Deep-reef fish assemblages of the Great Barrier Reef shelf-break (Australia)

Tropical mesophotic and sub-mesophotic fish ecology is poorly understood despite increasing vulnerability of deeper fish assemblages. Worldwide there is greater fishing pressure on continental shelf-breaks and the effects of disturbances on deeper fish species have not yet been assessed. Difficult to access, deeper reefs host undocumented fish diversity and abundance. Baited Remote Underwater Video Stations (BRUVS) with lights were used to sample deeper habitats (54–260 m), in the Great Barrier Reef (GBR), Australia. Here we describe fish biodiversity, relative abundance and richness, assessing the prediction that depth would drive assemblage structure in the GBR. Distinct groups of fishes were found with depth whilst overall richness and abundance decreased steeply between 100 and 260 m. Commercially-valuable Lutjanidae species from Pristipomoides and Etelis genera, were absent from shallower depths. Few fish species overlapped between adjacent depth strata, indicating unique assemblages with depth. We also detected new location records and potential new species records. The high biodiversity of fish found in shelf-break environments is poorly appreciated and depth is a strong predictor of assemblage composition. This may pose a challenge for managers of commercial fisheries as distinct depth ranges of taxa may translate to more readily targeted habitats, and therefore, an inherent vulnerability to exploitation.

A number of small-bodied fishes were recorded and are likely an underestimate of true abundance and richness. Both Terelabrus rubriovittatus and Cirrhilabrus roseafascia appeared in a large proportion (17%) of the sites. Other frequently sighted smaller fish include small Bodianus species (25% of sites) and Pentapodus species (19%).
Species richness and abundance with depth. Strong depth-related patterns of relative species richness (number of species per 60 minutes of video) and total fish abundance (sum of MaxN of all species per deployment per 60 minutes of video) were detected and these differences were significant according to ANOVA (Table 2). There was no interaction between depth and site (@ > p = 0.25) and therefore the interaction was pooled into the factor depth. Species richness and abundance generally decreased from shallow to deep although patterns varied by reef (Fig. 1). Comparing Shallow (50-115 m), Mid (128-160 m) and Deep (179-260 m) fish assemblage groups for species richness (t-tests), Shallow-Mid (p = 0.08, NS) and Mid-Deep (p = 0.06, NS) were not significantly different groups, but Shallow-Deep was (p = 0.02*). Tukey's HSD highlighted the same differences in overall species richness between the depth groups: Shallow-Mid (p = 0. 21 145 ) and shapefiles provided by the Great Barrier Reef Marine Park Authority (http://www.gbrmpa.gov.au/resources-and-publications/spatial-datainformation-services 146 ). a similar pattern, with non-significant differences Shallow-Mid (p = 0.47, NS) and Mid-Deep (p = 0.18, NS), and Shallow-Deep was a significant change (p = 0.004*) in pairwise t-tests. Post-hoc Tukey's HSD Shallow-Mid (p = 0.33, NS), Mid-Deep (p = 0.14, NS) and Shallow-Deep (p = 0.004*). Variation of relative species abundance within depth strata was high, as indicated by standard error (SE) of 27-63% of the mean abundance per depth (Fig. 3). There was also variation in relative species richness within depths, SEs 19-49% mean richness. For both richness and abundance there was a general decrease in the variation between sites from shallow to deep (Fig. 3). However, the variation within strata was not great enough to obscure strong depth-related patterns. The decline in relative species abundance was mirrored in some families, with carangids, labrids and lethrinids decreasing in abundance with depth ( Fig. 4). Lutjanidae exhibited depth-related zonation between species, with species Lutjanus bohar and L. sebae found at shallower depths and species from Pristipomoides and Etelis genera only in deeper depths. Lethrinidae species Gymnocranius euanus, G. grandoculis and Wattsia mossambica occurred at depths down to 150-160 m, other lethrinid species occurred in 128 m or shallower. Some fish species were only present at depths greater than 100 m (i.e. Pristipomoides aureofasciatus, Wattsia mossambica, Lipocheilus carnolabrum, Paracaesio kusakarii; Table 1).
There was high species variation within depth strata and a number of single-species occurrences (i.e. species only recorded at one site). Fifty-eight species identified were only present in one site, resulting in high among-site diversity. Of single species occurrences, MaxN (the maximum number of a species within a single video frame) ranged from 1-85 individuals.
