How Do Continuous High-Resolution Models of Patchy Seabed Habitats Enhance Classification Schemes?

Abstract

Predefined classification schemes and fixed geographic scales are often used to simplify and cost-effectively map the spatial complexity of nature. These simplifications can however limit the usefulness of the mapping effort for users who need information across a different range of thematic and spatial resolutions. We demonstrate how substrate and biological information from point samples and photos, combined with continuous multibeam data, can be modeled to predictively map percentage cover conforming with multiple existing classification schemes (i.e., HELCOM HUB; Natura 2000), while also providing high-resolution (5 m) maps of individual substrate and biological components across a 1344 km² offshore bank in the Baltic Sea. Data for substrate and epibenthic organisms were obtained from high-resolution photo mosaics, sediment grab samples, legacy data and expert annotations. Environmental variables included pixel and object based metrics at multiple scales (0.5 m–2 km), which improved the accuracy of models. We found that using Boosted Regression Trees (BRTs) to predict continuous models of substrate and biological components provided additional detail for each component without losing accuracy in the classified maps, compared with a thematic model. Results demonstrate the sensitivity of habitat maps to the effects of spatial and thematic resolution and the importance of high-resolution maps to management applications.

1. Introduction

Classification schemes are common tools to support spatial decision-making. In marine spatial planning and conservation, they help to inform and prioritize management actions including design of monitoring programs, risk assessments, placement of marine protected areas, etc. [1,2]. Furthermore, classification schemes can provide advances in transnational efforts towards common strategies for nature conservation and protection (see Strong et al. [3] for a review and comparison of several classification schemes). Building a classification scheme includes important decisions on spatial and thematic resolutions and the way to represent structures [4]. As exemplified in this paper, it is critical for mapmakers to understand and communicate the implications of these decisions in order to inform management actions and evaluations.

At the European level, the European Union Nature Information System (EUNIS) [5], the Baltic Marine Environment Protection Commission Underwater Biotope and Habitat Classification System [6] (HELCOM HUB, hereafter referred to as HUB) and Natura 2000 [7] are common classification schemes used to report against the European Marine Strategy Framework Directive (MSFD) [8] and the Habitat Directive [9]. The intended use of these schemes is to classify marine areas through several steps. Classification is first done according to the environmental setting (e.g., vertical zone), then by the geological features (soft/hard substratum), and finally by the biological components (e.g., communities, dominant species).

Similarly, in Sweden, classification schemes that combine both geological and biological information are required to support ongoing marine spatial planning initiatives (national and regional), design of marine reserves and other ecosystem-based fisheries management actions. In particular, the Geological Survey of Sweden (SGU) and the Swedish Agency for Marine and Water Management (SwAM) cooperate to provide detailed HUB and Natura 2000 maps in priority areas in the Baltic Sea. Ideally, these products should be backward compatible with SGU legacy maps, and include percent coverage of substrate and benthic species used in the national marine spatial planning cumulative impact assessment tool Symphony [10]. The methods presented in this paper are grounded in the need to develop such multiuse products.

The HUB classification (

Table 1. The Baltic Marine Environment Protection Commission Underwater Biotope and Habitat Classification System [6] (HELCOM HUB, hereafter referred to as HUB) split rule summary. Examples of HUB level 1–6 biotopes are found in the in the Supplementary Materials (Legend S1, Table S3).

Level	Description
1 Baltic (Letter)	Baltic
2 Vertical Zone (Letter)	Photic Aphotic
3 Substrate (Letter)	Coverage of a specified substrate type ≥90% Coverage <90%, Mixed
4 Community Structure (Number)	Coverage of Macroscopic vegetation or sessile macroscopic epifauna ≥10% Coverage > 0% < 10%, Sparse Coverage = 0%, No vegetation or macro fauna present
5 Characteristic Community (Letter)	Coverage of a specified taxonomic group ≥10% Coverage ≥ 10% but not of a specified taxonomic group, Mixed community Coverage = 0%, No macroscopic community
6 Dominating Taxa (Number)	Biomass/bio volume of some specified taxa ≥50%

), a relatively new classification scheme with few existing maps [11], has been developed to be compatible with EUNIS and is specifically designed to create a common understanding of the Baltic Sea habitats, biotopes and communities [6]. HUB is a hierarchical classification scheme that consists of six levels with biotopes defined by applying split rules similar to a dichotomous key. The first three levels describe habitats, whilst levels 4–6 describe biotopes. HUB codes contain a maximum of six characters with each character referring to a specific level where a split rule has been applied. Biological communities are characterized by HUB Level 5, which is most comparable to EUNIS Level 4. The spatial scale for HUB biotopes is not strictly defined. The guidance given in the technical report is that a biotope patch should be large enough to function as a biotope [6], implying that the community is a distinct unit with some distinct functions. For instance, the EUSeaMap project has used a practical application of a 5 × 5 m scale as a minimum patch size [12].

While the efforts at European level are a first attempt towards establishing a network of transnational nature conservation and protection areas, providing a platform to better cope with the challenges posed by climate change [13], there are a number of limitations of such broad scale classifications. For example, as noted by Galparsoro et al., [1] the EUNIS classification needs to include new habitat classes and develop the classifications of less well-represented areas (e.g., deep-sea areas). Regardless of how refined such classification schemes become, they will remain a simplification of nature, for example by neglecting transitional areas, reducing local complexities (and neglecting some important feature at a local scale) and creating artefacts [14].

This paper focuses on the advantages of using high-resolution continuous models of substrate and biota based on multibeam data, sediment samples and photos, to overcome some of the shortcomings in existing classification schemes. A recent study in the Baltic have looked at predicting percent coverage of substrate and biota using similar data [15], and found the method to be useful, but does not use the prediction to also provide thematic habitat maps. We present a novel approach to prepare and predict high resolution habitat maps comparing both a thematic model workflow with predefined HUB and Natura 2000 classes, and a continuous model workflow that allows classified maps to be created for multiple scales and definitions, as well as providing percent coverage maps of substrate and biological components for more detailed analysis. In addition, we estimate the importance of different predictor variables; for instance, the performance of the models using backscatter data versus a variety of metrics (pixel and object based) from high-resolution terrain models. Such information is quite relevant in Sweden where backscatter is not yet provided as a standard for multibeam surveys. We also investigate if the substrate models can be improved by supplementing the training data with legacy observations and expert annotations, and if the resulting substrate maps can help improve the biological models. Finally, we look closer at some of the end-user implications of using thematic and continuous maps at different spatial scales. Throughout the paper, we refer to continuous models as the result of the percent cover of each thematic class (for example percent cover of the substrate class sand).

2. Materials and Methods

The analytical workflow follows a logical sequence of steps: (1) data acquisition and processing; (2) processing of environmental predictors; (3) development of thematic and continuous models; (4) combining models into classified maps; and (5) map accuracy assessments.

2.1. Acquisition and Processing

2.1.1. Study Area

The Baltic Sea offshore area Hoburgs Bank (Figure 1) was selected as a pilot project for high resolution mapping in Sweden. The survey covered a 1344 km² area of the bank (10–60 m depth) with the seafloor containing a complex mix of substrates: mainly sand, gravel, hard clay, rock and boulder, shaped by glacial processes, waves and currents. Hoburgs Bank is part of a newly expanded Natura 2000 marine protected area, motivated by the population of seabirds and recent findings on the threatened Baltic Sea harbor porpoise [16]. The most apparent human-caused threat to the area is a major shipping lane that crosses the bank’s northwestern part, and a second shipping lane that runs southeast of the bank.

