From detection to complexity: AI boulder mapping enables structural analysis of Baltic Sea reefs

Advances in AI-based boulder detection using hydroacoustic data enable detailed characterization of geogenic reefs. As AI-based detection approaches the level of accuracy of human interpretation in small-scale test areas, it opens up the opportunity to efficiently analyze and characterize larger regions. Geogenic hard substrates are a key habitat for diverse benthic communities that provide crucial ecosystem services. Current classification schemes for boulder fields in the German Baltic Sea, using three categories (0 boulders, 1-5 boulders, and >5 boulders) inadequately capture habitat complexity, thus limiting our understanding of these critical geogenic reefs. Convolutional neural networks were used to detect individual boulders on side scan sonar backscatter mosaics with 25 cm resolution across four study sites, covering an area of 306 km² in the German Baltic Sea. Region-specific AI models detected about 6.7 times more boulders than previous automated methods. A maximum of 550 boulders per 50 $\times$ 50 m grid cell were detected. A novel metric, the Boulder Field Complexity Index (BFCI), was developed to describe the complexity of boulder fields. The BFCI integrates boulder count with the transition of sharpness and the spatial distribution of individual boulders. Compared to conventional approaches, the BFCI enables the characterization of boulder field complexity on a continuous scale and reveals significant complexity differences between study sites. The coastal and shallow study site Plantagenet Ground demonstrates the highest level of complexity, whereas the offshore and deepest study site, Western Rönnebank, exhibits the lowest. The abundance of boulders is negatively correlated with water depth, with the highest densities occurring in shallow waters. The spatial variability in BFCI values reflects the heterogeneous nature of glacial till deposits and differential erosion processes that have shaped boulder field distribution. This approach provides the foundation for linking habitat heterogeneity to benthic community patterns and ecosystem functioning.

1 Introduction

Geogenic hard substrates are a key habitat for diverse benthic communities. These structures function as biodiversity hotspots, providing crucial settlement space for sessile organisms, shelter for mobile invertebrates, and serving as feeding, spawning, and nursery areas for fish, birds, and marine mammals (Franz et al., 2021). Hard substrates form a unique niche, thereby providing crucial ecosystem services, especially in environments dominated by soft bottoms (Hoffmann et al., 2022). These vulnerable habitats are increasingly under threat from a variety of human activities, including coastal engineering, resource exploitation, nutrient discharge, and climate change (Halpern et al., 2008; Franz et al., 2021). They are consequently worthy of protection or restoration (Wilms et al., 2021; Casabona et al., 2024).

Therefore, it is important to understand the distribution of geogenic hard substrates. In marine environments, optical remote sensing is limited by water depth and clarity (Kulha et al., 2024). Thus, hydroacoustic methods are essential for seafloor mapping (Brown et al., 2011), especially hard substrate mapping (Papenmeier and Hass, 2018). Boulders forming geogenic reefs can be identified due to their elevation from the surrounding seafloor, as determined by bathymetric measurements (multibeam echo sounder measurements), or due to their impact on acoustic backscatter (multibeam echo sounder and side scan sonar measurements). Side scan sonar systems are especially suited for boulder detection. They are towed closely above the seafloor and operate by emitting fan-shaped acoustic pulses perpendicular to the sonar’s travel path, recording the intensity of echoes reflected from the seafloor. Objects such as boulders produce strong acoustic returns and characteristic acoustic shadows. These acoustic signatures allow for the identification of individual boulders (von Rönn et al., 2019; Papenmeier et al., 2020).

Based on the seafloor maps produced by acoustic remote sensing, hard substrate habitats are classified differently across various European regulatory frameworks and classification systems. The EU Habitats Directive lists them as ‘reef’ (Annex one code 1170), the Marine Strategy Framework Directive refers to them as ‘rock and biogenic reefs’, the European Nature Information System (EUNIS) uses ‘rock and other hard substrata’, while the HELCOM Underwater Biotope and Habitat classification system (HELCOM HUB) employs the term ‘rock and boulders’ (European Commission, 1992; European Commission, 2008; Helcom, 2013; European Environment Agency, 2025). Since all of these classifications lack detailed descriptions concerning spatial dimensions, habitat delineation, and substrate characteristics, member states of the EU have developed their own criteria. In Germany, reefs are classified according to the “Leitfaden zur großflächigen Abgrenzung und Kartierung des Lebensraumtyps Riffe” (Guidelines for Large-Scale Delineation and Mapping of Reef Habitat Types), which establishes specific criteria for identifying and delineating geogenic reefs in German waters (Heinicke et al., 2021). For the EEZ of the Baltic Sea, 50 $\times$ 50 m grids are used to analyze backscatter mosaics with a resolution of 25 cm to identify individual boulders larger than 50 cm. The grids are then categorized into three classes based on the number of boulders: Class 1 (no boulder), Class 2 (1–5 boulders), and Class 3 ( $>5$ boulders). Boulder distribution maps for the German EEZ are publicly available (Papenmeier, 2024), and several areas in German coastal waters have been analyzed according to this procedure (Darr et al., 2022; Feldens et al., 2023; Marx et al., 2024b).

Recent progress in applying machine learning, specifically neural networks, to detect objects and features in hydroacoustic data is a milestone for seafloor mapping (Feldens, 2020; Arosio et al., 2023; Garone et al., 2023; Lundine et al., 2023; Rajani et al., 2023). For large-scale boulder detection, Hinz et al. (2024) demonstrated how AI-based methods can bridge the gap from experimental applications to practical implementation. Traditional manual interpretation of hydroacoustic data for boulder detection is known to be time-consuming, tedious, and subjective, making it nearly impossible to locate all boulders across extensive areas. While human experts can effectively classify boulder density into basic categories (none, 1–5, >5 boulders per grid cell), previous AI detections showed that actual boulder densities can reach several hundred boulders per 50 $\times$ 50 m area (Franz et al., 2021; Feldens et al., 2023). The application of convolutional neural networks like YOLOv4 and YOLOv8 has achieved detection accuracies of up to 77.8% for side-scan sonar data (Hinz et al., 2024), enabling rapid, consistent, and large-scale mapping. These automated approaches not only improve mapping efficiency but also enhance our understanding of fine-scale habitat heterogeneity, which has been shown to significantly influence benthic biodiversity (Coolen et al., 2015; Romoth et al., 2023).

