Augmenting Seafloor Characterization via Grain Size Analysis with Low-Cost Imagery: Minimizing Sediment Sampler Biases and Increasing Habitat Classification Accuracy

Abstract

Bottom-grab samplers have long been the standard to describe nearshore marine habitats both qualitatively and quantitively. However, sediment samplers are designed to collect specific grain sizes and therefore have biases toward those sediments. Here, we discuss seafloor characterizations based on grain size analysis alone vs. grain size analysis augmented with quantitative benthic imagery. We also use both datasets to inform a prevalent benthic habitat classification system. The Coastal and Marine Ecological Classification Standard (CMECS) was used to test this hypothesis. CMECS was adopted by the federal government to standardize habitat classification in coastal U.S. waters. CMECS provides a hierarchal framework to define and interpret benthic habitats but does not prescribe specific sampling methods. Photography has been utilized for many decades in benthic ecology but has rarely been employed in habitat classification using CMECS. No study to date has quantitatively examined the benefit of incorporating benthic imagery into the classification of biotopes using CMECS. The objective of this study is to classify a roughly 1 km² subtidal area within Herring Cove in Provincetown, MA with CMECS and quantify the benefit of augmenting classification with low-cost imagery. A benthic habitat survey of the study area included grab sampling for grain-size analysis and invertebrate taxonomy, benthic imagery, water quality sampling at 24 sampling stations, and acoustic mapping of the study area. Multivariate statistical analyses were employed to classify biotic communities and link environmental and biological data to classify biotopes. The results showed that benthic imagery improved the classification and mapping of CMECS components. Furthermore, the classification of habitats and biotopes was improved using benthic imagery data. These findings imply that the incorporation of low-cost benthic imagery is warranted in coastal benthic biotope classification and mapping studies and should be regularly adopted. This study has implications for coastal benthic ecologists classifying benthic habitats within the CMECS framework.

1. Introduction

Accurate descriptions of marine habitats are an integral part of marine management [1], and the need for coastal and estuarine habitat classification is growing [2]. However, adequately classifying discrete subtidal habitats spatially is challenging as they are composed of continuous gradients of individual components (e.g., sand, gravel, and submerged aquatic vegetation) and are temporally dynamic [3]. Consequently, habitat descriptions may vary based on temporal and spatial resolution, as well as the methodology of data collection [4].

A marine benthic habitat can be defined as a physically distinct area of the seafloor distinguished by its abiotic and biotic characteristics, operating at a particular, but dynamic, spatial and temporal scale [1,5]. They are inextricably linked to the greater ecosystems in which they reside through the dynamic exchanges of energy, mass, and nutrients between benthic and pelagic habitats (e.g., predation, migration, and bioturbation) [6]. Thus, the temporal and spatial arrangement of benthic habitats and the resulting benthic-pelagic coupling shape coastal and estuarine ecosystems and the ecological services they provide [6].

A wide variety of tools and technologies to classify and map benthic habitats are available and their use directly influences the scale and resolution of the resulting ecological classifications [4,7].

A standardized system of classification and mapping is necessary to not only compare habitat information that varies spatially and temporally but also to efficiently communicate findings between studies that employ different methodologies [2,8]. Consequently, in 2012, the Coastal Marine and Estuarine Classification Standard (CMECS) was approved by the Federal Geographic Data Committee (FGDC) as a U.S. federal standard (FDGC 2012). At its core, CMECS is a catalog of hierarchical terms that provide a means of classifying ecological units using a standardized format [2]. The four components of CMECS that provide specific tools for describing sampling sites are the Water Column Component, Geoform Component, Substrate Component, and Biotic Component [2]. Although CMECS provides a framework in which habitat information can be compared in a standardized way, it is “data agnostic” in that it does not prescribe a certain set of methodologies to employ in data collection [2]. It is imperative then that researchers take advantage of tools that help maximize the resolution and accuracy of habitat classifications, and that they fully utilize new and existing data [2].

Benthic imagery has been utilized in marine science for many decades as an alternative to the traditional “kill ‘em and count ‘em” method [9] and presents many opportunities as a sampling technique, and consequently, is used in a wide variety of applications (e.g., hard-bottom areas, coral reefs, and artic habitats) [10,11,12,13,14,15,16]. Benthic imagery also allows for the rapid collection of data over large spatial scales [17].

Benthic imagery, due to its non-destructive and minimally invasive nature, is particularly valuable for collecting data on sensitive habitats like coral reefs [13,14,16] and coastal benthos in high latitudes, where traditional sampling might cause significant disturbance [15]. For example, Smith et al. (2015) combined the use of a multibeam echosounder and sidescan sonar with video transects to map benthic habitats in a 50 km² study area in coastal Antarctica without compromising the slow-developing biological assemblages. This technique also allows for rapid data collection over large areas, such as estimating scallop populations across 70,000 km² of the northeastern U.S. continental shelf [17], where traditional methods would be impractical and disruptive [17,18]. Incorporating benthic imagery into traditional sampling can also reveal ecological patterns that might otherwise go unnoticed, such as the habitat preferences of black seabass recruits within Crepidula reefs [19], which physical sampling alone could not reliably capture.