2.1.2. Hydrographic Survey

In autumn 2016 and spring/summer 2017, we conducted a full-coverage hydrographic survey of the Hoburgs Bank study area using SGU’s vessel S/V Ocean Surveyor that was equipped with a multibeam EM2040D (Kongsberg Maritime AS, Kongsberg) and a sub-bottom profiler (SBP) (Echoes 3500 T3, iXBlue, Natick). Hydrographic data collection was generally conducted from late afternoon to early morning, with sampling operations occurring during the daytime. We used a moving vessel profiler to collect sound speed data, and daily CTD casts to update the absorption coefficient for multibeam backscatter data in the acquisition software Seafloor Information System (Kongsberg Maritime AS, Kongsberg). Bathymetric data from the EM2040D multibeam were processed in CARIS HIPS and SIPS (v 10.3.1, Teledyne CARIS, Fredericton) using the Combined Uncertainty and Bathymetric Estimator (CUBE) surface generation method [17]. Seafloor backscatter data were processed by implementing the geocoder engine [18] in the Fledermaus Geocoder Toolbox (FMGT) (v7.7.6, QPS, Zeist). Both depth and backscatter were initially exported at 2-m resolution for use in on-board sampling design. Final post processing depth grids and backscatter mosaics were exported at 0.5-m resolution (Figure 2), together with depth uncertainty, survey track lines, as well as all angular range analysis (ARA) backscatter statistics available in FMGT. More details on the acquisition are available in SGU’s survey report [19].

2.1.3. Sampling and Underwater Observations

Stratified random sampling was used to select the locations for sediment grab samples and underwater observations (using a drop-camera) via the ArcGIS 10 Sampling Design Tool [20], based on (1) depth intervals, (2) substrate types (using backscatter intervals) and (3) complexity intervals (a combination of depth standard deviation and bathymetric position index). The classified depth, substrate and complexity raster were combined into polygons, and slivers were removed. The sampling design tool was used to randomly choose at least one of each thematically unique polygon found within a ~ 3 × 12 km area (approximately what we could sample and survey during 24 h in optimal conditions), with the samples placed randomly within each polygon. The process was repeated for a total of 34 blocks that divided the survey area (see Supplementary materials Map S1). We visited 559 locations for the drop-camera, which also included sediment samples at 434 of the locations (some boulder sites were excluded from sediment sampling). Distribution of the drop-camera samples divided into HUB 3 classes and HUB 4–6 classes are available in Appendix A

Table A4. Distribution of the ground-truth dataset divided into HUB 3 classes.

HUB 3	UV obs. 1	Legacy 2	Expert 3	Total
Rock and boulder	67	34	16	117
Hard clay	5	0	83	88
Coarse sediment	72	89	138	299
Sand	195	159	252	606
Mixed	220	167	61	448
Total	559	449	550	1558

¹ Drop camera samples (2016–2017), testing (30%) and training (70%); ² Drop camera, video transects, sediment samples (2005), training only; ³ Expert annotation (based on depth and backscatter), training only.

and

Table A5. Distribution of the ground-truth dataset divided into HUB 4–6 classes.

HUB Level 4–6	UV obs. 1
Characterised by perennial algae	73
Dominated by perennial filamentous algae	94
Characterised by epibenthic bivalves	42
Dominated by Mytilus spp.	59
Characterised by cnidarians	25
Characterised by annual algae	2
Mixed epibenthic community	16
Sparse epibenthic community	79
No epibenthic community	169
Total	559

¹ Drop camera samples (2016–2017), testing (30%) and training (70%).

.

Positioning was collected via a real-time kinematic (RTK) gps (Seapath 330 GNSS-RTK, Kongsberg Maritime AS, Kongsberg) with corrections from SWEPOS base stations, and adjusted to the cable breakpoint on the moonpool (grab-sampler) and A-frame winches (drop-camera), located in the middle and rear of the vessel respectively, about 15-m apart. A dynamic positioning system retained the vessel’s position over the site.

While the position uncertainty of the grab samples and drop-camera varied with depth and currents, it was generally within ±2 m when distinct features such as clay reefs spurs were visited. Fine sediment (from silt to gravel) was sampled with a Van Veen grab (0.1 m²), whereas coarser sediment and hard substrates were sampled with an Orange Peel Bucket (~0.5 m²). Onboard geologists interpreted all samples (origin, composition), and submitted a representative subset (n = 117) of fine sediment types (silt to gravel) from surface sediments (~0–5 cm) to sieve grain size analysis.

For the underwater observations, we used a custom-built drop-camera system with a 360°rotating and 90° tilting full frame Canon EOS 6D DSLR camera, ~75 cm above the seafloor. In order to capture images in all directions, we applied an automated photo and light protocol through which we were able to capture 100% sub-millimeter resolution coverage of ~15m² of the seafloor (see Figure 3 for example) and additional images of the surrounding seascape. We also used a GoPro Hero 4 camera to aid the photo interpretation with video, and two parallel red lasers placed 30 cm apart to provide a scale reference.

We analyzed the images using a point intercept method in order to obtain 100 points per site of substrate and biological coverage, as well as presence absence data. Presence of ripples, shells, crawl tracks, and fish were also noted. We relied on sediment samples, video imagery, as well as high-resolution multibeam data to better annotate images when substrate components were covered with benthic organisms. The point intercept annotations could not differentiate between finer grain-sizes than the difference between gravel and sand; hence the sand class from the photomosaic also included finer sediment fractions when present (

Table 2. Grain size ranges for the surface substrate compositions, and their sampling methods.

Size Range (mm)	Hard/Soft	Substrate	Grab-Sampler	Drop-Camera
		Bedrock 1		Point intercept
>600	Hard bottom	Large boulders		Point intercept
200–600		Boulders		Point intercept
60–200		Large stones		Point intercept
20–60		Pebbles & stones		Point intercept
2–20		Gravel		Point intercept
0.6–2	Soft bottom	Coarse sand	Sieve analysis	Point intercept 2
0.2–0.6		Medium sand	Sieve analysis
0.006–0.2		Fine sand	Sieve analysis
0.002–0.06		Silt	Sieve analysis	Point intercept 2
<0.002		Soft clay	Sieve analysis	Point intercept 2
<0.002	Hard bottom	Hard clay		Point intercept

¹ not found in the study area.² clay, silt and sand fractions were classified according to the dominant grain size (always sand in the study area).

). More detailed information on fine grain sizes were obtained from the sieved surface sediment samples, which was a separate dataset (

Table 3. List of sources of training and testing data according to purpose of application. Table A4 and Table A5 in Appendix A show the distribution of HUB classes within the data.

Data Source	N	Method	Target
Drop camera samples (2016–2017) 1	559	Point intercept	Percent coverage of benthic organisms.
Drop camera samples (2016–2017) 1	559	Point intercept	Percent coverage of substrate fractions (i.e., sand, gravel, pebbles, stones, boulders, large boulders and hard clay)
Legacy data (drop camera, video transects and sediment samples from 2005) 2	449	Estimated coverage	Percent coverage of substrate fractions.
Expert annotation (depth, backscatter) 2	550	Estimated coverage	Percent coverage of substrate fractions.
Grab samples (2016–2017) 2	117	Sieve analysis	Percent coverage of fine substrate fractions (i.e., soft clay, silt, fine - coarse sand)
Grab samples (2016–2017) 2	434	Expert	Thematic substrate classes (i.e., silty gravelly sand)

¹ Testing (30%) and training (70%), split using stratified random sampling based on HUB level 1–6 classes. ² Training data only.

). While we did not conduct a statistical uncertainty estimate of the image analysis process, a second expert double-checked all interpreted sites in order to minimize human error.

2.1.4. Legacy Data and Expert Annotations

In order to further improve the training dataset for the substrate models we used lower quality legacy data and expert annotations. Data collected during a survey in 2005 by SGU using drop-camera, samples and video transects [21] were processed and re-annotated with substrate information (449 sites). ArcGIS (ArcGIS Desktop v10.5.1, ArcGIS Pro v2.1.2) was used to display and interpret the samples/photos and video transects together with high-resolution backscatter and multibeam data in order to verify, and if needed modify, the location of some sites. For areas with a small number of samples, under-represented habitats and obvious artifacts in early model outputs, we also conducted additional expert annotations (550 sites) of substrate components based on the high-resolution multibeam data and available ground truthing. Distributions of the legacy and expert annotations divided into HUB 3 classes are available in Appendix A Table A4.

