Marine benthic images are commonly used to quantify habitat composition, ground-truth remote data and predict the extent of habitat types (Pelletier et al., 2020). Such imagery is now widely used to calibrate spatial analyses such as distribution models and change-over-time mapping (Mastrantonis et al. in review). Benthic images captured by platforms such as divers, drop cameras, towed-video, Remotely Operated Video (ROV), and Autonomous Underwater Video (AUV) are generally acquired from downward-facing cameras, with a field of view that is relatively constrained (~70o x ~40o) and covers a small area per sample unit (~ 1 m2, Bennett et al., 2016; Sheehan et al., 2016). Horizontal-facing images, using the same field of view, have a larger area (~25 m2) and are useful in a variety of situations and ecosystems (Bennett et al., 2016). Downward-facing images generally provide higher taxonomic resolution for sessile assemblages and sub-canopy species than horizontal-facing images, and improved estimates of mobile invertebrate numbers (Perkins et al., 2020). However, the larger area per sample unit of horizontal-facing images better aligns with resolutions of remote sensing products such as bathymetric lidar (~25 m2) and optical remote sensing platforms (~100 m2). Obtaining ground truthing data at a commensurate scale to remotely sensed products is an important consideration when modelling extent or community composition (Mastrantonis et al. in review). Horizontal-facing imagery is also more effective for monitoring the cover of erect habitats including canopy algae and corals (Bennett et al., 2016; Vergés et al., 2016), particularly if stereo images are captured allowing the dimensions of biota to be measured (Langlois et al., 2021). Stereo images further allow the sample unit to be standardised across varying visibility (Broad et al. 2023; McLean et al. 2016). The structural dimensions (i.e. height) of benthic biota can be an indicator of anthropogenic and environmental impacts, with imagery from Baited Remote Underwater stereo-Video (stereo-BRUV) surveys being successfully used to measure the recovery of soft-coral height after the cessation of trawling across an area of continental shelf (Langlois et al., 2021), and the impacts of marine heat waves on macroalgal canopy height (Vergés et al., 2016).
Spatially-balanced survey designs can increase sampling efficiency by evenly spreading samples in space and across the range of covariates of interest (e.g., depth and relief) (Robertson et al., 2013). Typical platforms for collecting benthic images (i.e. divers, towed-video, ROV, and AUV) have logistical constraints that result in them generally being deployed along transects, or in discrete patches or mosaics (Sheehan et al., 2016). By contrast, drop cameras provide point-samples, providing a more spatially independent method of gathering benthic data (Robertson et al., 2013). Where rapid repeated deployments are possible, drop cameras are suited to ground-truthing relatively large spatial areas (Pelletier et al., 2020) and sites requiring validation can be chosen based on covariates of interest (Mastrantonis et al. in review). Transect-based sampling can also be used in a spatially balanced manner, but care must be taken to account for spatial dependence within transects and clusters of transects (Foster et al., 2020). Regardless, transect-based and locally-dense sampling can introduce clusters of samples within similar environmental settings, or spatial bias, that can weaken subsequent statistical analyses (Robertson et al., 2013). While drop cameras have clear logistical and efficiency advantages for sampling larger areas, due mainly to the brevity of their deployments and relative ease of obtaining independent observation units, deeper water environments (>200 m) increase time for deployment and create logistical challenges. Below these depths, multi-platform swarms, either of AUVs and ROVs conducting transects, are likely to be more cost-effective (Liu et al., 2023).
We have developed a remote wide-field drop camera system, called the Benthic Observation Survey System (BOSS), with a combined field of view of approximately 270˚ (Figs 1-2), amenable to stereo- or mono-camera configurations (Fig. 3). The design originated from an integrated fibre-optic camera system developed by Rick Starr at Moss Landing Laboratories for sampling demersal fish assemblages, that developed from rotating stereo-video landers (Starr et al., 2016). The system was adapted to be able to be rapidly deployed and retrieved from a variety of vessels into water depths of 2 to 200 m and is self-righting on the seabed (Figs 1-3), with a single deployment in 30 m of water taking just 8 minutes with a 5-minute bottom time. This tool is suited to the collection of widespread georeferenced point samples, enabling the cost-effective sampling of broad areas using spatially-balanced sampling designs, to produce benthic habitat coverage predictions (Fig. 1) or inform other environmental assessments (i.e. benthic biota dimensions). We demonstrate this method through a project led by Traditional Owners of the south-west of Australia to characterise the habitats associated with ancient submerged coastline features across the continental shelf, to inform further detailed analysis (Langlois et al. ). We provide a standard operating protocol (SOP) for the BOSS with information on system design, field operation, image annotation, data validation, and examples of a workflow to generate a habitat map product (Fig. 1). We highlight the benefits of using multiple horizontal fields of view to characterise benthic habitat heterogeneity but also suggest that future studies should investigate the potential of collecting demersal fish assemblage information comparable to Starr et al. (2016).
Figure 1: BOSS workflow for benthic composition ground truthing and production of predictive spatial models. a) Spatially balanced design with inclusion probability, b) drop camera, c) imagery annotation, d) quality control, e) predictive modelling and validation to produce f) probabilities of occurrence for individual habitat classes and g) categorical habitat predictions.