While benthic imagery has limitations, such as reduced taxonomic resolution compared to physical sampling [10], it remains an essential tool when traditional methods are unfeasible or unethical [10]. Coupling imagery with physical sampling, as demonstrated by Zajac et al. [20] in studies mapping habitats in Long Island Sound, enhances the resolution of both biotic and abiotic descriptions of the seafloor. The continued development of technologies like Structure-from-Motion (SfM) photogrammetry, which enables the creation of 3D habitat maps [14,16], is expected to further increase the use of underwater imagery in marine science. Despite its long history of use [9], the integration of benthic imagery into frameworks like the Coastal and Marine Ecological Classification Standard (CMECS) remains limited [21,22,23,24,25,26], but its potential benefits suggest a growing role in future ecological studies.

The primary objective of this article, which was part of a larger study [27] funded by the US National Park Service (PMIS# 201780), is to quantify the benefit of low-cost benthic imagery to better classify bottom sediment type and grain size and augment traditional benthic habitat mapping techniques. In particular, the utility of attaching a consumer-grade action camera (GoPro Hero 5^TM) when collecting bottom-grab samples. Sediment samplers are often biased toward particular grain sizes due to the size of the apertures and closing mechanisms of the sediment samplers themselves and may preferentially collect finer sediments while excluding coarser grains or vice versa. The resulting bias occurs because the physical design of the sampler limits the range of sediment sizes that can be efficiently collected, leading to a biased representation of the sediment distribution and skewing the analysis toward the sizes the sampler is optimized for. While Van Veen samplers, similar to the one used in this study, are well-suited to collect sediment samples in nearshore sandy coastal areas [28], one pebble-sized grain can prevent the sampler from closing. Thus, a bias away from gravel-sized material may exist when using these samplers. Benthic habitat surveys were conducted to classify the CMECS Substrate, Geoform, Water Column, and Biotic Components. The accuracy of habitat classifications and the efficacy of environmental predictor variables were compared between low-cost benthic imagery and traditional grab sampling using multivariate statistical techniques. Augmenting physical sampling with imaging resulted in the identification of meaningful biological–physical linkages, which helped to classify and map biotopes within the study area. Benthic imagery was used to supplement CMECS Component classifications and improve the classification of biotopes. The results from this study have implications for all researchers mapping coastal or estuarine benthic habitats, especially those aiming to rapidly and inexpensively maximize informational output and accuracy of classifications.

2. Materials and Methods

2.1. Study Area

The study site (Figure 1), Herring Cove in Provincetown, MA, USA, is dominated by subtidal sand flats and banks, comprised of very coarse sand (median grain size ~1.2–2.0 mm), with interspersed gravel reefs [27]. The offshore extent of the flats is a steep slope, in which depths quickly drop off from 10 to 45 m [29]. Herring Cove is a mesotidal coast, with mean and spring tidal range of 2.83 m and 3.07 m, respectively (https://tidesandcurrents.noaa.gov/ accessed on 21 July 2021), in a partially sheltered environment due to its fetch-limited southwest orientation toward Cape Cod Bay and is located within the Cape Cod National Seashore.

2.2. Benthic Grab Sampling

A stratified random sampling scheme was designed using ArcMap v.10.x placing 24 stations in 1 km². The stations are positioned along approximately 2.5 km of shoreline. The survey vessel, the R/V Marindin, was anchored before samples were collected. GPS data points were collected for each sample using a Garmin^® 78x and exported as a CSV for analysis.

A Young-modified Van Veen grab sampler (bottom penetration: ~0.1 m, sample volume: 0.004 m³) was used to collect samples in triplicate from the seafloor. Samples were subsequently sorted and identified by trained personnel according to Mittermayr et al. 2020. The sediment remaining after invertebrates was sorted, sieved at 4 mm, and retained for further analysis. The organisms collected were identified to family level [30,31,32] using the best available identification keys and literature [33,34,35,36,37], enumerated, and archived.

2.3. Sediment Processing

Sediment samples were taken from the top 5 cm of the grab and processed for grain-size analysis and organic matter content analysis (following Heiri et al. 2001 [38]). Grain-size analysis of sediments < 4 mm in diameter was performed using a Beckman-Coulter LS 13 320 Laser Diffraction Particle-size Analyzer (Beckman Coulter Inc., Brea, CA, USA) and a Retsch Tech Camsizer (Microtrac Retsch GmbH, Haan, Germany). Sediment > 4 mm was not analyzed via particle size analyzers for this study. All particle-size analyzer outputs were subsequently processed with the software GRADISTAT v8 [39] and reported using the Wentworth grain-size scale [40]. GRADISTAT computed summary statistics such as the mean, median, mode(s), standard deviation, skewness, and kurtosis based on the log-normal distribution according to Folk [41]. Percentages of mud, sand, and gravel were calculated for each sample, which were used for the CMECS Substrate Component classification [2,39].

2.4. Benthic Imagery

A GoPro Hero 5™ (San Mateo, CA, USA) was mounted to the sampler and recorded video each time it was lowered into the water. The camera was mounted facing downward to collect data on seafloor habitat characteristics. Still images were captured from the video using Adobe Premier CC™ v24 at either the moment before the sampler made contact with the seafloor or just after sediment turbidity had settled. Care was taken to minimize the disturbance of sediments, which can generate turbidity clouds that obscure imagery quality. Still images were captured at a point during the sampling video that yielded the best image clarity for that particular video file. Videos were collected regardless of the retention of the grab sample.

