This study was done along a ~25km stretch of coastline in south-eastern, Australia (Figure 1). At the time of sampling, ships anchored in an unregulated anchor roadstead in this region, over an area spanning ~220km² (Davis et al., 2022) in water depths ranging from 35-60m. The seabed in this depth range is characterized by extensive, predominantly low-profile rocky reef, including channels containing mixed reef with boulders, gravel and sand (Linklater et al., 2019).

To delineate the anchor-roadstead and identify anchor-free reference areas we used a similar approach to Broad et al. (2023). We identified vessels at anchor using positional information (Automated Identification Systems (AIS) between September 2012 to June 2016 – just prior to fish sampling. This spatial information was then overlaid on high resolution maps with 5m gridded bathymetry derived from multibeam echosounder (MBES) surveys (www.aodn.org.au) allowing us to identify ‘anchored’ and appropriate ‘anchor-free’ reference locations on rocky reefs within a depth range ~35-50m. Conservative estimates based on the AIS data confirmed that reference areas had not been anchored upon for > 4 years.

Baited Remote Underwater Video (BRUV) was used to test the hypothesis that fish assemblages would differ across reefs ‘anchored’ and ‘anchor-free’ areas. BRUV has been used routinely for sampling fish assemblages as it is non-destructive and provides useful measures of species richness as well as relative abundance estimates of diver-shy and economically important fish assemblages (Watson et al., 2005; Kelaher et al., 2014; Rees et al., 2018a, b; Rees et al., 2021). As bait attracts fish, BRUVs also reduce the problem of zero-inflated datasets (Schramm et al., 2020). BRUV was particularly suitable in this study as units can be deployed in water depths exceeding 30m and considered too deep for diver-based assessments, particularly over large spatial scales (km’s).

The BRUV system used in this study consisted of a single GoPro HERO3+ camera (www.gopro.com) mounted within a waterproof housing camera housing. BRUV’s were deployed across four sites: two sites of ‘anchored’ reef and two sites that were ‘anchor-free’. Anchored treatments were interspersed in space and time and within each site we deployed 16 BRUV replicates (Figure 1) for a total of 64 BRUV deployments in the study. Each BRUV was deployed for 35 minutes to allow for settlement time and to achieve a 30-minute video sample. Previous research has demonstrated that 30 minute is an appropriate set time to provide a representative sample of temperate rocky reef fish assemblages (Harasti et al., 2015). Each sample was separated by a minimum distance of 200 m, consistent with other assessments using this method in the region (Kelaher et al., 2014; Rees et al., 2018a; Knott et al., 2021). Sampling was interspersed over several days during the late Austral winter and early spring (July to September, 2016) with all collections done in daylight hours (0900 to 1600) to alleviate potential effects of crepuscular feeding behaviors (Wraith et al., 2013). BRUV deployments used ~500 grams of thawed and crushed pilchards (Sardinops sagax) which has been determined as the optimal bait across a range of feeding guilds (including herbivores) (Harvey et al., 2007; Wraith et al., 2013) and was replenished for each subsequent deployment.

Videos were analyzed in the laboratory by a single experienced observer (AB) using the BRUV analysis software EventMeasure (http://www.seagis.com.au). All fish were recorded if they swam within 2 meters of the bait bag, providing a standardized (FOV) ~9 m³ (Malcolm et al., 2007). Species richness, total Max N, and Max N of each fish species were recorded for each video sample. Species richness (SR) is the sum of all the species recorded for each video sample and Max N is an estimate of the total abundance observed in one frame of a given species. Fish that displayed sexual dimorphism (e.g. Labrids) or protruding sexual organs (e.g. claspers in male elasmobranchs) provided improved estimates of these metrics and we adjusted MaxN counts of those species (see Supplementary Table 1). The total relative abundance for each sample, Total MaxN, was determined by summing the MaxN for all species over the 30 min sample. After video analysis, fish were classified into taxonomic groups (subclass/family) and trophic guilds to test for differences (for a full list with definitions see Supplementary Tables 1, 2). The later was based on their diet, previous research (Wraith et al., 2013; Swadling et al., 2019) or review of other supporting literature (Bell et al., 1978; Choat and Ayling, 1987; Bulman et al., 2001). In instances where no trophic information was available for a particular species, a similar, closely related species was used for inference. Species were grouped into one of three trophic groups; (i) herbivores; (ii) zooplanktivores and (iii) generalist carnivores.

