A quantitative comparison of towed-camera and diver-camera transects for monitoring coral reefs

Study site

The Rowley Shoals is a group of three isolated oceanic shelf reefs near the edge of the continental shelf ~260 km from mainland north west Australia (Fig. 1). The reefs—named Mermaid, Clerke and Imperieuse—are separated from each other by >25 km of oceanic water (Wilson, 2013).

Benthic surveys were conducted in March 2018 and October 2019 at 10 permanently marked monitoring sites (Fig. 1). Sites were stratified by four habitats: ‘slope’ on the reef slope at ~6 m depth (at lowest astronomical tide, LAT); ‘crest’ on the reef crest at ~3 m LAT, both on the outer eastern boundary of Mermaid, Clerke and Imperieuse reefs; ‘lagoon floor’ at the base of coral bommies in the lagoon at Clerke and Mermaid reefs (9–12 m LAT); and ‘bommie’ on the top of coral bommies in the lagoon at Clerke and Mermaid reefs (~1 m LAT) (Fig. 1). Surveys were prioritized during slack tides, when winds were below 20 knots and swell was less than 1 m.

Permits to conduct these surveys were received from the Australian Government Department of the Environment (Mermaid Reef Commonwealth Reserve), the Western Australian Department of Primary Industries and Regional Development (DPIRD #2999) and the Western Australian Department of Biodiversity, Conservation and Attractions.

Survey methods

At each site there were fixed continuous transects, each separated by approximately 10 m. The total length of transects was 250 m on the reef slope (composed of five 50 m-long transects), 150 m on the reef crest (three 50 m-long) and 120–150 m at the lagoon floor and coral bommie sites (three 50 m-long in 2018 and six 20 m-long in 2019) (Fig. 2). Reef crest and slope transects ran along depth contours parallel to the reef crest, while the lagoon floor and bommie transects wrapped around the base or top of the bommies respectively (Figs. S1.3 and S1.4).

Sites were surveyed on scuba following the standard operating procedures (SOP) of the AIMS long-term monitoring program (Jonker, Johns & Osborne, 2008). The dive team required 2–3 personnel (photographer and dive buddy, surface attendant if needed). The divers used a digital Canon PowerShot G1 X Mark II (12.8 MP, 1.5 inch sensor) in an underwater housing, photographing the substrate every metre along the transects, with the camera held perpendicular ~50 cm from the substratum. The resulting images each capture an area of approximately 23 × 30 cm. It took ~45–60 min to collect images along five 50 m end-to-end transects, depending on the substratum complexity of the reef and the water conditions, such as surge or current. After the images were collected, a geo-referenced track of the transects was obtained by a swimmer following the transect line at the surface with a GPS.

The ‘tow-camera’ was designed to be hand deployable from a tender or small vessel for the specific purpose of surveying shallow reef areas that have traditionally been surveyed by divers. The system is therefore lightweight, battery powered, and much smaller than more traditional tow-video systems (e.g. Carroll et al., 2020), measuring 25 × 27 × 75 cm. The tow-camera system consists of an aluminum frame, to which an oblique-angled high-definition Sony analogue video camera is attached, providing a live video feed via a 30 m loadbearing cable to a monitor (Atmos) on the boat. A downward-facing Panasonic Lumix DMX-LX100 Mk I (13 MP, Four Thirds sensor) digital camera, was set to photograph at 3 s intervals with an INON Z240 strobe (Fig. S1.1). The tow-camera was deployed from a 6 m rigid hulled inflatable boat and minimum of two people conducted the surveys: one adjusted the height of the system to approximately 0.5 m from the substratum while viewing the live video feed (Bowden & Jones, 2016; Carroll et al., 2020) and one drove the boat along the GPS track of the diver transects at a speed of 1–2 knots. Slower speeds were preferable, but the boat needed to move fast enough to stay on track. The optimum time to complete a 50 m transect, assuming the boat maintained exact course along the transect, would be 2.5 min, corresponding to a speed of 0.33 ms⁻¹ (0.65 knots) and with a photo interval of 3 s, one photo per 1 m. Each transect was repeated 2–4 times, to increase image replication, so total survey time for each site was approximately 30–60 min. Images were geo-referenced based on the image timestamp matched to the GPS on the boat, which gave an approximation for the position of the tow-camera, which followed behind the boat at approximately 5–10 m. We estimate the distance of the tow-camera from the substrate varied between 25 and 175 cm, meaning the area captured in an individual image likely varied between approximately 13 × 18 cm and 81 × 108 cm.