The percent cover of sand, gravel, shell material, and benthic macroalgae (BMA) was visually quantified using a random point overlay on still images with the program Coral Point Count with Excel Extensions (CPCe) [42]. Sand was defined as fine unconsolidated mineral sediments < 2 mm in size [2,40]. Gravel was defined as coarse unconsolidated mineral sediment > 2 mm in size [2,40]. The shell material was defined as shell hash and rubble, living or non-living, >2 mm in size [2]. The BMA category included filamentous, sheet, and leathery/leafy benthic macroalgae. The percent cover was calculated for BMA as a whole, but the presence and absence were recorded for individual growth morphologies (e.g., filamentous, sheet, and leathery/leafy benthic macroalgae) of benthic macroalgae.

To identify the optimal number of random points, or optimal point count (OPC), to use for benthic cover estimation, random point sample sizes and associated margins of error were estimated using the binomial probability distribution based on Larsen and Marx [43]. The initial reconnaissance was conducted on 10 images (out of 92 total images), which were strategically chosen for being representative of different levels of benthic cover categories (e.g., images with high, low, and intermediate levels of sand, gravel, SAV, and shell material). Percent cover analyses were replicated 3 times using 5, 25, 50, 75, and 100 random points for each of the 10 reconnaissance images. The mean percent cover and standard deviation were calculated for each image from the 3 replicate analyses. Additionally, the amount of time required to analyze the percent of cover was recorded.

First, a binomial probability (p) of 0.5 and an alpha value (α) of 0.05 were used to estimate the sample sizes necessary to achieve a margin of error (d) of 0.01, 0.05, 0.1, and 0.15 in benthic cover estimations. Next, the benthic cover estimate precision achieved by using the different random point sample sizes (5, 25, 50, 75, and 100 random points) was quantified. Using a binomial probability of 0.5 provides the most conservative estimate of sample sizes and precision because the confidence intervals are greatest when p = 0.5 (E. Gallagher, pers. comm.) and the margin of error (d) is ½ the 95% confidence interval in the binomial probability distribution [43].

2.5. Vessel-Based Acoustic Surveys

Acoustic surveys were conducted in the study area at a water depth of between 1 and 8 m in October 2017 using a bow-mounted Edgetech™ 6205 (EdgeTech, West Wareham, MA, USA) phase-measuring sidescan sonar, which collects both dual-frequency sidescan imagery and swath bathymetry. This instrument also collects uncalibrated bathymetry-mode backscatter [44], and its use in this study was outside the scope of the project. The operating frequency for sidescan imagery is 550 kHz and 1600 kHz, and in shallow water, this system has been shown to yield absolute localization of targets of 0.41 ± 0.26 m [45]. The bathymetric range and vertical resolution were each 1 cm (Edgetech, 2019). Motion data were collected with a Teledyne™ TSS DMS-05 motion reference unit (Teledyne Technologies, Thousand Oaks, CA, USA) mounted onto the sonar housing, which measured heave to 5 cm and roll and pitch to 0.05° [46]. The position was collected using a Trimble^® R8 GNSS system using real-time kinematic GPS (RTK-GPS) (Trimble Inc. Westminster, CO, USA). For more details about these instruments, see Borrelli et al. [47]. A bathymetric raster was interpolated with 0.5 m resolution. The depth of each sampling station was extracted from the bathymetry raster in ArcGIS.

2.6. CMECS Geoform Component

The CMECS Geoform Component was classified by applying NOAA’s Benthic Terrain Modeler (BTM) toolbox in ArcMap v.10.7 [48] to the bathymetric raster data. The BTM toolbox accomplishes this by calculating the difference between the depth of a raster cell and the mean depth of all cells in a defined annulus (donut shape) surrounding the cell [48]. The broad-scale BPI was calculated using an annulus with an inner radius of 25 m and an outer radius of 250 m. The fine-scale BPI was calculated using an annulus with an inner radius of 5 m and an outer radius of 25 m. The radii were selected based on previous work conducted in the area [29,49]. The BTM toolbox then takes the values from the slope, fine-scale, and broad-scale BPIs and, by applying a user-defined classification dictionary, classifies each raster cell to a geoform. The classification dictionary used in this study was developed by Borrelli et al. [29] for previous habitat mapping projects using CMECS within the same region.

2.7. CMECS Substrate Component

The CMECS Substrate Component was first classified using only data from grain-size analysis. CMECS uses both median grain size and percentage thresholds to define benthic substrate (FDGC 2012). Although particle-size analyzers can compute grain-size distributions with high precision, many are mechanically limited to analyzing substrates < 4 mm in diameter. Thus, data from benthic imagery analysis were used to augment the CMECS Substrate Component to provide information on coarse substrates > 4 mm in diameter. The benthic imagery data were expressed as percent cover values for sand, gravel, shell material, and BMA and equaled 100%. In most instances, it was impossible to determine what underlying substrate the benthic macroalgae were attached to. For this reason, the random points landing on BMA were removed from percent cover calculations (for the CMECS Substrate Component only). Furthermore, points landing on the shell were removed from percent cover calculations to ensure the benthic imagery data were directly comparable to the grain-size data. For example, sampling stations that had >30% relative cover of gravel were classified as “Sandy Gravel” and sampling stations that had >80% relative cover of gravel were classified as “Gravel”, in accordance with CMECS. The results were graphically displayed using ArcMap. The point data were interpolated into a raster layer using the function ‘Spline with Barriers’.