There were great differences in group membership by depth. However, in some cases there was species overlap in group memberships with depth (Table 3). Indicator species analysis of four pre-defined depth groups and multilevel pattern analysis attributed 130 species to a group or groups based on transformed species abundance. Twenty-three species were selected as having significant differences with depth: 13 were assigned to unique groups and ten species were assigned to two groups. No species were assigned to more than two groups. The upper mesophotic group (54-65 m) had a total of 36 unique species, of which seven were significantly attributed to only that depth strata (p < 0.05). The middle mesophotic group (85-115 m) was assigned 30 species  with three significant. The lower mesophotic (128-160 m) had 18 species assigned, two were significant. The sub-mesophotic group (179-260 m) was assigned 13 species, only one was significant. There was a greater shared assemblage between the upper and middle mesophotic (11 species total), then between the upper and lower or the upper and sub-mesophotic groups. Middle and lower-mesophotic shared 11 species; the lower mesophotic and sub-mesophotic sites shared six species. The genus Parapercis (Family Pinguipedidae) was unusual in that it may be a depth-generalist genus, found in all three mesophotic groups (0.462, p = 0.765). Further, the highly mobile Gymnosarda unicolor (Family Scombridae) was found throughout the deepest groups (0.622, p = 0.363). Presence-absence data revealed almost identical results, out of 130 species 24 were selected: 12 were assigned to a unique group, 12 assigned to pairs of groups.

Discussion
We found strong differences in fish assemblages with depth with high variability among reefs and sites within reefs. Further, we found distinct assemblages of fishes in mesophotic and sub-mesophotic habitats of the GBR, and these contrasted greatly with those of shallower shelf-habitats (e.g. soft bottom 20-90 m) 37 75 . In the tropics, where the food security of many countries is uncertain, deeper reefs may be next in-line for greater fishing pressure. Many tropical coastlines that have limited shallower fishing areas are targeting deeper fisheries 76 . This is concerning as deeper environments are thought to be vulnerable 9, 76, 77 and fish assemblages are poorly described 78,79 , which may compound the problem. In general, deeper fish assemblages are thought to be diverse, valuable and vulnerable 80 . Since many of these species only occur at deeper depths, it is critical to consider these depth zones as distinct. Bycatch is one of the immeasurable impacts of fishing, therefore, it is important to inventory the biodiversity and value we may lose when we target deeper fisheries. High single-species occurrences can indicate the relative rarity of the fish taxa, but this can only be answered with future sampling and greater spatial replication. It is imperative, therefore, to obtain thorough baseline information on deeper tropical ecosystems before these species and habitats are compromised.
Some of the key indicator species per depth strata were commercially important species. Deep Lutjanidae (snappers from the genera Aphareus, Etelis and Pristipomoides), serranids, carangids and sharks are among the "largely unexplored fauna" of the Townsville area and continental slope 81 , and important for "regional food futures" 81 . Australia shares fauna with the south-western Pacific islands and the larger Indo-Pacific region 30 . As human populations increase across Australia and Indo-Pacific islands nations, pressure will be added to fish stocks throughout the region and sustainable fisheries management will increasingly become a major international political issue [81][82][83] .
In many Pacific nations, there are long-standing and emerging deep bottomfish fisheries and there is growing concern that these data-limited fisheries are vulnerable to the effects of overfishing 77,84,85 . In Hawaii, deep-reef lutjanids, serranids and carangids form the second largest fishery behind the tuna fishery 48 . For the majority of these fishes, biological information is lacking, but limited life history information demonstrate overall low production (see review 86 ). "Essential Fish Habitat" has been set aside to reduce the impacts from fishing in Hawaiian waters 74 and in other countries where these species are targeted similar precautionary measures should be made.
In Australia, deep-reef fishes are targeted by multiple methods along an extensive tropical coastline spanning Queensland, Western Australia and the Northern Territory. In the Northern Territory and Western Australia, mixed gear is used to target Pristipomoides species, primarily Pristipomoides multidens 87,88 , however, often multiple species are marketed under the same common name "Goldband snapper" 89 . In Western Australian waters deepwater demersal trawl gear is also used to target deep-reef fishes 90 . Fishing methods which target >50 species in ~200 m depths unfortunately catch many species as bycatch. In Queensland, while fishing pressure in deeper habitats of the GBR is comparatively lower than in shallow waters, more comprehensive information on deeper based on the Australian Faunal Directory 58 and California Academy of Sciences' Catalog of Fishes 59 63 were consulted for reported depth range. Where differences in these references occurred, the maximum depth range is reported. Climate and known distribution information from FishBase. New record information was compared to reported data from FishBase, Fishes of Australia and Atlas of Living Australia 64 databases and cross-referenced with John Pogonoski (CSIRO).
habitats will help to extend conservation strategies for the GBR World Heritage Area 35,91 and the adjacent Coral Sea 79, 81 to incorporate deeper habitats.