Image selection

The tow-camera continuously photographed (every 3 s) along the full length of all transects at a site, so individual transects were not distinguished at the time of the tow (Fig. 2). Two to four replicates were collected by turning the boat and repeating the same track over the transects. The geo-location of each tow-image was then used to classify each image into one of the transects; polygons were created from the GPS diver track in QGIS to represent each transect with a gap between (see Fig. S1.2). Any images that were in the gap between transects and those collected when the boat was turning around were removed. Images from replicated tows were pooled into their respective transects based on image position.

The shortest distance between each tow-camera image and the diver GPS track was measured in metres using ‘distance to nearest hub’ in QGIS Desktop 3.10.0. The mean and range of distances from the diver track were calculated for each tow and any tow-camera images more than 5 m from the diver track were removed from further analyses. On the diver transects, one image was taken every metre, i.e. 50 images per 50 m transect, or 20 images per 20 m transect (Table 2). For the tow transects, the total number of images varied depending on the number of images within the 5 m buffer surrounding the diver track, the speed of the boat and the number of replicate tows at a site.

Image analysis and benthic categories

Two observers experienced in coral reef taxonomy analysed the images using point sampling, where the substratum under each of five fixed points overlaid on each image was assigned to the finest taxonomic resolution that could be confidently identified (see Jonker, Johns & Osborne, 2008); for corals this was usually to genera.

Benthic data were classified and analysed at two taxonomic scales: (i) the ‘benthic community’, which grouped all hard corals together and considered the proportions of abiotic substratum, macroalgae, turf/coralline algae and soft corals as separate categories, as well as the proportions of unidentified substratum and points that did not have sufficient image quality to be identified; and (ii) the hard ‘coral community’. A combination of genera, family and growth form described the coral groups to best represent the distinctive groups present at the Rowley Shoals (see Table 1).

Covariates

We were interested in whether the slope, depth, or substratum complexity accounted for differences between the data from the two methods (Table 2). We used laser airborne depth sounder data (LADS, https://www.navy.gov.au/ran-aviation-history/laser-airborne-depth-sounder-lads), to calculate the mean and standard deviation of depth along the transects, and to estimate substratum complexity. The LADS point cloud was interpolated to form a raster surface using a unidirectional and optimised universal kriging model in the Geostatistical Analyst module in ArcGIS^™ version 10.6. The final raster resolution was 2.66 m² pixels which was a stable estimation of the variogram model sill point and finer scale interpolation would have introduced stochastic error. From the same LADS raster, we used the ArcGIS^™ surface analysis tool in Spatial Analyst to calculate the mean slope of each transect, as the maximum rate of change between each cell and its neighbours in degrees.

We also considered whether there was a relationship between the number of tow-camera images per metre and the difference between the two methods, to test whether more tow-camera images increased the accuracy of the tow-camera relative to the diver.

Multivariate statistical analyses

Benthic and coral community assemblage data were analysed by transect using multivariate and univariate methods. The percent cover of each benthic group was calculated from the number of points assigned to each group in the images divided by the total number of points per transect. All multivariate analyses were conducted in PRIMERv7 (PRIMER-e) using Bray–Curtis dissimilarities calculated from square-root transformed data (Clarke & Gorley, 2015). Community assemblage data were visualised with non-metric multidimensional scaling (nMDS) (Clarke & Gorley, 2015) to examine potential separation between groups associated with the four main factors of interest: method, habitat, reef and year. Vector overlays of Pearson correlations >0.6 were used to identify the most influential benthic and coral groups driving separation in the data. Additional vector overlays with Pearson correlations >0.15 were used to visually assess the extent to which the physical covariates—depth, depth variation or slope—were associated with separation in the data cloud. These physical variables were transformed in cases where their distribution was particularly right skewed: slope and substratum complexity were log_e transformed.