2.8. CMECS Biotic Component

The CMECS Biotic Component was classified using abundance data of benthic invertebrate infauna and epifauna collected from the biological replicate grab samples, as well as the cover of macroalgae collected from benthic imagery. Biotic communities were defined using hierarchical agglomerative clustering in PRIMER-e v.7 [31].

Two taxa drop rules, according to Gallagher and Grassle [50], were applied to these data. First, all generic taxonomic designations at taxonomic levels higher than family were dropped if there were more than two valid lower-level designations for that group. Second, all known meiofaunal taxa were dropped from analysis as the sampling methods used do not adequately capture taxa smaller than 1 mm [2,51]. Finally, because CMECS biotic communities and biotopes are characterized by dominant taxa, only those taxa contributing to the top 95% of the total observed abundance were included in the analyses. A Pearson correlation verified that the data matrices comprising 100% and 95% were correlated more than would be expected from chance alone (p = 0.998; p-value < 0.001), and the 95% abundance matrix was found to be representative of the dataset. These data were then fourth-root transformed to reduce the influence of numerically abundant taxa.

The taxa–station data matrix (95% most abundant taxa; fourth-root transformed) was converted into a resemblance matrix using the Bray–Curtis coefficient of similarity. To group together stations with similar community structures, hierarchical agglomerative clustering using group-average similarities was performed with the CLUSTER tool [31,52]. An unweighted pair group clustering method using an arithmetic mean (UPGMA) was employed. To test whether the clustering of stations that resulted was due to more than chance alone, a similarity profile routine (using the SIMPROF tool) was employed using 10,000 random permutations [52]. Next, a non-metric multi-dimensional scaling (nMDS) ordination was performed to visualize the clustering of stations [31,52]. Finally, a similarity percentage (SIMPER) tool was utilized to quantify the contribution of each taxon to the within-cluster similarity and identify typifying species within each cluster [31].

To incorporate benthic macroalgae into the classification of biotic communities, communities with ≥29.3% cover of BMA were classified as having a benthic macroalgal bed as a co-occurring element. Normally, when defining biotic communities from benthic imagery within the CMECS framework, the classification is based on dominance [2], but because benthic macroalgae was the only biotic category being measured with imagery, 29.3% cover was chosen as the threshold as it represents 2× the station average within the study area. Stations within clusters that had ≥29.3% cover of BMA were reviewed to distinguish which biotic group (e.g., filamentous, leathery/leafy, sheet algae, etc.) was representative of the algal bed.

2.9. Biotopes

Multivariate statistical analyses employed to quantitatively describe biotopes in the study area were conducted in PRIMER-e v.7. The two primary statistical tools utilized were distance-based redundancy analysis (dbRDA; McArdle and Anderson 2001 [53]) and constrained binary divisive clustering [31]. In PRIMER-e, dbRDA is performed with the DISTLM tool and is used to partition variation in biotic data, described by a Bray–Curtis resemblance matrix, according to a multiple regression model [54]. DISTLM quantifies the amount of variation in the benthic invertebrate data matrix that is described by each environmental predictor variable and builds sets of predictor variables that maximize the variation explained [54]. Constrained binary divisive clustering is performed with the LINKTREE tool in PRIMER-e and works by successively dividing the full set of biotic samples into new subgroups constrained by a specific threshold value of a particular predictor variable or set of predictor variables [31].

Data on environmental variables were normalized by subtracting the mean and dividing by the standard deviation for each variable separately [31]. The percent cover of BMA was included as an environmental predictor variable as it provides a biogenic habitat to the benthic environment. The benthic invertebrate data were fitted using depth, grain-size statistics, and benthic imagery data.

Highly collinear predictor variables (r = 0.95) were removed following the results of the iteratively run Pearson’s correlations in PRIMER-e. To test how related the environmental and biotic data were, a nonparametric similarity matrix relatedness routine (using the RELATE tool) was used with and without the inclusion of benthic imagery data to examine the quantitative impact these data have on the relatedness of the biotic and environmental data matrices. To examine which environmental predictor variable(s) optimized the relatedness between these data, the BEST tool was used on the biotic and environmental data matrices. BEST identifies the highest possible matrix correlation by iteratively searching through all possible combinations of environmental predictor variables [31]. However, it does not provide any information on the amount of variation explained by each predictor variable or the overlap in variation explained by sets of predictor variables [54]. Consequently, dbRDA was performed using the DISTLM tool in PRIMER-e to achieve quantitative partitioning of the multivariate variability that is explained by each predictor variable and use the amount of overlap in variation explained by sets of predictor variables to build a parsimonious model [54]. Burnham and Anderson [55] suggest using the small-sample-size-corrected Akaike Information Criterion (AICc) when the ratio of samples (n) to variables (v) is <40. After collinear variables were removed from the data matrix, the ratio of n/v ≈ 2, so AICc was used as the information criteria. To build a linear model using the best possible combination of predictor variables, DISTLM was performed for data matrices including and excluding benthic imagery to quantify the benefit of incorporating benthic imagery into distance-based redundancy analysis.