Variation in fish assemblages was strongly correlated with depth and a combination of biotic and abiotic environmental conditions may contribute to this pattern. The thermocline and changing temperature and productivity with depth, may correlate with food for planktivores and piscivores 92,93 . Position of the thermocline is probably a key factor driving the distribution of fishes 5,94 . Our CTD data indicate temperatures rapidly decline below 100 m to 150-200 m, and similar profiles have been previously recorded in this area 95 ; a steeper change than recorded in other tropical mesophotic regions 3,5,94 . This depth corresponds with a transition from the lower mesophotic assemblage to the sub-mesophotic fishes. Variation in physical properties (i.e. nutrients, light, oxygen and temperature) of the water column along with position and intensity of the thermocline influence species abundances in shallow tropical waters 96,97 and this appears to apply in deep shelf-waters 98 . Competition is also a powerful influence on species richness and abundance in shallower waters [99][100][101] and more research on mesophotic competitive interactions is needed.
We found strong patterns of fish abundance with depth, but there was also some variation among reefs that may reflect depth-related patterns of habitat structural complexity [102][103][104] . Decreases or changes in fish diversity within depth strata may be linked to differences in available habitat similar to shallow water environments 36,[105][106][107][108] . Environmental drivers, such as currents and thermal stratification, will affect physical characteristics of the environment (i.e. temperature, sedimentation and food availability), which influence abundance and species diversity 109 . These abiotic factors affect the benthic community (the biotic structures, e.g. hard coral), which combined with the geomorphology, constitutes the habitat available to fishes 110 . Our results indicate inter-reefal habitats had lower relative species richness than those neighbouring reefs, suggesting the importance of the habitat type on diversity. Habitat quality may also explain some variation in relative species richness and abundance among reefs sampled.
Of the information necessary for conservation strategies, worldwide current species inventories and distributions are incomplete 111 . Further, data-poor locations inhibit the ability to monitor and record range extensions and distributional records. Analogous to the tropicalization of temperate waters 112,113 , shallower species may extend their range and begin to inhabit deeper depths 114 . There is little information on how thermal tolerances may change fish distributions or behavior, such as changing spawning locations or moving deeper to avoid warm waters 6 . Distributional records and documented range extensions can be used as a "canary in a coalmine"; fishes as sentinel species can indicate the relative health of the broader ecosystems.  Table 2. Species richness and abundance decreased with depth across all reefs pooled (one-way ANOVA). Shelf-break environments may be priority conservation hotspots, with high proportions of endemics 21,22 or species with restricted depth-ranges 33,115 . Australia has high total endemism and up to a third of its demersal fishes may be endemic 30 , therefore, there may also be high endemism in its demersal shelf-break fish assemblages. We may also be underestimating the Australian shelf-break's conservation value, as key bioregions including the upper continental slope of Queensland and the inter-reefal areas of the GBR are missing comprehensive fish assemblage information 31 . As genetic tools are increasing the resolution of cryptic speciation, there are likely differences detected between eastern and western Australian populations, and within species-complexes from neighboring regions 30    there were strong depth zonation patterns with characteristic and distinct demersal fish assemblages below 40 m. However, there was a "disjunction" at the shelf-edge between the continental shelf and slope bathomes assemblages (>40 m and <200 m), possibly due to "edge effects near the shelf break" 31 . We hypothesize that further investigation of shelf-edge habitats will demonstrate high diversity and distinctive communities. Shelf-break habitats should be considered intrinsically unique and a source of unforetold biodiversity and value.
There has been a rapid proliferation of reporting new species and new geographic records from mesophotic regions (e.g. refs 5, 20, 21, 25-27, 116-124). Even though underwater video cannot collect taxonomic samples 125,126 , it can be a useful method for identifying hotspots for conservation priorities 32 . There were species we were unable to identify. While these represent a small percentage (<5%) of fish species identified from BRUVS deployments, the observations indicate there are other new species at depths previously unrecorded in the GBR. In our study, fish identifications can be scrutinized as images are listed by CAAB (Codes for Australian Aquatic Biota) codes in the AIMS database for future re-assessments of the identifications.