Fine-scale habitat features for each BRUV replicate were analyzed using TransectMeasure (www.seagis.com.au; see Supplementary Tables 3, 4) following the method described in McLean et al. (2016). A 5 x 4 grid was overlaid on a high definition still frame from each BRUV deployment delineating cells to be scored (n=20). For each of the 20 cells, the dominant habitat type and vertical relief was scored following the CATAMI classification scheme (Althaus et al., 2015). The main habitat classifications were defined as either; unconsolidated sediments (sand or gravel), consolidated rock, macroalgae (any >5cm), sponges, ascidians and open water. For each cell containing benthos, an estimate of vertical relief was scored between 0 – 5 (Wilson et al., 2007). The BRUV Field of View (FoV) was also scored as ‘open water’, ‘video facing-down’, ‘facing-up’ or ‘limited FoV’ e.g. obstruction by a rocky outcrop. Data were exported from TransectMeasure using R scripts available from Langlois (2017). Cells of open water were removed before calculating the percentage cover of habitat and the mean vertical relief for each deployment. Depth (m) was recorded for each BRUV deployment in the field using the research vessel’s sounder.

Multibeam surveys (MBES) in the vicinity of the Port Kembla roadstead were completed by NSW state government during the period Oct 2014 – May 2017 and made available via the Australian Oceanographic Data Network (AODN: https://www.aodn.org,au). Survey methods are detailed in Linklater et al. (2019) utilizing both a Geoswath 125khz (Teledyne GeoAcoustics, U.K) and an R2Sonic 2022 (R2Sonic, USA) with a POS MV vessel reference system for motion correction. Bathymetric data were issued as depth-weighted averages of cleaned (Cube-modelled) multibeam soundings binned at 5m resolution and reported relative to Australian Height Datum. Data were imported to ArcMap (ESRI, USA) and additional layers of hillshade, slope and ruggedness calculated using tools including Benthic Terrain Modeller. These layers were then used to inform the selection of BRUV deployment sites over reef across the northern impact, southern impact and southern control locations. Bathymetric data for the section of the seabed in the northern control location were unavailable at the time of fieldwork planning. In this instance, a random location for each BRUV drop was nominated and then the presence of reef confirmed using the vessel’s onboard echosounder when on site. To calculate measures of reef structural complexity, the standard deviation of bathymetry values within 25 m, 50 m and 100 m radius buffers were calculated from the MBES data for each BRUV deployment.

The influence of anchoring and fine-scale habitat variables on the fish assemblage was analyzed using Generalized Additive Mixed Models (GAMMs; Hastie and Tibshirani, 1986; Hastie, 1990). GAMMs were chosen over Generalized Linear Mixed Models (GLMMs) to allow for non-linear relationships between the response and explanatory variables. A full-subsets modelling approach was used to determine which of the available explanatory variables were most important in influencing the fish assemblage (Fisher et al., 2018). Using this approach a complete model set was constructed and the selection of models was based on Akaike Information Criterion optimized for small sample sizes (AICc). Models within two delta AICc units (ΔAICc < 2) of the model with the lowest AICc were selected (Burnham and Anderson, 2002). If there were multiple models within two delta AICc unit, we selected the most parsimonious model(s) with the fewest explanatory variables. The relative importance of each explanatory variable was calculated by summing the AICc weight values for all models. Variables with higher summed values represent increased importance of that explanatory variable (Burnham and Anderson, 2002). To avoid issues of collinearity, candidate models could only contain explanatory variables with Pearson correlation coefficients < 0.28 (Graham, 2003). The maximum number of explanatory variables for each candidate model was limited to three to reduce over-fitting and ‘site’ was included as a random effect in all models to account for overdispersion and correlation in the data (Harrison, 2014). In all candidate models, explanatory variables were fitted with smoothing splines, with the ‘k’ argument limited to four.

Models were run on species richness, total relative abundance (total MaxN), abundance of trophic guilds, taxonomic groups and the abundance or presence/absence of common species present in >30% of samples (~16 spp.). Untransformed measures of the fish assemblage were used as response variables. The models for species richness were fitted with a Gaussian error distribution while models for total abundance included a negative binomial error distribution. GAMMs for the abundance of Ophthalmolepis lineolata and Bodianus unimaculatus were fitted with a Poisson error distribution. A Tweedie error distribution was fitted to all other relative abundance models due to the large number of zeroes (Tweedie, 1984). Taxa observed in low relative abundances were transformed to presence/absence and modelled using a binomial error distribution; this method has been shown to be an effective method for extracting ecologically meaningful data in rare or cryptic species (Joseph et al., 2006). Prior to analysis, the distribution of explanatory variables was plotted. Explanatory variables displaying a skewed distribution were transformed to ensure an even spread of values across the observed range. All data manipulation, analyses and plots were developed using the R language for statistical computing with the ‘dplyr’ (Wickham et al., 2020), ‘ggplot2’ (Wickham, 2016), ‘mgcv’ (Wood, 2011) and ‘FSSgam’ (Fisher et al., 2018) packages.