3. Results

3.1. Benthic Imagery

A total of 93 videos were captured, constituting 4.1 h of video footage. From these videos, a total of 138 still images were captured. Still images from sampling locations in which all the grab samples were unsuccessful and discarded (and the R/V Marindin had to be moved to an alternate sampling location) were excluded from the analysis. Still images of unsuccessful grab attempts were included in the analysis only if the grab attempt happened at the sampling location in which the successful sample was later collected. This ensured that still images used in the analysis were representative of the habitat directly surrounding the successful grab samples that were used for sediment and biological analyses. A total of 68 videos (2.77 h of footage) from successful sampling locations yielded 92 still images that were used in the analysis. A total of 25 videos (1.33 h of footage) from unsuccessful sampling locations yielded 46 still images, which were excluded from the analysis.

To detect changes in benthic cover as small as 1% (d = 0.01) with an alpha value (α) of 0.05 and a binomial probability (p) of 0.5, the sample size would need to consist of at least 9604 random points. This was tested by analyzing 5 sets of random points (5, 25, 50, 75, and 100) using 10 OPC reconnaissance images (3 replicates). The number of points necessary decreases as larger margins of error are allowed (

Table 1. To detect changes in benthic cover as small as 1% (d = 0.01) with an alpha value (α) of 0.05 and a binomial probability (p) of 0.5, the sample size would need to consist of at least 9604 random points. This was tested by analyzing 5 sets of random points (5, 25, 50, 75, and 100) using 10 OPC reconnaissance images (3 replicates). The number of points necessary decreases as larger margins of error are allowed (Table 1).

Margin of Error (%)	Number of Points Necessary to Achieve	Number of Points	Resulting Margin of Error
1	9604	5	43.8
5	384	25	19.6
10	96	50	13.9
15	43	75	11.3
		100	9.8

). Congruently, the binomial distribution was also used to estimate the margin of error (d) given different sampling sizes (n) (Figure 2). Using 50 random points was deemed adequate because (a) the margin of error was <15%, (b) the minimal increased precision (4%) was between 50 and 100 random points, (c) no statistically significant differences were observed in percent cover estimations for any substrate type between 50 and 100 random points, and (d) there was <1.5 min of analysis per image.

Processing benthic imagery for percent cover using 50 random points in Coral Point Count resulted in a mean of 70.9% cover of sand, 11.5% cover of gravel, 14.7% cover of BMA, and 2.9% cover of shell material within the study area.

3.2. Acoustic Surveys and CMECS Geoform Component

Acoustic surveys conducted on 23 October 2017 covered an area of 1.04 km² with a mean depth of 6.0 ± 0.9 m. Only Level 1 Geoforms were defined within the study area. A depth modifier was used in the classification dictionary to distinguish between intertidal and subtidal flats, as well as shallow and deep slopes [27]. “Subtidal flats” were dominant, comprising 90% of the study area, equivalent to 0.94 km², with the remaining 10% classified as “Banks” (Figure 3).

3.3. Sediment Characterization and the CMECS Substrate Component

The CMECS Substrate Component was first described using only data from grain-size analysis (GSA). The mean grain size per station ranged from 1.2–2.0 mm, median (D₅₀) grain size ranged from 1.2–2.1 mm, and mode grain size ranged from 1.2–2.4 mm. The percent of mud ranged from 0–0.0002%. According to Wentworth [40], this sediment is classified as very coarse sand. The percent of sand ranged from 46–90% and the percent of gravel ranged from 10–54%.

Based on the results from GSA, two CMECS Substrate Subgroups were present in the study area: “Gravelly Sand” (at 22 stations) and “Sandy Gravel” (at 2 stations). “Gravelly Sand” is defined in CMECS as containing < 30% gravel (>2 mm), with sand (0.625–2 mm) comprising > 90% of the remaining sand–mud mix. “Sandy Gravel” is defined in CMECS as containing 30–80% gravel, with sand comprising > 90% of the remaining sand–mud mix.

Benthic cover data from benthic imagery analysis (BIA) were used to classify grain size and augment CMECS substrate classifications. Sampling stations 14, 15, 18, and 23 had >30% mean cover of gravel and were classified as “Sandy Gravel” instead of “Gravelly Sand”. Sampling station 16 had >80% cover of gravel and was classified as “Gravel” instead of “Gravelly Sand”. The biggest difference between GSA and BIA was observed at station 16. The percent of gravel according to GSA was 14.1%, while the mean percent cover of gravel according to BIA was 98% (Figure 4).

Augmenting GSA with BIA resulted in one additional substrate subgroup classification (“Gravel”) and the reclassification of 25% of the sampling stations. “Gravelly Sand” remained the dominant substrate classification, comprising 17 of 24 sampling stations. “Sandy Gravel” was present at 6 of 24 sampling stations, and “Gravel” was present at 1 of 24 sampling stations. The use of benthic imagery improved the accuracy of CMECS substrate classifications within the study area.

3.4. Benthic Invertebrate Sampling and the CMECS Biotic Component

In total, 82 taxa were identified from 48 biological replicate samples. After the drop rules were applied, 72 taxa remained. The 95% most abundant taxa represented 17 different families and 30,671 individuals. The hierarchal agglomerative clustering resulted in a hierarchal structuring of sampling stations, from which two clusters (a and b) were identified as being due to more than chance alone (p = 0.01; 100,000 permutations).