We used permutational multivariate analysis of variance (PERMANOVA; Anderson, 2001) to test whether method, habitat, reef, survey year, or two-way interactions between these factors affected the composition of benthic and coral communities. Higher order interactions were excluded. Models were reduced by excluding two-way interactions that had a p-value > 0.25 before re-running the analysis. Residuals were permutated (9,999 permutations) under a reduced model, and tests were based on Type III (partial) sums of squares. We tested the hypothesis that the ability of the tow-camera to detect similar benthic and coral community assemblages to the diver method may vary between habitats and therefore specifically examined pairwise tests on the interaction of habitat and method.

Univariate statistical analyses

Differences in the percent cover detected between methods of each of the broad benthic groups and hard coral taxa were investigated using linear models in R (R Development Core Team, 2018). The analysis was conducted at the transect level and designed to investigate whether mean depth, slope, substratum complexity, habitat and/or number of tow-camera images per transect correlated with the difference in the percent cover between the two methods, $\mathrm{\Delta }\mathrm{c}\mathrm{o}\mathrm{v}\mathrm{e}\mathrm{r}{}_{i,j}$, where

$$\mathrm{\Delta }\mathrm{c}\mathrm{o}\mathrm{v}\mathrm{e}\mathrm{r}{}_{i,j}=\mathrm{c}\mathrm{o}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{d}\mathrm{i}\mathrm{v}\mathrm{e}\mathrm{r}{}_{i,j}-\mathrm{c}\mathrm{o}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{t}\mathrm{o}\mathrm{w}{}_{i,j}$$

and i is the benthic group and j is the transect in question.

$\mathrm{\Delta }\mathrm{c}\mathrm{o}\mathrm{v}\mathrm{e}\mathrm{r}$ was modelled as the response variable and habitat, reef and year were included as fixed effects. Site was not included directly in the model, because there was only one site per combination of these fixed effects, and therefore the combination of these factors determined the sites indirectly, thus accounting for spatial autocorrelation of transects. The response variable $\mathrm{\Delta }\mathrm{c}\mathrm{o}\mathrm{v}\mathrm{e}\mathrm{r}{}_{i,j}$ was bound by a minimum of −100% and a maximum of 100%, and a Gaussian error model was used. Homogeneity of variance was checked visually by plotting the residuals against the fitted values. Residuals were examined and were satisfactory for conforming to the normal distribution assumption in most cases. For the coral taxa groups that only appeared on a handful of transects—Diploastrea, Fungiidae, Pavona and unidentified coral—the residuals were not normally distributed, and we did not have confidence in these models, and therefore did not include these statistics. While we may have considered a zero-inflated Gaussian model, or presence-absence models, we did not because there were so few occurrences with which to provide any robust conclusions about these groups. Collinearity among the continuous predictor variables was assessed: slope and substratum complexity were highly correlated (r > 0.75) and therefore only slope was included in the models. Predictor variables were transformed to reduce the skewness where appropriate: the slope and the number of images m⁻¹ were log_e-transformed (Table 1). A full model containing all variables was fit first and sequentially simplified based on Akaike Information Criterion.

We focused on the intercept of the models when predictor variables were at their mean to test whether there was a significant difference between the two methods under average conditions. Therefore, we mean-centred the continuous variables—number of images, slope and depth—and used sum-to-zero contrast matrices for the categorical variables habitat, year and reef. The model intercepts were directly interpreted as $\mathrm{\Delta }\mathrm{c}\mathrm{o}\mathrm{v}\mathrm{e}\mathrm{r}{}_i$ under average conditions for continuous variables and marginalised over the factors.