2.2. Environmental Setting

Multibeam bathymetry and backscatter were the backbone of the environmental variables that we developed (hereafter referred to as predictors), and we used both pixel metrics applied across the seascape at multiple spatial scales and OBIA to make the most of this information. We also included in situ oceanographic data collected during the survey (salinity, temperature, and current speed and direction near the seafloor), as well as interpreted SBP data (postglacial sand deposits). In addition, we included geographical coordinates (northing/easting), which have been found to account for autocorrelation in the model [22]. All predictors were resampled to the modeling resolution of 5 m, which roughly corresponded with the scale of the underwater observations ~15 m² while also providing a practical minimum mapping unit for the end users (See Supplementary materials Tables S1 and S2 for a complete list of the predictor variables).

Bathymetric and backscatter information were recalculated to include multiple scale resolutions relevant for the model building (0.5 m, 1 m, 2.5 m, 5 m, 20 m, 50 m, 100 m, 200 m, 500 m, 1 km and 2 km), hereafter referred to as multiscale metrics. To represent scales smaller than 5 m, we first developed terrain metrics or morphometrics [23] in ArcGIS (standard deviation, profile, planform, and standard curvature, slope, slope of the slope (a measure of the magnitude of slope change), terrain-surface roughness and surface area to planar area) from the high resolution grids (0.5 m, 1 m, 2.5 m) then aggregated the information to 5 m using minimum, maximum, range and mean values. The same metrics were also run using 5 m bathymetry and backscatter (with the addition of median value and bathymetric position index, BPI). To represent scales larger than 5 m, we generated a set of smoothed depth grids at 5-m resolution using a Gaussian low-pass filter to remove details at different neighborhoods (i.e., 20 m to 2 km), followed by calculations of selected terrain metrics (i.e., slope, slope of slope, BPI and slope direction). In order to adjust for potential angle artifacts, uncertainty data retrieved from multibeam bathymetry as well as the Euclidean distance to the survey routes were included as predictors.

In addition to the pixel based metrics, we used object based image analysis (OBIA) in ENVI Feature Extraction Module (v 5.4, Harris Geospatial Solutions, Broomfield) to generate segments with statistics from depth and backscatter data, a process similar to previous semi-automated coral reef habitat mapping work [24,25]. Segments were developed from a composite image (principal component compressed depth metrics in 5 bands and backscatter in one) at 2.5-m resolution. ENVI segmentation thresholds were set to segmentation 11 and merge 91, using an edge detection algorithm. The resulting segment statistics (shape, textural and spectral attributes, [26] were exported as raster images and aggregated to the modeling resolution of 5 m.

We also applied another alternative way for treating habitat features as objects by calculating the Euclidean distance to features (often consisting of moraine ridges) seen in fine scale BPI (inner radius 5 m, outer radius 25 m, BPI ≥ 2), as well as distance to fine-medium sand patches (using a threshold reclassified backscatter mosaic).

In order to reduce collinearity and the amount of predictors, we grouped similar predictors and used principal component analysis (PCA) to extract the components that contributed to >5% of the total variance within each group (see supplementary Script S1). Selected predictor variables, such as depth, where excluded from the PCA process to better understand their importance in the models. Furthermore, expert evaluation in the model fitting stage helped in reducing the total number of predictors from 184 to 41.

2.3. Thematic and Continuous Models

One of the goals with the analysis was to compare thematic models (i.e., HUB 3–6, Natura 2000) and thematic maps classified from continuous models (hereafter referred to as reclassified maps), in this case we always used the same testing and training data, tuning grid and predictor variables.

For the general data preparation, we split the drop-camera data from 2016–2017 into 30% testing data (n = 154) and 70% training data (n = 405), using stratified random sampling based on HUB level 1–6 classes (reclassified from percent cover of substrate and biota). The same testing data was used for all models (substrate and biota), except models using sediment grab samples which used all data for training due to the low number of sieve-analyzed samples (n = 117). See Table 3 for an overview of the data and the targets they were used for. One of the strengths of our approach was the possibility to integrate different data sources, such as results of grain size analysis from samples (which captured fine grain size fractions not visible in photos) and the percent coverage from images, (which captured the distribution of coarser grain sizes, and the total coverage of fine grain sizes).

For the thematic models, we first reclassified the testing and training data into thematic classes for HUB and Natura 2000, before the models where run for each level (HUB level 3, HUB level 4–6, and Natura 2000 reef/sandbanks). For the continuous models, we prepared the training data by assigning 0.1% coverage to presence with <1% cover and 0.001% coverage to absence, followed by logit transformation (all individual models). We also evaluated continuous compositional models. The modeling framework of the latter approach explicitly takes into account that substrate components must sum up to 100 % at each location [27].

Model algorithms were run with R v3.4.1 [28] (see supplementary Script S2–S5). For both thematic and continuous models, we used boosted regression trees (BRTs) implemented by the gbm [29] and caret package [30]. BRTs is a flexible nonparametric statistical machine learning algorithm suitable for regression and classification problems that can handle nonlinear interactions between predictors [31,32,33], and has been successfully applied in different contexts including coral reefs [23,34] and marine protected area design [35,36]. High performance in BRTs is reached by assembling a large number of simple trees [37], starting from a tree produced from a random subset of the data and sequentially adding further trees to find the best model fit [31,32,33], as measured by loss in predictive performance (e.g., deviance) [29]. The final prediction is calculated as a weighted average across all trees [29]. We used the number of trees, interaction depth and shrinkage as parameters to tune the models [33], and relied on root mean squared error (continuous models) and overall accuracy (thematic models) to choose the best combination of parameters that reduce overfit as suggested by Elith et al. [29]. In fact, we relied on the recommendation by Hastie et al. (2009) that setting the shrinkage parameter in BRT “tends to eliminate the problem of overfitting, especially for larger data sets”. Final parameters used to tune the models are available in the Supplementary materials (Script S3–S5). Although overfitting has been recognized as an issue in machine learning techniques that reduce model’s generality [29,38], Zaki and Meira [39] stated that some overfit is actually helpful, especially for those classes that might go underrepresented. Futhermore, we followed Schiele et al. [11] and ran models separately to keep the number of classes at each run relatively low.

We applied BRTs (Gaussian distribution) to predict percent coverage of finer substrates (i.e., soft clay, silt, fine to coarse sand from sieved samples, see Table 2 for size ranges), percent coverage of coarser substrates (i.e., sand (which also included subfractions of silt and clay not visible in the imagery), gravel, pebbles, stones, boulders, large boulders and bedrock, see Table 2 for size ranges) and percent coverage of benthic organisms according to HUB level 5 and 6 groups (

Table 4. List of HUB level 5 and 6 groups in the study area, and the species included.

HUB Level 5 and 6	Species included in Group
Characterized by annual algae	Ceramium tenuicorne, Ceramium sp., Chorda filum, Dictyosiphon/Stict\yosiphon (complex), Ectocarpus/Pylaiella (complex), Filamentous Phaeophyceae, Haliosiphon tomentosus
Characterized by perennial algae	Battersia arctica, Coccotylus/Phyllophora (complex), Delesseria sanguinea, Filamentous Rhodophyceae, Furcellaria lumbricalis, Polysiphonia/Rhodomela (complex), Unidentified Rhodophyceae.
Dominated by perennial filamentous algae	Battersia arctica, Filamentous Rhodophyceae, Polysiphonia/Rhodomela (complex).
Characterized by epibenthic bivalves	Mytilus spp.
Dominated by Mytilidae	Mytilus spp.
Characterized by cnidarians	Hydrozoa (Cordylophora caspia)
Characterized by moss animals	Electra sp.

). We also applied BRTs (multinomial distribution) to predict thematic maps of HUB 3–6 and Natura 2000 classes directly from reclassified training data.

We assessed the outputs of the substrate models to spot obvious artifacts (e.g., due to survey routes), followed by expert annotation, which helped to improve the training dataset. The latter step was repeated until the draft maps were approved. The substrate maps where then used as part of the predictor variables for the biological models, as well as the fine sediment fraction models (i.e., in addition to the 41 predictor variables also used in the substrate models). Finally, we reran several models with different predictor variables removed to better understand how different groups of the predictors affected map accuracy. BRTs, as an ensemble approach, use a high number of trees to turn possibly weak predictors into a strong prediction (Kuhn et al. 2013). Being quite robust to collinearity, BRTs handles many predictors as it ignores non-informative predictors when fitting trees [29].

