used as an upper limit of optimal grain size, when the ecological question being investigated through the sampling has to do with the identification of patterns, since we have found that these are discernible at scales of tens to hundreds of meters. Our findings indicate that samples of epibenthic megafauna of some regular shape, representing an area of approximately 314 m2 (i.e., that of a circle of 20 m in diameter), up to approximately 5,024 m2 (diameter of 80 m) will efficiently capture relationships between environmental factors and biota, and appropriately describe the system; samples much smaller (at least one order of magnitude difference) will be more affected by stochastic processes, and samples much larger than that are likely to smooth out faunal discontinuities and depict, instead, regional trends.

Reliability of our results is enhanced by the fact that several (12) correlograms were built from our data (which in turn represented the co-variation of some 150 variables), making it possible to supply a probability distribution for this estimate. Furthermore, and perhaps most enlightening, values were tightly distributed around the mean, even more so if the intercept for KH9 (81 m), which strongly suggests a different behavior, is removed. Notwithstanding the width of the distribution, it remains a robust method to determine the size of the influence zone of local ecological processes. The fact that the range was relatively narrow in our system points to further ecological (or possibly geological) questions, related with the factors driving these patterns, deserving of further attention in future studies.

How much smaller than 20 m across should/could grain size be, while still ensuring that micro-scale noise is mostly absent from the data? In this case, the underlying issue is that observations must be large enough so variability among samples is maximal (Kenkel et al. 1989). Indeed, when we performed, in an exploratory phase of the study, an Analysis of Similarities (ANOSIM) on the frame-by-frame data, setting site as factor, the test returned significant, implying that on average there was more variation between sites that within sites (i.e., variation between video frames was very small). Only after grain size was increased, did variation appear at the desired level (i.e., between sites). Our approach was a step in the right direction, but fine-tuning of the sample sizes would be useful, as the “bubble” scale did not completely remove the track (site) signal in the patterns of variation. Thus, the answer to the above question becomes: sample size should be as close to 20 m as time and budgetary constraints allow.

The heterogeneity of benthic communities has been successfully explained by invoking the action of small-scale temporal disturbances; to the extent that they create a mosaic of micro-environments to whose spectra of varying conditions, different sets of species are well adapted. This model was first proposed by Johnson (1970). In our study area, patches of hard substrate are found interspersed with areas of unconsolidated sediment, limiting the presence of the more specialised organisms to the areas where their suitable habitat appears. Similarly, depth imposes varying ambient pressure conditions on the benthic biota. The alignment of the structure of benthic communities with this mosaic of environmental conditions is then readily predicted by the theory. Most striking is the finding that the inclusion of mobile species, such as fish and shrimps, did not deter spatial structure from appearing. Further studies are needed to unravel the relative importance of the different factors driving this pattern, among which must be “the heterogeneity of the physical environment, the vagaries of larval recruitment, and biological interactions” (Johnson 1970).

(Re)sampling the benthic environment at the scales stemming from the spatial autocorrelation analysis has enabled us to define a meaningful level to describe patterns of association between species. Interspecific associations are known to vary greatly as grain size is changed, and as a result species classifications and ordinations are also prone to change (Kenkel et al. 1989). A procedure that attempts to maximize between-sample variation, while minimizing the variation within the sample, is especially well suited to carry out subsequent mapping because it can readily discriminate between types of biotopes, taking into account that properties such as species composition, diversity, and environmental factors are gradational in space, and that however one defines classes, there will be variation among classes and within them (Levin 1992). Because patches displayed by maps can only be interpreted with respect to the scale of the sampling program, it is vital that this is in accord with the scale of processes shaping the structure of the landscape (Legendre and Legendre 1998). The approach proposed here ensures simultaneously sampling efficiency, and maximal mapping resolution.

It is worth driving home the point that “scale should not be considered an intrinsic property of the ecosystem, but rather, a methodological decision we must make as observers, keeping in mind that most ecological processes are multiscaled in nature” (García-Charton and Pérez-Ruzafa 2001, p. 918). In other words, there is no correct scale for describing a system, but the description of variability is contingent upon the window through which the system is viewed. In a homogeneous environment, spatial variability will be a function of grain size, in such way that as grain size is increased variability decays, and the rate at which this will happen is in turn determined by the way spatial autocorrelation falls off with distance (Levin 1992). In this paper, we have presented a robust method upon which to base such a fundamental decision as the size of the window through which to look at a system, justifiably balancing the amount of simplification permissible and the amount of aggregation needed, and thereby determining the level of detail necessary to understand and further explore patterns.


This study was financially supported by Naturvårdsverket’s project Korterhavet, and FORMAS through contract 217-2006-357 to M. Lindegarth. Many thanks to Per Nilsson for his useful comments, and to Andy Hennell for writing and re-writing many MATLAB scripts.

Copyright information

© Senckenberg, Gesellschaft für Naturforschung and Springer 2009