In interpreting the results, a mean positive difference in cover ( $\mathrm{\Delta }\mathrm{c}\mathrm{o}\mathrm{v}\mathrm{e}\mathrm{r}{}_i$ > 0) meant the diver images detected a higher cover of i than the tow-camera images. Conversely, a mean negative difference in cover ( $\mathrm{\Delta }\mathrm{c}\mathrm{o}\mathrm{v}\mathrm{e}\mathrm{r}{}_i$ < 0) meant the diver images detected a lower cover of i than the tow-camera images.

Benthic community structure

Variation in benthic community structure was explained by the method of sampling, habitat, reef, survey year and interactions between these factors (Table 3). The main effect of method was significant, indicating that, over and above the interactions, there was an overall difference in the benthic community structure detected between the tow-camera and diver methods (Table 3). The greatest components of variation (a measure of the magnitude of the effect) were for the interaction between reef and habitat, and habitat as a main effect (Table 3). A significant method-by-habitat interaction indicates that the differences between the methods varied among the four habitats. Significant differences in the benthic community composition exisited between the sampling methods within all four habitats, but to different extents (Table S3.1). Among the habitats, the benthic communities separated mainly due to differences in the cover of hard corals; in particular, the reef crest and slope had more hard corals than the lagoon floor and bommies (Fig. 3A). Across all transects, the tow-camera method recorded a lower percent cover of crustose coralline algae and hard coral, a higher cover of abiotic substratum, and a higher proportion of images that could not be assigned to a group due to poor image quality (‘unidentified substratum’).

Correlations between the site covariates (slope, substratum complexity, depth) and the multivariate ordination axes were weak (Fig. 3A). Univariate analyses indicated that steeper slope (correlated with substratum complexity) was associated with the tow-camera underestimating the cover of hard corals, while obtaining a higher cover that could not be identified (Table S3.4; Figs. 4 and 5A). Increases in depth correlated with the tow-camera recording lower cover of crustose coralline algae, and more abiotic substratum relative to the diver method. The two methods became more comparable in their estimates of hard corals as depth increased (Fig. 5B).

Coral community structure

The coral community structures identified by the two methods were more similar than the broader benthic community structures (Fig. 3; Table 3). As expected, habitat was the most important correlate of variation in coral community structure, but there was also a significant interaction between habitat and method (Pseudo-F = 2.38, df = 3, p = 0.0016; Table 3), indicating differences between the methods depended on the habitat. Pair-wise tests showed no differences in the coral assemblages identified by the two methods at the lagoon floor (t = 1.10, df = 32, p = 0.32) and bommie sites (t = 1.18, df = 30, p = 0.25), but significant differences at the reef crest (t = 1.95, df = 31, p < 0.001) and reef slope sites (t = 2.33, df = 55, p < 0.001; Table S3.1).

The most notable differences in the mean cover of coral taxa between methods was a lower cover of encrusting corals, Isopora, Merulinidae, Millepora and digitate and corymbose Acropora detected by the tow-camera, and a lower cover of tabular Acropora in the diver images (Fig. 4; Table S3.4). Differences in the cover of coral taxa between methods varied among habitats (Fig. 6), reefs and survey years (Table S3.4). For example, the tow-camera surveys detected lower cover of encrusting hard corals relative to the diver surveys at bommie sites but more cover of this group at the lagoon floor sites (Fig. 6; Table S3.4). The diver surveys tended to detect more Merulinidae, Millepora, Pocilloporidae and Porites than the tow-camera (Fig. 4; Table S3.4) as slope increased and more Merulinidae and digitate and corymbose Acropora as depth increased (Fig. 5).

For several benthic and coral groups, an increase in the number of images per metre increased the comparability of the tow-camera data to the diver data (hard coral: t = 0.72, p < 0.001; crustose coralline algae: t = 0.83, p < 0.001; abiotic substratum: t = 0.86, p < 0.001; Merulinidae: t = 0.19, p = 0.028; Millepora: t = 0.21, p = 0.030; Figs. 5C and 4F; Table S3.4).