In total, we observed 68 species of temperate reef fish across 35 families from 64 BRUV deployments (4684 individuals; Supplementary Table 1). Overall, there was very strong evidence that ‘anchored’ rocky reefs influenced the total abundance of temperate reef fish observed, with both ‘anchored’ treatment and consolidated rock (%) included in the top model (Table 1; Figure 2). There were twice as many fish on ‘anchored’ reefs (98 ± 12 SE) compared to those that were ‘anchor-free’ (48 ± 6 SE) (Figure 3), with ~ 40% of the fish taxa examined containing the ‘anchored’ treatment in the top model (Table 1). Despite trends in fish abundance, anchor treatment had no effect on the richness of species occurring on rocky reefs, although as reefs increased in structural complexity, so did fish diversity (Figure 2).

Fishes within trophic guilds showed contrasting responses to anchor treatments. The zooplantivores, showed a distinct, almost four-fold increase in abundance on ‘anchored’ reefs (75 ± 12 SE) compared to ‘anchor-free’ reefs (~20 ± 4 SE). Increasing reefal habitat was also a strong predictor of abundance for this guild (Figure 3). Similarly, herbivores showed preference for greater rocky reef habitat (Figure 3), although they were indifferent to anchor treatments. In contrast, the null model was the most parsimonious for the abundance of generalist carnivores, indicating that neither anchoring treatment, nor any habitat features had any effect on the distribution of this trophic guild (Table 1; Figure 2).

Just six of the sixteen common taxa examined (observed in >30% of samples) revealed ‘anchored’ reefs to be a strong predictor of their abundance (Table 1; Figure 2). The top model for the common shoaling, zooplanktivore, Atypichthys strigatus included ‘anchored’ reefs explaining 20% of their distribution with abundance measures more than three times greater (40 ± 14 SE) than ‘anchor-free’ reefs (12 ± 5 SE) (Figure 4). Notably, there were three top models within 2 AICc for A. strigatus with all models including the ‘anchored’ treatment (Table 1). In contrast, the probability of detection of benthic elasmobranchs and the leatherjacket, Meuschenia freycineti were three and two times more likely to be observed on ‘anchor-free’ reefs (Table 1; Figure 4).

While the most parsimonious model for the occurrence of the economically valuable sparid - snapper, Chrysophrys auratus similarly included ‘anchored’ reefs and the habitat variables, mean relief and structural reef complexity (<25m) derived from bathymetric data (Table 1; Figure 2). Overall, these variables explained 30% of the abundance of snapper. Although, ‘anchored’ reefs was included in the top model, the effect size was small, and estimates had high variability (Figure 5).

In comparison, the Grey Morwong Nemadactylus douglasii included ‘anchor-free’ reefs as well as reef structural complexity (<100m) in its top model and these variables explained 38% of its abundance (Table 1; Figure 2). Evidence for preference of grey morwong for undisturbed reefs was very weak, with little distinction in abundance between ‘anchored’ (0.94 ± 0.2) and ‘anchor-free’ reefs (1.14 ± 0.2) (Figure 5). Likewise, for the Maori wrasse, Ophthalmolepis lineolatus, the top model included ‘anchored’ reefs, however the effect was weak with their abundance on ‘anchor-free’ reefs being 3.3 (± 1) compared to 2.5 (± 0.7) for those disturbed (Table 1; Figure 2). In contrast, water depth was a much stronger predictor, where the abundance of Maori wrasse declined with increasing depth (Figure 5).

Importance scores for two of the zooplanktivores, Caesioperca lepidoptera and Scorpis lineolata indicated weak support for ‘anchored’ reefs as a predictor for their distribution (Table 1; Figure 2). Although habitat variables best explained their distributions with the top models for C. lepidoptera including increasing consolidated rock and reef structural complexity <100m (Figure 6). While the top models for S. lineolata included shallower reefs and increased reef structural complexity <100m (Table 1; Figure 2).

Anchoring was not an important predictor for several species, including Pseudocaranx dentex, Trachurus novaezelandiae, Mecaenichthys immaculatus, Hypoplectrodes maccullochi, and Bodianus unimaculatus with habitat metrics providing more explanatory value in their spatial distribution (Table 1; Figures 2, 6). In contrast, no habitat variables included in our assessment, nor either anchor treatment, were of explanatory value in the abundance of several taxa, including Aulopus purpurissatus, Meuschenia scaber, Parma microlepis and Upeneichthys lineatus (Table 1; Figure 2).