Cluster a resulted in an average group Bray–Curtis similarity of 76.8% and was characterized by consistently high abundances of small burrowing polychaetes (Syllidae and Dorvilleidae spp.), burrowing clams (Myidae, Mactricidae, and Tellinidae spp.), and small burrowing sea cucumbers (Synaptidae spp.). The group-average percent cover of benthic macroalgae was 4% in cluster a. Cluster b resulted in an average group Bray–Curtis similarity of 71.1% and was characterized by consistently high abundances of burrowing fringe worms (Cirratulidae spp.), mobile gastropods (Calyptraeidae and Columbellidae spp.), and lesser abundances of small burrowing polychaetes (Syllidae and Dorvilleidae spp.). The group-average percent cover of benthic macroalgae was 32.5% in cluster b.

These results were used to classify the clusters in terms of CMECS biotic communities. Cluster a was classified as a “small surface-burrowing polychaete bed” with the co-occurring element “clam bed” and comprised 15 of 24 sampling stations. Cluster b was classified as “Cirratulid bed” with the co-occurring elements “Calyptraeid bed” and “filamentous algal bed” and comprised 9 of 24 sampling stations.

3.5. Biotopes

The identification of highly collinear predictor variables resulted in the removal of four variables: % sand, % gravel, mean grain size, and % cover of sand. The remaining environmental variables included depth, % organic matter (by sample weight), median grain size (D₅₀), mode grain size, sediment sorting, sediment skewness, sediment kurtosis, mass of coarse gravels retained from biological replicates (‘>4 mm’), % cover of gravel, % cover of BMA, and % cover of shell material.

Multivariate statistical analyses were conducted on two datasets, one including benthic imagery data and one excluding benthic imagery data. When benthic imagery was included, the matrix relatedness routine (RELATE) resulted in a Spearman’s rank correlation coefficient of ρ = 0.46 (p < 0.001) between the invertebrate similarity matrix and environmental distance matrix, as opposed to ρ = 0.34 (p < 0.001) when benthic imagery was excluded. Additionally, the matrix relatedness optimization routine (BEST) resulted in % cover of BMA maximizing Spearman’s rank correlation coefficient at a value of ρ = 0.67 (p = 0.001). When benthic imagery variables were removed from the environmental data matrix, the set of depth and % organic matter resulted in the highest Spearman’s rank correlation coefficient (ρ = 0.45, p = 0.002)

Nonparametric testing shows that benthic imagery predictor variables, particularly % cover of BMA, help to explain the invertebrate community structure within the study area. To quantify the amount of variation explained by each environmental predictor variable, distance-based redundancy analysis (dbRDA) was conducted using the DISTLM tool.

The sequential tests, in which sets of variables were built to minimize AICc values, resulted in the identification of depth, % organic matter, and % cover of BMA as the best overall model to fit the environmental data cloud (AICc = 142.4; R² = 0.59). When benthic imagery variables were excluded from dbRDA, the best overall model was the set of depth and % organic matter (AICc = 147.7; R² = 0.43). According to rules of thumb described by Bolker [56], including benthic imagery variables in dbRDA produced a distinguishably better model: 59% vs. 43% explanatory power. However, to ensure the effectiveness of the distance-based linear models, the most parsimonious models were chosen, which included only one predictor variable each: % BMA cover (explaining 45% of the variation for the dataset including benthic imagery) and % organic matter (explaining 32% of the variation for the dataset excluding benthic imagery).

The constrained binary divisive clustering routine resulted in five significant clusters (or biotopes) of sampling stations (p < 0.05) for model 1 (including benthic imagery) and four significant clusters (or biotopes) of sampling stations (p < 0.001) for model 2 (excluding benthic imagery). The SIMPER tool subsequently revealed the contributions of individual taxa to within-cluster similarities and between-cluster dissimilarities and allowed the clusters derived from LINKTREE to be described as biotopes (

Table 2. Distance-based linear model 1 (including benthic imagery) resulted in five biotopes. The most important predictor variable was calculated to be % cover of benthic macroalgae (BMA) derived from benthic imagery.

Model 1 Biotope	Stations	BMA Cover (%)	CMECS Biotope Description
A	4, 11, 21	0	Small Surface-Burrowing Fauna with co-occurring element Clam Bed found in predominantly sandy substrate with 0% SAV
B	2, 3, 5, 6, 7, 17, 22	1–2	Small Surface-Burrowing Fauna with co-occurring element Clam Bed found in predominantly sandy substrate with 1–2% cover of SAV
C	8, 19, 20, 24	3–10	Small Surface-Burrowing Fauna with co-occurring element Holothurian Bed in predominantly sandy substrate with 3–10% cover of SAV
D	1, 9, 12, 15, 18	11–30	Small Surface-Burrowing Fauna with co-occurring element Larger Deep-Burrowing Fauna found in mixed substrate with 12–30% cover of SAV
E	10, 13, 14, 16, 23	>31	Mobile Molluscs on Hard or Mixed Substrates with co-occurring elements Larger Deep-Burrowing Fauna and Mobile Crustaceans on Hard or Mixed Substrates found in coarse substrates with little sand and >30% cover of SAV

and

Table 3. Distance-based linear model 2 (excluding benthic imagery) resulted in four biotopes. The most important predictor variable was calculated to be % organic matter derived from sediment samples.