2.4. Combining Models into Classified Maps

To combine the models into classes, we adjusted all substrate and biological components to ensure that the model outputs corresponded with the rules of the point intercept analysis. The substrate components had to sum to 100% in each pixel, which was realized by dividing each component by the total sum of all components. Similarly, the fine fraction components modeled from grab samples (soft clay to coarse sand) where adjusted to fit the sand model, realized by multiplying the proportion of each fine fraction component with the sand component. Biological components were adjusted in a similar way; uncolonized substrate and biological cover had to sum to 100% or more, and no single biological component when summed with uncolonized substrate could be >100%.

Next, we reclassified the outputs of the percent coverage models using conditional statements according to HUB (e.g., as seen in Table 1) [6], Natura 2000, and the SGU substrate scheme [40] (R script available, see supplementary S6–S8). For cases with unclear definitions as percent coverage or terrain metrics (i.e., Natura 2000 subtypes, SGU substrate scheme), we relied on text descriptions to define a conditional statement. Since scale was not clearly defined for any of the classification schemes, we used the full resolution of the models (5 m), but also created classified maps in other resolutions up to 250 m pixels by aggregating the percent coverage models before the conditional statements were run.

2.5. Accuracy Assessment

We validated all maps using a stratified random (HUB 1–6) subset of data retrieved from the high-resolution drop-camera (154 testing locations extracted from the original 559 sites). The substrate models were also evaluated with training data using bootstrap prior to the spatial prediction.

Thematic models were assessed with confusion matrices of observed vs. predicted [41] for each level of HUB and Natura 2000 maps. Furthermore, Kendall rank and correlation coefficients (tau) were calculated [42].

Continuous models were assessed using R², root mean square error (RMSE), mean absolute error (MAE) and mean error (ME, observed—predicted), as well as regression of observed vs. predicted values [41], which included 95% confidence band between observed and predicted coverage. In addition, the predicted values were grouped into intervals (0–10%, 10–50%, 50–90%, 90–100%) and the observed vs. predicted for each group was assessed using ME and standard deviation (SD). Finally, the same groups, as well as absence/presence were evaluated as classes using the same confusion matrix as the thematic models (overall accuracy and Tau). The continuous models of fine fractions were only evaluated using bootstrap prior to the spatial prediction due to the relatively low number of sieved grab samples.

Spatial patterns for model uncertainty was estimated combining all model outputs (generated using different predictor variables and training data in the development phase) for our different substrate models (10 different models for each component, which included both individual and compositional model runs) and all outputs for the biological models (5 different models for each component using different predictor variables) and produced mean standard deviation maps for substrate and biological models in ArcGIS.

3. Results

Below we describe the results from the study structured along two main themes: (1) technical section with modeled maps and accuracy (Section 3.1, Section 3.2, Section 3.3 and Section 3.4); and (2) End-user examples (Section 3.5), including the effect of scale and predefined thresholds. In addition to the images shown in the result, a geospatial pdf showing maps and survey data in more detail are provided in the Supplementary materials (Map S1).

3.1. Thematic vs. Reclassified

When comparing thematic maps (i.e., thematic models) and the reclassified thematic maps (i.e., from the continuous workflow) our results indicated that the accuracy was similar across all levels of the classification schemes regardless of which BRT workflow we used (

Table 5. Observed vs. predicted accuracy from 154 testing sites for thematic maps (thematic distribution BRT models) and reclassified thematic maps (from continuous BRT models).

	Overall Accuracy (%)		Tau
	Thematic	Reclassified	Thematic	Reclassified
HUB 3	77.9	80.5	0.72	0.76
HUB 4	79.9	81.8	0.70	0.72
HUB 4–5	63.0	62.3	0.55	0.55
HUB 4–6	55.8	53.2	0.49	0.47
Natura 2000	87.7	87.7	0.75	0.75
Mean	72.9	73.1	0.64	0.65

). This observation was further strengthened by studying the spatial output of the maps together with high-resolution backscatter and bathymetry data, taking visual aspects such as survey artifacts into account.

Map accuracy was highest for the maps with fewer classes (i.e., Natura 2000) with decreasing accuracy as the number of classes increased (Table 5 and

Table 6. Number of unique classes found in the reclassified thematic maps versus the thematic maps (same number of classes as in the training data). Number of classes with less than 10 predicted pixels are put in brackets (full list of classes and pixel count for HUB 1–6 available in Supplementary Materials Table S3).

	HUB 3	HUB 4	HUB 4–5	HUB 4–6	HUB 1–6	N2000	N2000 Subtypes
Thematic maps	5	3	7	9	29	2	-
Reclassified maps	5	3	8	12	59 (8)	2	12

). The number of classes in the reclassified map products was generally higher than the original classes identified in the training data, while the thematic maps were restricted to the number of classes in the training data set. For the complete HUB map level 1–6 the number of classes doubled in the reclassified map from 29 to 59 classes, however 8 of these HUB 1–6 classes had fewer than 10 predicted pixels (Table 6), of which some could be the result of pixel mismatch between the combined models. Very few pixels had obviously illogical combination of classes (based on our observations and knowledge of the survey area), the only clear example was 2 pixels classified as sand dominated by Mytilus spp. (full list of classes and pixel count for HUB 1–6 available in Supplementary materials Table S3).

3.2. Predictors and Training Data

We investigated how different types of predictors and training data contributed to the thematic HUB 3 substrate models (

Table 7. Comparison of HUB 3 classification model overall accuracy (OA), Cohen’s kappa and Tau, when using different predictor variables and training data. The accuracy was calculated from the testing data (30% of 2016–2017 drop camera samples), and additionally by bootstrapping the training data in the model fitting stage (may show inflated accuracy numbers in case of overfitting). Predictor variables (all groups contained northing/easting): (1) Depth metrics 0.5 m–2 km, OBIA metrics and backscatter metrics. (2) Depth metrics 0.5 m–2 km, backscatter. (3) Depth metrics 5 m, backscatter mosaic. (4) Depth metrics 0.5 m–2 km. (5) Depth metrics 0.5 m–5 m. (6) Depth metrics 5 m.

Predictors	Training Data	Bootstrap (Training Data)	Validation (Testing Data)
1) All	Survey, legacy, expert	OA 77%, kappa 0.68	OA 77,9%, Tau 0.72
2) Reduced	Survey, legacy, expert	OA 76%, kappa 0.66	OA 77.2%, Tau 0.71
3) Reduced	Survey, legacy, expert	OA 70%, kappa 0.58	OA 73.4%, Tau 0.67
4) Reduced	Survey, legacy, expert	OA 68%, kappa 0.55	OA 64.3%, Tau 0.55
5) Reduced	Survey, legacy, expert	OA 64% kappa 0.49	OA 61.0%, Tau 0.51
6) Reduced	Survey, legacy, expert	OA 63% kappa 0.47	OA 55.2%, Tau 0.44
1) All	Survey	OA 73% kappa 0.60	OA 58.9%, Tau 0.49

). Models trained using the same predictors, but only the point intercept dataset for training, showed a large decline in accuracy (testing data overall accuracy 77.9% vs. 58.9%, while bootstrap of training data showed overall accuracy 77% vs. 73%, indicating overfitting in the model with only survey data included). These results highlight that legacy data and expert annotations were important to the model fit. Expert annotations took approximately 5 min per site, legacy data approximately 10 min, while substrate and biota point intercept annotations of the drop camera data (2016–2017) averaged around one hour (annotations of 100 points/site ranged from 5 min to 2 h depending on site complexity).