In conclusion, we found that depth was a strong predictor of fish assemblages at mesophotic and sub-mesophotic depths of the GBR. Our findings on the GBR align with other tropical and sub-tropical studies in deeper habitats. Distinct fish assemblages and high species diversity was found along the depth gradient and this potentially contributes to high levels of endemism in Australian fishes and other parts of the world. These narrow depth distributions may constitute an inherent vulnerability to targeted fishing pressures and should be incorporated in future regional management strategies.

Materials and Methods
BRUVS deployment. Three reefs were sampled along the shelf-edge (Myrmidon, Unnamed and Viper) and one inter-reefal transect using a depth-stratified sampling design (Fig. 1). Two identical BRUVS units rated to 300 m were used, with an aluminum elliptical roll-bar frame enclosing a camera-housing with a flat acrylic front port and battery-powered spotlight (white) mounted above the top roll-bar. Sony high-definition Handicams HDR-CX110 were used, with focus set to manual infinity to maximize the field of view. Using a bridle-rope configuration with twice the water depth of attached line per deployment, each BRUVS was marked by surface floats and flags for retrieval. The bait arm consisted of a plastic conduit to a plastic mesh bag filled with ~1 kg of crushed pilchards (Sardinops sagax, see review for the effect of bait 127,128 ).
Forty-eight deployments were made in May, June and Sept 2014 on three cruises. All deployments were placed during daylight (50-300 m of water depth; 0700-1800) with most of the effort targeting 100-300 m in transects at each reef with three targeted depth strata. Our hypothesis was that there would be differences in the fish assemblage with depth. BRUVS were deployed in shallow (~100 m), mid (~150 m) and deep (~200 m) strata at each reef. Viper Reef is on a shallower sloping shelf-edge, so depths of >200 m were not available without travelling substantially further offshore. Instead, BRUVS were deployed shallower to get a similar bathymetric depth gradient (50-150 m) over a similar spacing between deployments (i.e. differences would be due to depth, not increased distance from shore). Within depth-strata BRUVs were haphazardly-spaced several hundred meters apart.

Fish identification and analysis of video metrics. Underwater imagery was read using Australian
Institute of Marine Science (AIMS) purpose-built software. The following details were noted: time on the sea-bed, time of first appearance of each species, and abundance N of each species until time MaxN (highest number of Table 3. Key fish indicator species per depth strata (multilevel pattern analysis). IndVal index (0-1) is accompanied by significance levels: *p ≤ 0.05, **p ≤ 0.01, ***p ≤ 0.001; "a" for species abundance data, "o" for occurrence (presence-absence) data.
individuals of a species per frame) reached, until the end of sampling (when the video left the bottom or when the tape finished recording). MaxN is a conservative estimate of abundance to eliminate the possibility of re-counting fish swimming in and out of the field-of-view 65 . Videos were read to its full length (27 to 84 minutes, average soak of 54 minutes) and later standardized for length of time of sampling (number of species present-absent per site for species richness, and number of fish per species for relative abundance, per 60 minute increment). Fish were identified to lowest possible taxa, with the assistance of fish experts, fish identification books and Fishbase.org 60 . Every effort was made to identify large, conspicuous fish in addition to smaller, cryptic species. Fish identification photographs and BRUVS deployment metadata are archived in the Australian Institute of Marine Science database and can be accessed by request.
Depth patterns. Species were summed across all sites for species richness and abundance. Where standardized values of total abundance and richness were used, the estimates were standardized by number of species per 60 minutes of sampling time. For our analyses two depth classification systems were used. For the one-way ANOVA which required a balanced design, three depth categories "Shallow" (50-115 m), "Mid" (128-160 m) and "Deep" (179-260 m) were used. For other analyses "Shallow" was further divided to two smaller categories to investigate the differences 50-115 m. Our sites were categorized in four depth strata: "upper mesophotic" (50-65 m), "middle mesophotic" (85-115 m), "lower mesophotic" (128-160 m) and "sub-mesophotic" (179-260 m). These strata represented breaks in the depth-stratified sampling design, but also aligned with previously documented transitional boundaries, including the ~150 m lower depth-limit of Mesophotic Coral Ecosystems (MCEs) 129 . Analyses were performed using several packages in R statistical software 130 (CRAN ver. 3
To evaluate the general trend of how species richness and abundance varied with depth, standardized richness and abundance were square-root transformed and data were tested for any significant deviation from normality (Shapiro-Wilks: species richness Wilks = 0.98, p = 0.66; abundance Wilks = 0.95, p = 0.07) to meet the assumptions of ANOVA. In our original design we had the factors 'Depth' (a = 3) and 'Reef ' (b = 3; Myrmidon, Unnamed, Viper) and site (n = 4) with an interaction between depth and site. The interaction was weak (p < 0.25), therefore, the factors were pooled as recommended by Underwood (1997) 131 . The factor 'Reef ' was pooled for a stronger test for the factor 'Depth' . ANOVA was performed for Depth (a = 3, n = 14) for both richness and abundance and two-tailed t-tests between depth groups with a Bonferroni correction was applied.