Model 2 Biotope	Stations	% Organic Matter	CMECS Biotope Description
A	3, 4, 5, 11, 17, 20	<0.43	Small Surface-Burrowing Fauna with co-occurring element Clam Bed found in substrate with <0.43% organic matter
B	1, 2, 7, 18	0.44–0.47	Small Surface-Burrowing Fauna with co-occurring element Clam Bed found in substrate with 0.44–0.47% organic matter
C	6, 8, 21	0.51–0.52	Small Surface-Burrowing Fauna found in substrate with 0.51–0.52% organic matter
D	9, 10, 12, 13, 14, 15, 16, 19, 22, 23, 24	>0.56	Larger Deep-Burrowing Fauna with co-occurring elements Mobile Molluscs on Hard or Mixed Substrates and Small Surface-Burrowing Fauna in substrate with >0.56% organic matter.

and Figure 5).

4. Discussion

The second-most important surrogate for benthic communities after water depth is substrate/sediment type [1]. The Van Veen sampler used in this study is well-suited to collecting sediment samples in nearshore sandy coastal areas and is often used in studies to characterize sediment [28]. However, one pebble-sized or larger grain can prevent the sampler from closing, thus a bias away from gravel-sized material may exist when using these samplers. At several sampling stations, the Van Veen returned to the surface having not fully closed due to a single pebble-sized sediment grain. The sampler was reset and returned to the seafloor and the failure was repeated. After several attempts, the anchored boat was moved and the pattern repeated until the boat moved out of the area with high percentages of gravel. This had the effect of skewing the results toward sand-sized material as can be seen via the comparison of the sediment type maps based on GSA and GSA augmented by BIA (Figure 4).

We believe investigators should make every effort to add a low-cost video camera to the sampler to ground-truth the sediment type at every drop location whether or not a sample is retrieved. In fact, it may be critically important to incorporate the video when the sampler fails to document that sediment type in order to properly characterize the sediment type in areas where gravel-sized material is present.

It was determined that five stations (14, 15, 16, 18, and 23) were misclassified using only the GSA from the collected samples. The GSA for this project was limited to <4 mm, but the critical 2–4 mm pebble-sized material, classified as gravel in the Wentworth scale, was reduced due to the sampling bias caused by the difficulty of retrieving complete samples in areas with significant percentages of gravel-sized material.

A tradeoff is evident between GSA and BIA in terms of classifying substrate using the CMECS framework. GSA by means of a particle-size analyzer is more precise than estimating percent cover using a random-point overlay. However, a relatively small subsample (100 mL) is taken from the sampler for GSA and all information on substrates > 4 mm is lost. In contrast, percent cover values derived from BIA are averaged over multiple images from a single sampling station, which provides a more complete description of the surficial substrate features present at each sampling station. However, measuring grain-size statistics (e.g., mean, median, mode, sorting, skewness, etc.) from images is currently impractical. In these ways, the two methodologies complement each other when used together.

The classification and mapping of benthic habitats and biotopes is an important first step in ecosystem-based management [57] and an integral part of marine spatial planning [1]. As both the number of marine benthic habitats that are classified and the technologies used to classify them grows, so will the importance of using a standardized system of classification, such as the Coastal and Marine Ecological Classification Standard (CMECS). Likewise, it is imperative that benthic habitats are classified and mapped accurately and that researchers take advantage of any widely available and inexpensive tools that can help them do so. This study attempted to quantitatively describe the benefit of utilizing low-cost benthic imagery (a GoPro Hero 5™ attached to a Young-modified Van Veen grab sampler) to augment the classification and mapping of benthic habitats and biotopes with CMECS. Within the study area, benthic imagery helped improve the accuracy of the CMECS Substrate Component classification by providing information on substrate > 4 mm in diameter, information that is lost using traditional grain-size analysis with a particle-size analyzer. The use of benthic imagery also improved the classification of the CMECS Biotic Component by providing information on benthic macroalgae, which are not captured in traditional benthic grab sampling. Furthermore, benthic imagery variables (namely, % cover of BMA) quantitatively improved the classification and mapping of biotopes, as well as the ecological relevance of the biotopes.

When determining the optimal number of random points to use in benthic cover analysis, a tradeoff was evident between increasing the sample size of random points and the effort required to process the images. Using 50 points was deemed sufficient, as only large differences in benthic cover were important for the classification and mapping of benthic habitats and biotopes, and the resulting margin of error < 13% was adequate. Researchers should consider the desired margin of error, the type and frequency of substrates in the study area, and the amount of time they are willing to dedicate to image processing when deciding on the number of points to use in benthic cover estimation. A caveat of using random point overlay is that the benthic cover will always equal exactly 100%, even when benthic cover appears layered e.g., macroalgae attached to gravel, thus potentially underrepresenting the lowest layer. This caveat will likely permeate any approach to estimating benthic cover from plan-view or even oblique images and researchers should pool these substrate categories when attempting to identify areas with coarse substrates.