The highest accuracy was observed for the model using all training data (including legacy and expert annotations) and all types of predictor variables (including backscatter, multiscale metrics 0.5 m–2 km and OBIA). When testing how OBIA, multiscale metrics and backscatter affected model accuracy (in addition to 5 m depth metrics which was always included), we observed that models that included only a backscatter mosaic had relatively high accuracy (73.4%), showing that backscatter is an important predictor and that relatively simple models with 5 m depth and backscatter metrics can perform well. The accuracy improved further (+4.5%) when OBIA and multiscale (0.5 m–2 km) metrics were also included, with OBIA having the highest percent contribution in the models (25.8%, Table S2). Performance reduced in all models when backscatter was not included, even if more advanced depth metrics were used. However, some of the loss in accuracy was regained by including multiscale (0.5 m–2 km) depth metrics (OBIA was not tested due to the inclusion of backscatter in the segmenting process). The percent contribution for all predictors used in the thematic models (HUB 3–6) is available in Appendix A

Table A1. Percent contribution of predictors (grouped by theme, see Supplementary Materials, Table S2 for more detailed breakdown) in thematic BRTs with and without substrate (sub) models included in the predictor variables.

Predictor Group	HUB3 % contr.	HUB4–6 no sub % contr.	HUB4–6 with sub % contr.	Mytilus no sub % contr.	Mytilus with sub % contr.	Average no sub% contr.	Average with sub % contr.
BS mosaic metrics 0.5 m–1 m	3.1	4.3	3.0	3.3	1.2	3.5	2.1
BS mosaic metrics 5 m	10.0	4.6	2.0	3.8	0.7	6.1	1.3
BS ARA	2.8	4.9	1.4	3.1	0.6	3.6	1.0
BS mosaic multiscale 20 m–2 km	4.0	4.1	2.6	3.2	1.1	3.8	1.9
Pg.s sediment thickness	1.4	1.4	1.4	1.4	1.4	1.4	1.4
Depth 5 m	4.4	3.8	3.7	2.1	0.8	3.4	2.3
Terrain metrics 0.5 m–1 m	8.5	9.4	5.6	10.1	2.7	9.3	4.1
Terrain metrics 5 m	5.4	7.7	5.3	6.7	2.4	6.6	3.9
Terrain metrics multiscale 20 m–2 km	16.3	20.6	15.5	14.7	7.3	17.2	11.4
Distance to reefs and sand	9.6	7.4	3.2	4.9	1.4	7.3	2.3
OBIA based on terrain metrics and BS 2.5 m	25.9	19.7	7.3	40.0	3.0	28.5	5.1
Oceanography	2.4	2.5	2.2	1.3	0.6	2.1	1.4
Uncertainty (depth, backscatter)	2.0	3.1	2.3	2.0	1.3	2.4	1.8
Northing/easting	4.3	3.5	3.1	1.7	0.9	3.1	2.0
Substrate models	-	-	40.2	-	75.2	-	57.7

(summary) and Supplementary Materials Table S2.

Comparisons between biological models that included substrate models as predictors and those that did not, indicated that the inclusion of substrate models developed using a larger training data set (i.e., survey, legacy and expert annotations) could be used to increase the accuracy of biological models. For example, accuracy of Mytilus spp. models improved with the addition of substrate models, and the total mussel cover increased from 8% to 10.7% (

Table 8. Accuracy table and mean percent cover for Mytilus spp. models with different predictor variables. Accuracy was measured (testing data, observed vs. predicted) with R², root mean square error (RMSE), mean absolute error (MAE) and mean error (ME). In addition, the predicted values were grouped into intervals (0–10%, 10–50%, 50–90%, 90–100%) and evaluated with overall accuracy (OA) as classes and absence—presence. Predictors (1) all refer to both substrate model outputs (developed using all available training data and predictor variables) and survey metrics (Supplementary materials Table S2), (2) same as (1) but excluding substrate model outputs, (3) same as (2) excluding all backscatter derived predictors.

Predictors	R2	RMSE	MAE	ME	OA Classes	OA Abs-pres.	Cover (Mean)
1) All	0.62	15.53	9.10	2.30	75 %	89 %	10.7 %
2) No substrate	0.60	16.69	9.84	5.61	69 %	88 %	8.0 %
3) No backscatter	0.49	19.60	11.83	7.96	60 %	75 %	4.4 %

). The total contribution of substrate predictors in the model was 75% (Appendix A Table A1, Supplementary materials Table S2), further highlighting the strong link between substrate and Mytilus spp. A visual comparison between the outputs showed that the greatest change in mussel cover occurred in flat areas composed of hard substrates, and areas with coarse soft substrates where it was difficult to differentiate between the two types of environments in the depth and backscatter data. When testing the same models but also removing backscatter the performance decreased further and the mean mussel cover changed to 4.4% (Table 8), showing the importance of backscatter as a predictor variable for the biological models. For the thematic models of biological HUB classes, the overall accuracy increased slightly for all levels when substrate models where included as predictors (HUB 4 +0.7%, HUB 4–5 +1.3%, HUB 4–6 +1.9 %).

3.3. Continuous Percent Coverage Models

3.3.1. Substrate Components

Initially, two different ways to predict substrate components using BRTs were explored i.e., individual and compositional models. The two modeling approaches showed similar overall accuracy, with some slight differences, e.g., small patches with a high cover of hard clay were under predicted in the compositional model, which was not the case for the individual model. Therefore, we present the simpler individual continuous modeling approach (hereafter referred to as continuous models). Furthermore, most individual models improved slightly when adjusting the models to sum to 100% (by dividing with the total sum). A comparison between individual and compositional substrate models can be found in Appendix A,

Table A2. Substrate component accuracy assessment using the validation data set (n = 154) for four model types, individual models (ind), individual models adjusted to 100% sum (i.adj), compositional models (comp), and combined compositional and individual models (hard clay component) adjusted to sum to 100% (comb). All models were run using Gaussian distribution boosted regression tree models with the same training data and predictor variables.

	R2				RMSE				MAE				Bias
	ind	i.adj	comp	comb	ind	i.adj	comp	comb	Ind	i.adj	comp	comb	ind	i.adj	comp	comb
Sand	0.80	0.83	0.81	0.81	20.4	17.9	19.2	19.3	11.8	10.1	10.0	10.0	5.7	−1.4	0.9	0.9
Gravel	0.46	0.53	0.49	0.49	12.6	12.6	12.6	12.7	6.9	7.4	7.1	7.1	2.8	−2.3	0.4	0.3
Pebbles	0.23	0.34	0.28	0.28	17.7	15.7	18.2	18.2	10.1	9.3	10.6	10.5	6.0	3.4	−0.6	−0.6
L stones	0.25	0.32	0.32	0.33	13.5	13.4	16.4	16.2	7.3	7.4	9.4	9.4	4.1	−0.2	−3.6	−3.5
Boulders	0.44	0.61	0.55	0.56	18.8	15.4	16.8	16.7	9.8	8.7	8.7	8.5	5.6	−0.4	0.4	0.6
L boulders	0.44	0.36	0.62	0.64	9.7	11.5	7.8	7.6	4.0	4.7	3.2	3.1	1.1	−0.2	1.5	1.6
Hard clay	0.66	0.61	0.23	0.36	5.7	5.8	9.2	8.8	1.8	1.8	2.3	2.0	1.5	1.1	1.1	0.7
Mean	0.44	0.51	0.51	0.52	15.4	13.2	15.2	15.1	8.3	7.1	8.2	8.1	4.2	0	−0.2	−0.1
SD	0.20	0.19	0.21	0.19	5.3	3.9	4.5	4.6	3.6	1.8	3.3	3.4	2.1	1.8	1.7	1.7

where we also combined the two approaches with some success.

The overall spatial patterns of the modeled substrate components (Figure 4) were as expected based on knowledge from the survey. For example, the model of hard clay successfully mapped clay features also seen in the multibeam data and the large boulder model captured pixels where distinct boulders could be identified in the multibeam data. Survey artifacts (mainly due to sound velocity errors in the outer multibeam swath, as well as nadir artifacts in the backscatter signal) where suppressed in the models, but remained a challenge (seen as striping in SW/NE direction, Figure 4i,j), especially in flat areas of mixed substrates where backscatter was observed to be less useful to differentiate between coarse sand-gravel and low relief hard bottom areas.

Regression analysis of observed and predicted substrate cover showed varying degrees of fit dependent on the substrate component with R² ranging from 0.32 (large stones) to 0.83 (sand) (Figure 5,

Table 9. Accuracy of percent coverage substrate components (individual models adjusted to sum to 100%). Distribution of the ground-truth dataset divided into HUB 3 classes are available in Table A4. Accuracy was measured (testing data, observed vs. predicted) with R², root mean square error (RMSE), mean absolute error (MAE) and mean error (ME), in addition, the predicted values were grouped and evaluated with overall accuracy (OA) as classes (0–10%, 10–50%, 50–90%, 90–100%) and absence - presence (<absence threshold, >absence threshold-100%).