Mean standardized richness and abundance were also plotted in relation to depth strata separately by reef (Myrmidon, Viper and Unnamed; varied number of replicates within stratum). In addition, deployments were made along an inter-reefal transect (60-200 m, one replicate per depth). Shallower BRUVS sets from Viper Reef, one from on top of the submerged unnamed deep reef and the inter-reefal transect were included as an additional (50-65 m depth strata, n = 4). For analysis of separate families, we separated the Lutjanidae family into "deep" members (Etelis and Pristipomoides genera) and "other" (all other member species). Family analyses followed the one-way ANOVA for species richness and abundance.

Investigating fish assemblages.
We also wanted to investigate species associations as they may be better predictors of environmental conditions than species individually. This is often difficult because of positively-skewed frequency distributions and the high frequency of zeros in larger community composition datasets 132 . Species abundances (summed MaxN, maximum number of fish per species per site) were fourth-root transformed, which down-weights highly abundant species and reduces the skew in the distribution for each species 133 .
We used a Principle Coordinate Analysis (PCoA) ordination to visualise the differences between sites. Eliminating single-species occurrences (species only occurring at one site) from this analysis (58 of 130 species), we used 47 of the sites with 72 of the fish species in a Bray-Curtis dissimilarity matrix (packages vegan 134 , ecodist 135 ). Agglomerative heirarchical unconstrained clustering revealed 12 significant clusters (SIMPROF; packages cluster 136 , clustersig 137 ). For the ordination we color-coded the sites with the depth strata from the previous constrained univariate analyses and size-coded the symbols to correspond with species richness in the resulting biplot (functions capscale, vegdist). Capscale revealed ordination distances that were analogous to the original dissimilarities and is similar to redundancy analysis but can utilise non-Euclidean dissimilarities 134 . To determine which fish species corresponded with the variance between sites, we plotted the 15 species with the highest species scores.
We used species abundance data to perform multilevel pattern analysis of species by depth (functions multipatt, package indicspecies 138 ). This method first lists species associated with particular groups of sites and then indicator species analysis is independently conducted for each species 139 . This method requires multiple testing, but can help to predict the likelihood of individual species to attribute to that depth assemblage 139 . Statistical significance is interpreted based on the IndVal index, which is a measure of association between the species and that depth group and tested through a permutation test 140 . An advantage of the function multipatt is that it looks for both indicator species for individual depth strata as well as combinations of strata 139 . We also repeated this analysis using presence-absence (occurrence) data using Pearson's phi coefficient of association, a measure of the correlation used between binary variables (values of 0 and 1) 133 . Because this analysis is independently conducted for each species, we chose to include all species. Further, rare or single-species occurrences can be important for ecosystem functioning 141,142 . We considered the inclusion of all species to align with our objective of describing complete assemblages, and rare species (sensu FishBase) are of higher conservation concern as they can be more sensitive to ecosystem stresses than common species 143 . Measurements of temperature with depth. On the outer shelf-edge off Myrmidon Reef, near the 300 m isopleth (Fig. 1), a Seabird Conductivity Temperature and Depth recording device was slowly lowered (<1 m/sec) by hand to an estimated maximum depth before retrieval. The instrument was calibrated for 60 seconds below the surface before deployment. Repeated samples were made in early August 2009, 2010 and 2013.
All methods in this study were carried out in accordance with local guidelines and regulations for the GBRMP. Experimental protocols were approved by the animal ethics committee at James Cook University. Methods were non-invasive and no animals were taken in this fieldwork. Data availability statement. BRUVS deployment information, recorded species and linked images are available by request from the Australian Institute of Marine Science. Map bathymetric contour lines from Dr. Rob Beaman and Project 3DGBR (www.deepreef.org); map shapefiles provided by the Great Barrier Reef Marine Park Authority (http://www.gbrmpa.gov.au/resources-and-publications/spatial-data-information-services).