4.1. CMECS Substrate Classification

The use of benthic imagery improved the classification of the CMECS Substrate Component by providing information on the presence or absence of substrates > 4 mm in diameter surrounding each sampling station. Substrates > 4 mm in diameter were either coarse gravel or shells and large shell fragments, all of which provide structure and complexity to the benthic habitat as well as a potential surface for attachment for benthic macroalgae and sessile or slow-moving benthic invertebrates. Because the study area is primarily composed of subtidal sand flats, the presence of such habitats provides a source of spatial heterogeneity and complexity to the seafloor and is of particular importance to the mapping of habitats and biotopes within the study area. However, Van Veen-style grab samplers are biased toward fine-grain sediments. Without benthic imagery, information on the presence or absence of coarse substrates within the study area would have been inaccurate. Because the number of benthic macroalgae was the best environmental predictor variable for explaining the benthic invertebrate community structure, it was clear that coarse gravel resting on the surface of the seafloor had a greater ecological impact on habitats and biotopes within the study area than coarse gravel buried in the sediment.

Benthic imagery helped improve the classification of the substrate within the study area by augmenting but not replacing grain-size statistics. Studies such as that by Harris and Baker [1] have classified surficial substrate without taking physical samples but had to estimate the median grain size by using typical values for different types of substrates. However, classifying substrate using benthic imagery techniques alone will likely result in lower-resolution classifications. The complementary approach of using both benthic imagery and traditional grain-size analysis provides precise sediment statistics and more broad-scale information on the surrounding habitat and coarse substrate features.

4.2. CMECS Biotic Classification

Benthic imagery improved the classification of the CMECS Biotic Component by providing information on benthic macroalgae, an important biogenic habitat feature. Benthic macroalgal communities provide shelter, shade, and detrital material to the benthic habitat, which is important to and can shape the structure of associated fauna [2]. Without clear guidance from CMECS on what percent of benthic cover of benthic macroalgae constitutes a co-occurring macroalgal bed, selecting a threshold value based on the mean value (e.g., 2× the average % cover) of macroalgae within the study area was deemed most appropriate and is likely a good rule of thumb for other researchers who run into this issue.

4.3. Biotopes

It was clear from the matrix relatedness routine, relatedness optimization routine, and distance-based redundancy analysis that benthic imagery variables were not only important for making biological-environmental linkages but performed better than depth and grain-size statistics at doing so. This is likely the result of sampling within a small, homogeneous study site.

When benthic imagery was included in the environmental data matrix, the most parsimonious model included % cover of SAV, which explained 45% of the variation in the invertebrate data. When benthic imagery was excluded, the most parsimonious model was % organic matter, which explained only 32% of the variation in the invertebrate data. However, the model including % organic matter was only a slightly better model than depth, which explained 31% of the variation in the invertebrate data. Not only did the model including % cover of SAV explain more variation in the invertebrate data than either model that resulted from the exclusion of benthic imagery but it also had an AICc score of 4–7 less than either depth or % organic matter, which, according to Bolker (2006), means that it is clearly distinguishable as a better model. It is clear from these analyses that statistically better models were produced when benthic imagery was included in the environmental data.

It is known that the amount of organic matter in sediment can have ecological implications for the structure of benthic invertebrate communities [58]. Benthic macroalgae are ecologically important as they provide shelter and shade for benthic invertebrates [2] and can serve as an important nursery habitat for many benthic species [59]. However, within the study area, organic matter made up a small proportion of sediment sample weights and varied little from station to station (ranging from 0.2–1.1%). Thus, while statistically significant, biotope model 2, based on organic matter content in the sediment (Table 3), lacks real-world applications.

Because organic matter and benthic macroalgae were both correlated with coarse substrates, while organic matter is usually associated with fine substrates (Gallagher, pers. comm.), it is likely that benthic macroalgae is a large source of sedimentary organic matter within the study area. Because benthic macroalgae require coarse substrate for attachment, the % cover of BMA acted as a proxy for the presence of coarse substrates and the absence of fine substrates (and vice versa). Cobble and boulder reefs provide three-dimensional structural complexity to the seafloor and are known to host diverse associated fauna [59,60,61,62,63]. Cobble and boulder habitats are also known as nursery grounds for many benthic species [59].

Despite its known utility, benthic imagery has only been incorporated into CMECS benthic habitat classifications in a handful of published studies [21,22,23,24,25,26,49]. No study to date has attempted to quantify the benefit of using imagery to describe benthic habitats within the CMECS framework.

5. Conclusions

Investigators should add a low-cost video camera to the sampler to ground-truth the substrate/sediment type at every drop location whether or not a sample is retrieved. In fact, it was shown here at locations where a sample was not received that the video data were critical to correctly characterize substrate/sediment type.

Including benthic imagery improves habitat mapping within the CMECS framework, thus resulting in more interpretable and meaningful classifications and map products. Benthic imagery can help identify marine debris (e.g., derelict fishing gear, trash, etc.), SAV, bedforms, and other benthic habitat features that may not be identified with traditional grain-size analysis but are important to the communities of invertebrates that inhabit them as well as other ecosystem services. Thus, the methodologies presented in this study should be adapted and improved by researchers working to characterize bottom types.

Benthic habitat classification and mapping play important roles in ecosystem-based management and ecosystem monitoring, and this study illustrates that the use of benthic imagery increases the accuracy of classification and mapping of substrate/sediment types and benthic habitats. The adoption of these methodologies could have a positive impact on the management of coastal marine ecosystems.