	R2	RMSE	MAE	ME	OA Classes	OA Abs-pres.
Sand	0.83	17.9	10.1	−1.4	72.1	84.4
Gravel	0.53	12.6	7.4	−2.3	74.0	77.9
Pebbles	0.34	15.7	9.3	3.4	75.3	81.2
Large stones	0.32	13.4	7.4	−0.2	77.9	80.5
Boulders	0.61	15.4	8.7	−0.4	74.7	85.1
Large boulders	0.36	11.5	4.7	−0.2	83.1	87.7
Hard clay	0.61	5.8	1.8	1.1	91.6	92.2
Hard bottom	0.79	18.2	9.4	0.3	75.3	92.9

). MAE ranged from 1.8% (hard clay) to 10.1% (sand) across all substrate component models, however the low MAE for hard clay was likely a result of a low occurrence of hard clay in the area. On average there didn’t appear to be a substantial amount of bias in the models with ME ranging from −2.3% (gravel) to 3.4% (pebbles). However, all models, to varying degrees, tended to underestimate toward the lower end of the cover scale and overestimate towards the upper end. This was exemplified by positive ME in the 0-10% class and a predominance of increasingly negative ME values in classes >10% (

Table 10. Accuracy of percent coverage substrate components (individual models adjusted to sum to 100%), divided into groups. Accuracy was measured (testing data, observed vs. predicted) with mean error (ME) and standard deviation (SD).

	0–10%			10–50%			50–90%			90–100%
	n	ME	SD	n	ME	SD	n	ME	SD	n	ME	SD
Sand	53	5.6	15.3	33	−2.1	20.8	20	−15.5	27.3	48	−2.7	7.2
Gravel	95	1.9	7.8	50	−9.1	14.2	9	−8.5	21.6	0	-	-
Pebbles	100	4.2	10.0	52	2.3	22.2	2	−11.4	24.0	0	-	-
Large stones	102	3.0	9.9	50	−5.7	16.4	2	−24.9	25.3	0	-	-
Boulders	105	2.0	5.9	29	−6.3	24.6	20	−4.6	27.2	0	-	-
Large boulders	134	1.3	5.5	15	−5.2	21.8	5	−25.1	38.6	0	-	-
Hard clay	152	1.3	5.0	1	−32.1	-	1	7.5	-	0	-	-
Hard bottom	86	4.8	13.3	21	3.6	22.4	24	−14.7	26.6	23	−4.4	8.9

). Furthermore, regression lines tended to be above the 1:1 line at the lower end of the scale and below the 1:1 line at the upper end indicating the aforementioned effect (Figure 5). Overall accuracy when predicted values were reclassified into classes 0–10%, >10–50%, >50–90% and >90–100% (OA classes) or presence >0.001% (OA abs-pres.) were generally high ranging from 72.1% (sand) to 91.6% (hard clay), and 77.9% (gravel) to 92.9% (hard substrates) respectively (Table 9).

For sand, which was the dominating substrate type on the bank, the highest accuracy was found in the 0–10% interval (which in most cases equals 90–100% hard bottom) and the 90–100% interval (i.e., 0–10% hard bottom), showing higher uncertainty in the mixed hard/soft bottom areas. The individual hard bottom components, which rarely occurred in the 90–100% interval, had the highest accuracy in the 0–10% interval, increasing uncertainty of predictions >10%.

3.3.2. Substrate Fine Fractions

The maps of sieved fine fractions adjusted to the main sand model (which included finer subfractions as well) are shown in Figure 6, with accuracy assessments using bootstrap of the sieved training data shown in

Table 11. Accuracy statistics (using bootstrap of training data, n = 25) for fine fraction proportions predicted from sieved sediment samples. R², root mean square error (RMSE), mean absolute error (MAE).

	R2	RMSE	MAE
Clay	0.03	2.02	0.92
Clay-silt	0.27	0.71	0.54
Silt	0.32	0.66	0.53
Fine sand	0.76	1.07	0.85
Medium sand	0.79	0.76	0.59
Coarse sand	0.58	1.41	1.02

. In a visual assessment of the outputs, we noticed that silt and fine sand accumulated in areas where post-glacial sand deposits were seen in the sediment echo sounder profiles. This observation often also corresponded with sediments with no ripples and the occurrence of Macoma balthica shell aggregations. The fine fraction models did not have very high accuracy overall and would benefit from a larger training dataset to describe such substrate features. However, the accuracy of the adjusted maps shown in Figure 6 is likely higher due to the adjustment to the main sand model (Table 9 and Table 10).

3.3.3. Biological Components

The individual outputs and corresponding accuracy assessment from the biological models included in the HUB 5 level are shown in Figure 7. Similar to the substrate components some survey artifacts carried over to the biological models (seen as striping in SW/NE direction along survey lines in Figure 7g–i, However most of the artifacts occurred in mixed sediment areas with low relief, while sand areas and more distinct reef features showed very few visual artifacts.

Regression analysis of observed and predicted HUB level-5 biological cover showed varying degrees of fit dependent on the biological component with R² ranging from 0.01 (epibenthic moss animals (Bryozoans)) to 0.84 (colonized substrate) (Figure 8,

Table 12. Accuracy of percent coverage biological components (HUB level-5 groups). Accuracy was measured (testing data, observed vs. predicted) with R², root mean square error (RMSE), mean absolute error (MAE) and mean error (ME), in addition, the predicted values were grouped and evaluated with overall accuracy (OA) as classes (0–10%, 10–50%, 50–90%, 90–100%) and absence - presence (<absence threshold, >absence threshold 100%).

	R2	RMSE	MAE	ME	OA Classes	OA Abs-pres.
Annual algae	0.37	7.1	1.5	1.5	94.8	91.6
Perennial algae	0.52	20.4	13.0	1.8	68.2	91.6
Mytilus spp.	0.62	15.5	9.1	2.3	75.3	89.0
Cnidarians	0.12	9.7	4.7	2.4	77.3	79.9
Moss animals	0.01	0.8	0.2	−0.2	100	90.3
Colonized substrate	0.84	17.9	8.0	−6.7	80.5	97.4

). MAE ranged from 0.2% (moss animals) to 13% (perennial algae) across all HUB level-5 biological component models, however the low MAE for moss animals was likely a result of extremely low cover of moss animals in the area. On average, bias in the models was relatively low with the exception of uncolonized substrate where ME was −6.7% (Table 12). Similar to substrate components, models for perennial algae, epibenthic bivalves, and epibenthic cnidarians, to varying degrees, tended to underestimate toward the lower end of the cover scale and overestimate toward the upper end (

Table 13. Accuracy of percent coverage biological components (HUB level-5 groups) divided into map percent cover intervals. Distribution of the ground-truth dataset divided into HUB 4–6 classes are available in Table A5. Accuracy was measured (testing data, observed vs. predicted) with mean error (ME) and standard deviation (SD).

	0–10%			10–50%			50–90%			90–100%
	n	ME	SD	n	ME	SD	n	ME	SD	n	ME	SD
Annual algae	153	1.4	6.9	1	15.7	-	0	-	-	0	-	-
Perennial algae	84	4.3	10.9	39	5.0	29.5	31	−8.8	23.0	0	-	-
Mytilus spp.	80	2.6	8.4	56	2.5	20.7	18	0.6	20.5	0	-	-
Cnidarian	139	3.4	8.5	15	−7.1	12.4	0	-	-	0	-	-
Moss animals	154	0.2	0.8	0	-	-	0	-	-	0	-	-
Colonized	27	−0.8	1.8	22	−9.8	14.2	23	−13.7	27.6	82	−5.9	15.3

and Figure 8b–d). The annual algae model did not behave in the same manner tending to show an underestimation of annual algae cover, however this was likely because 149 of 154 of the predicted cover values were <2% making it difficult to assess. The epibenthic moss animal model was difficult to assess as all predicted values were <0.9% and observed coverage was generally was low (<6%, Figure 8e). The uncolonized substrate model, on the other hand, tended to overestimate the cover of uncolonized substrate. This was exemplified by negative ME values across all classes (Table 13) as well as the regression line below the 1:1 line (Figure 8f). Overall accuracy where predicted values were reclassified into classes 0–10%, >10–50%, >50–90% and >90–100% (OA classes) were generally high ranging from 68.2% (perennial algae) to 100% (epibenthic moss animals) (Table 8). However, high values for epibenthic moss animals and annual algae were not surprising as few to no observed values were >10% (Table 8). Overall accuracy where predicted values were reclassified to presence (OA abs-pres.) was high, ranging from 79.9% (Cnidarians) to 97.4% (Table 8).

Both substrate and biological component generally showed higher uncertainty in the mixed and often patchy hard/soft bottom areas (Figure 5, Figure 8 and Table 10, Table 13). This was further confirmed by the maps of standard deviation between all model outputs (Figure 9).

3.4. Thematic Maps

The reclassified thematic maps according to HUB, Natura 2000, and SGU substrate maps at 5-m resolution are shown in Figure 10, Figure 11 and Figure 12. The accuracy of the HUB and Natura 2000 maps had overall accuracy ranging from 53.2% (HUB level 4–6) to 87.7% (Natura 2000). More details are available in Table 5, as well as in the user-producer matrix of HUB 3 (Appendix A

Table A3. Observed vs. predicted confusion matrix for reclassified HUB 3 map using the validation data set (n = 154).

	Coarse Substrate	Hard Clay	Mixed	Rock & Boulder	Sand	Sum (n)	User Accuracy %
Coarse substrate	12	0	10	0	0	22	54.55
Hard clay	0	0	0	0	0	0	0
Mixed	3	1	42	3	0	49	85.71
Rock & boulder	0	0	3	16	0	19	84.21
Sand	3	0	7	0	54	64	84.38
Sum (n)	18	1	62	19	54	154	0
Producer accuracy %	66.67	0	67.74	84.21	100	0	80.52

). No accuracy assessment was available for the SGU substrate maps and Natura 2000 subtypes, since the testing data did not include these classes due to the included terrain metrics (Natura 2000 reef subtypes) and sediment samples (i.e., fine sand class in SGUs substrate maps).

The Natura 2000 subtype map captured eight reef classes and four sandbank classes, revealing features and seascape patterns not seen in the main Natura 2000 map (two classes, sandbanks and reefs), including an extensive pattern of ridges (Figure 11). Similarly, the reclassified SGU surface substrate map captured 6 classes and showed considerably more detail than the legacy surface substrate map, especially for the areas mapped with 1:500 000 scale (Figure 12).

3.5. End-User Applications

In the following sections we investigate some of the end user applications of reclassified continuous maps.

3.5.1. Impact of Scale

The spatial resolution used when reclassifying continuous maps affected the percent coverage of the resulting thematic classes. When the HUB level 3 classification was applied to a growing-size window, small distinct features declined in cover and were replaced by more general classes (Figure 13,

Table 14. Percent coverage changes for HUB level 3 classes as the analysis window changes from 5 m to 250 pixels. Significant changes of cover for the rarer classes occur already at the 10-m pixels.

	5 m	10 m	25 m	50 m	250 m
Sand	36.4%, 469 km2	35.2%, 472 km2	35%, 465 km2	33.9%, 456 km2	31.3%, 420 km2
Coarse	16.2%, 217 km2	14.3%, 192 km2	13%, 179 km2	12,4%, 167 km2	10,0%, 133 km2
Mixed	46.4%, 623 km2	50,0%, 672 km2	51.8%, 696 km2	53.5%, 719 km2	58.7%, 789 km2
Rock and boulder	0.74%, 10 km2	0.56%, 7.5 km2	0.32% 4.3 km2	0.10%, 1.4 km2	0.01%, 0.12 km2
Hard clay	0.06%, 0.81 km2	0.02%, 0.22 km2	0.003%, 0.04 km2	0	0

).

In the Natura 2000 maps (reef/sandbanks), the reef class increased in cover as the window-size increased (Figure 14), moving towards 100% cover of reef when using mean cover for the whole bank. Predicted mean cover of hard bottom (19.3%) was below the threshold for reef (>50%) but Mytilus in the study area was above the threshold (>10% mussel cover), which then classified it as reef. Conversely, if the mean cover for Mytilus would have been < 10%, the mean cover of the study area would have been classified as sandbank, reversing the pattern of the growing analysis window.

High-resolution habitat maps generally allow the user to choose across many different options, therefore, aggregating the information to the appropriate-management scale. Applying a management scale of 250 m pixels to the HUB 3 classification made the classes rock and boulder and hard clay disappear (0.01% and 0%). However, if instead all 250 m pixels containing these classes at the 5 m scale were mapped, rock and boulder covered 32% and hard clay 17.8% (Figure 15).

3.5.2. Effect of Predefined Thresholds in HUB

The areas classified as characterized by epibenthic bivalves or dominated by Mytilidae (i.e., ≥10% and ≥50% mussel cover respectively) in the reclassified HUB map were 161 km² (12%) and 11.6 km² (0.9% of the area) respectively. The area using the same thresholds but calculated directly from the percent coverage map of Mytilus was 479 km² (36%) and 19.8 km² (1.5%) (Figure 16). Due to overlap with other benthic organisms with higher cover (primarily perennial algae), the HUB map was not accurate when analyzing percent cover of individual organisms.

4. Discussion

High-resolution continuous and thematic models can help to overcome some of the limitations of classification schemes. We found that the accuracy (overall accuracy and Tau of observed vs. predicted) of classified HUB and Natura 2000 maps were similar regardless of whether we used classification models directly, or reclassified models of continuous percent coverage. However, we also found that producing continuous maps and then applying classifications had a greater value due to the added possibilities, such as simple application of thresholds (conditions) or spatial analysis directly from continuous variables (e.g., through raster calculation).

In the first part of the discussion below, we discuss how the use of continuous maps affects the possibilities for end users. In the second part, we focus on technical aspects of this study and the accuracy of the maps we produced.

4.1. User Applications

Given that benthic habitat maps are often supported by similar types of information (e.g., substrate and benthic organisms), it is possible to map different classification schemes that share the same continuous components (but perhaps not the same combination of thresholds), without having to rerun any models. For the case where maps include components that are not suitable for continuous models, the classification can still use a combination of percent coverage predictions and terrain metrics (e.g., BPI, slope) to map geomorphological features such as ridges and walls (in this study exemplified by the ridge class in the SGU implementation of Natura 2000 subtypes, mapped from a combination of percent hard bottom cover prediction and BPI thresholds), or use additional thematic models for some specific features (in this study exemplified by the sand ripple class in the SGU implementation of Natura 2000 subtypes).

Moreover, the thematic maps that are based on continuous models are not limited to one spatial scale but can easily be adjusted to any window larger than the minimum mapping unit. Therefore, we provided all HUB and Natura 2000 maps at multiple scales to accommodate different definitions and applications. The scale and the minimum mapping unit can dramatically change the occurrence of different thematic classes [43,44], especially in patchy environments like our study area as shown in Figure 13 and Table 14. Though some guidelines of scale often do exist [3], clear directions for how to address the problem of thematic changes at different scales are missing from many schemes, including EUNIS, HUB and Natura 2000. The importance of such schemes makes this a priority to address. When the mapmakers and end-users are free to decide on the scales relevant for a specific habitat and biotopes of interest it could come with unwanted effects. Assuming, for example, that different legislation is tied to the Natura 2000 reef and sandbank classes on Hoburgs Bank, it would then be important to also specify the scale, since the bank can be classified as a sandbank, a reef, or both depending on the size of the analysis window. Similarly, the scale also needs to be considered when looking to monitor habitat (and ecological status) change. A focus on conservation features (e.g., reefs) can often overlook the ecological importance of connected patches types such as sand or seagrasses in the seascape, which can also be critical resources for mobile species [45]. Furthermore, when transnational classification schemes are used in multiple mapping projects, the lack of a clear definition on scale can hamper the overarching goal of creating a unified map. Transnational maps can be seen as a common governance effort across countries borders as discussed by Pecl et al., [13] for instance when considering habitat maps as tools to manage effects of climate change to ecosystems.

In this study we show that an important advantage of continuous high-resolution maps is that they provide more options when moving information across different scales, for example from mapping to the management scale. In Figure 15, we recalculated the HUB level-3 substrate map at 250-m resolution, which is the resolution used by SwAM to assess current and future cumulative impact on the marine environment, and to evaluate the effect of the proposed national marine spatial plan [10]. At this scale, the less common classes hard clay and rock and boulder disappeared into the more commonly occurring HUB classes sand, coarse sediments and mixed. In fact, the coverage of the hard clay and rock and boulder classes started to decline already at the 10 m scale. If small but distinct features associated with hard clay and rock and boulder classes have important ecosystem functions, despite their coverage, it could be key to include this information at the management scale. For example, SGU (observation by Gustav Kågesten during field season 2018 in the north middle-seabank) and Beisigel et al. [46] found that such features are used as spawning locations by herring and as shelter by cod. On the contrary, if only commonly occurring classes are of interest perhaps a coarser scale is appropriate as well. When high-resolution maps are available there are alternative ways to include such specific features at the management scale, for example by mapping management units that contain a presence of the features of interest. The HUB rock and boulder class in our study area covered nearly a third of the bank, when testing this approach for 250 m pixels (Figure 15). This could be essential information in a marine spatial plan, for example assuming that the mere presence of large boulders in a planning unit provides grounds for protection from bottom trawling.

Maps of continuous variables can have a broad variety of applications where classified maps may not be as suitable. An obvious example in the study area is to estimate mussel biomass and related ecosystem services [47] using the continuous prediction, thereby providing the percent mussel coverage with greater certainty than thematic maps, which rely on intervals of mussel coverage. However, if for instance a bird conservation action is being planned based on the total area suitable for birds feeding on blue mussels, and assuming that birds seek areas with at least 50% mussel cover according to their feeding strategy, one may wrongly think that the HUB classification can help in the conservation plan. Though classes at level 4 to 6 capture locations characterized by an organism (if it has more than 10% coverage) or locations dominated by an organism (if it has more than 50% coverage), a location would only be considered dominated by an organism if no other organism had a greater coverage in that cell (since holdfast and canopy count as overlay in our point intercept data the total cover can exceed 100%). In our study area, we observed that perennial algae often covered Mytilus, and this caused the HUB classification to be labeled as dominated by perennial algae instead of mussels (which often still had a coverage of >50%, though not captured in the HUB map). In this last case, the locations will incorrectly be considered as unsuitable for the bird conservation action. According to our assumptions, the total area of suitable bird habitat almost doubles when basing the calculation on the continuous mussel prediction instead of the HUB map (Figure 16). If a different coverage of mussels is set (e.g., 10–50%), the suitable habitat increased threefold when using the mussel percent cover map instead of the HUB map. Finally, if the suitable feeding location threshold is a number larger than 50% of mussel cover, the HUB map will not provide any suitable area with confidence.

Differences in the representation of the relative abundance of habitat types in the seascape can have important implications for management activities including identifying and quantifying spatial change, valuing natural capital and protecting resource designations such as essential fish habitat. Furthermore, when habitat maps are used as spatial predictors in species distribution models or to quantify spatial patterns the issue of spatial and thematic resolution must be considered specifically as a key variable and carefully evaluated [34,48,49].

4.2. Technical Applications

Our use of legacy data and expert annotations proved, as in previous studies [50], to have a high value in the models. For example, their use increased the HUB 3 thematic substrate model accuracy (overall accuracy observed vs. predicted) from 59% to 78% (Table 7). The additional training data concerning substrate components also indirectly improved the biological models, using substrate model outputs as predictor variables. The accuracy of the Mytilus spp. model (R², RMSE, MAE and ME of observed vs. predicted) and HUB level 4–6 maps improved when the substrate predictions were included as predictor variables (Table 7), even when high resolution bathymetry and backscatter were available, which to some degree can act as surrogate for substrate grain size [51]. We also noticed fewer survey artifacts in the biological models when substrate models were included as predictor variables. The full combination of substrate and biological models into HUB biotopes level 1–6 (see Supplementary materials Legend S1 or Table S3 for all 59 resulting HUB classes), are likely to have high uncertainties at the pixel level for the end user (generally, the more thematic classes combined the lower the accuracy for each combination as shown in Table 5), but also generated very few pixels containing obviously rogue classes (such as an illogical combination of soft bottom dominated by a hard bottom species). We suspect that this was due to the use of substrate outputs in the biological models, and from using the same set of predictors. Our results showed the possibilities of including ground truthing data from several quality levels (i.e., high resolution survey, legacy data and expert annotations), that data from sediment samples and images can be combined using continuous models, and the benefits of integrated geological and biological surveys and models.

The high resolution of our survey enabled us to capture features such as sand ripples and individual boulders, using a combination of backscatter statistics, multiscale bathymetry and OBIA metrics (which had the highest overall contribution in the models, Table A1, Table S2). While models using the full range of predictor variables performed best and is the approach we recommend for future studies prioritising high accuracy, we concluded that a simpler combination of bathymetry metrics and a backscatter mosaic at resolution of 5 m would have been sufficient to create our models (accuracy for HUB level 3 decreased from 78% to 73%). The combination of 5 m bathymetry metrics and the backscatter mosaic outperformed any further combination of multiscale metrics that did not include backscatter, indicating the importance of backscatter as a predictor to capture sediment characteristics. This is not surprising given that backscatter has long been recognized as a key predictor variable [52,53]. When backscatter was not included, we observed that the multiscale metrics were increasingly important (OBIA metrics was not tested in this case, but it is likely it would be useful as well considering that it was important in the models that included backscatter), highlighting that some accuracy can be gained using high-resolution (≤1 m) bathymetry metrics and rescaling techniques.

Though we provided several different statistical measures describing the accuracy of our predictions, covering the gap between reality and the accuracy assessment of our predicted maps remains a challenge [14]. Combining continuous models into thematic maps, as well as using substrate models to train biological models made techniques such as cross-validation impractical. Hence, we used the same testing data for all models, which to some degree lowers the confidence in the accuracy statistics produced. Figure 9 confirms that error is higher in patchy areas and lower in homogeneous areas for both substrate and biological models. The manual point intercept annotations of high-resolution drop-camera imagery were a source of unknown error, and also a major part of the post processing effort. Due to the inclusion of lower quality expert annotations and legacy data in the training data set, we were able to lower the per site annotation effort for the substrate models, with a positive spillover effect to some of the biological models. Even so, we recognize that the time and cost needed to implement continuous percent coverage annotations may be a limiting factor for mappers, compared to thematic annotations. We believe that further development and application of computer vision technologies that have potential to support expert annotations of ground truth data and estimates of uncertainty [54,55,56,57] will be an important field to lower the costs and further improve the approach presented in this study.

Our maps of Hoburgs Bank are intended to be living rather than static products that can be enhanced as soon as, for instance, more or better training data becomes available. However, even with future improvements (e.g., related to data quality, reduction of human error, and more sophisticated models and techniques), we will always live with some degree of uncertainty due to natural variability that our modeling will not be able to incorporate [14,58].

5. Conclusions

Maps with continuously varying data on substrate and benthic organisms added more valuable information than more simplified patch-mosaic thematic maps. Continuous maps provide more flexibility in application and can be scale-adjusted and resampled more easily to address a wider range of research questions and management applications. Our findings provide a clear example of how high-resolution continuous maps can be successfully generated by combining multibeam depth, backscatter, drop-camera and sediment samples, and that additional classified maps can be generated from this workflow without the loss of accuracy compared with thematic models. While we recognize the value and convenience of thematic maps, relying solely on such classification schemes can result in the neglect of habitats or features with great environmental value. It is therefore essential to understand the limitations of any classified maps for marine spatial planning and decision-making particularly with regard to scale and schema dependence and the associated errors that may be propagated through the mapping process.