While current classification schemes have established important foundations for boulder field mapping, classifying densities into distinct classes presents significant limitations. Our hypothesis is that currently applied manual methodologies, which consider only boulder abundance, are insufficient to capture reef variability and spatial heterogeneity across scales. With the achievements of AI-based boulder detection, we can now identify boulders in detail across large areas. This capability enables us not only to examine individual areas but also to compare multiple regions with each other, capturing the unique characteristics of each area while highlighting cross-regional patterns and similarities. This research aims to develop a novel approach to describe the structural complexity of boulder fields, which is not only controlled by the number of boulders but also by their spatial arrangement. Therefore, we introduce a novel metric that combines boulder count with the sharpness of transition and spatial distribution of individual objects within geogenic reefs.

2 Materials and methods

2.1 Study sites

The study sites (Figure 1) are distributed along the German Baltic Sea coast. The sites were primarily selected based on data availability and to represent diverse environments across the German Baltic Sea with different water depths and different coastal distances (1–30 km). In total, an area of about 306 km² was investigated. Kiel Bay (KB, 131 km²) is located in the western part and has water depths ranging from 15 m to 24 m. This study site was divided into two subparts (KB1 and KB2) to be able to use two different models for boulder detection. Mecklenburg Bay (MB, 27 km²) lies further east with depths between 8 and 20 m. Plantagenet Ground (PTG, 65 km²) is located in the central region, with water depths of 7–15 m. The easternmost site, Western Rönnebank (WRB, 83 km²), reaches depths of 23–35 m. Together, these sites cover a range of depths and geological settings typical for the German Baltic Sea. The boulders in the southern Baltic Sea primarily originate from Scandinavia. They were transported and deposited by Pleistocene ice sheets that advanced across the Baltic basin (Bressau, 1957; Matthäus, 2019). Since the last glacial maximum, the moraine deposits composed of boulder clay were reworked during sea level fluctuations (Lampe, 2002) in exposed areas. This has resulted in lag deposits containing exposed cobbles (64–256 mm) and boulders ( $>$ 250 mm) (Schwarzer et al., 2019). In more sheltered areas, moraine deposits were covered partly or completely by Holocene sands, in more offshore areas by mud. Sediment redistribution is ongoing in the Baltic Sea (Schwarzer and Bohling, 2008; Bohling et al., 2009) and can cause temporary covering or exposure of the hard substrates.

Figure 1

Figure 1. Overview map with the study sites located along the German Baltic Sea. KB: Kiel Bay, MB: Mecklenburg Bay PTG: Plantagenet Ground, and WRB: Western Rönnebank. Bathymetry source: provided by the Federal Maritime and Hydrographic Agency (BSH) for scientific use.

2.2 Hydroacoustic data

Backscatter mosaics based on side-scan sonar (SSS) with a spatial resolution of 25 cm were used to detect boulders across all study sites. Boulders were classified following the Wentworth scale ( $>$ 256 mm) (Wentworth, 1922). Given the 25 cm mosaic resolution, the effective lower detection limit was 50 cm (corresponding to two pixels). The data collection efforts were conducted over the period from 2006 to 2024 and involved multiple institutions and projects, including the Leibniz Institute for Baltic Sea Research Warnemünde (IOW), the Federal Maritime and Hydrographic Agency of Germany (BSH), the Federal Agency for Nature Conservation (BfN), and the State Agency for Environment, Nature Conservation and Geology Mecklenburg Vorpommern (LUNG MV). Table 1 provides details on the specific equipment, acquisition dates, and institutional sources for each site. All data were processed using SonarWiz Version seven software from Chesapeake Technology to apply standard geometric corrections, bottom tracking, layback correction, and empirical gain normalization. The backscatter mosaics were exported as GeoTIFF. Although the data originated from different sources and exhibited mixed quality characteristics due to variations in acquisition years (2006–2024) and equipment, overall quality was sufficient for AI-based boulder detection. Data for Mecklenburg Bay and Plantagenet Ground had previously been successfully utilized in boulder detection studies by Feldens et al. (2023).

Table 1

Table 1. Overview of data origin and side-scan systems used.

2.3 Automatic boulder detection using AI

For automatic boulder detection, the workflow for side-scan backscatter data described by Hinz et al. (2024) was used. The detection pipeline is based on the convolutional neural network YOLOv4, implemented in PyTorch. The model is trained on GIS-based experts’ annotations. Detected objects are automatically post-processed to remove, e.g., duplicates. The output of the workflow is a geopackage with centroids for each detected object. Two different models were applied to account for the varying characteristics of boulder fields across the study sites. In case of the Kiel Bay area, the study site was divided because of its different boulder densities in two sub-areas. Model performance was evaluated using two standard object detection metrics: Intersection over Union (IoU), which quantifies the overlap between predicted and actual boulder locations, and mean Average Precision at IoU threshold 0.50 (mAP50), which measures detection accuracy at $\ge 50\%$ overlap. Model one was trained using 5,909 boulder annotations from the Western Rönnebank site, including empty areas with no boulders to reduce false positives. The training process involved approximately 30,345 iterations, resulting in performance metrics of Intersection over Union (IoU) = 66.72%, mean Average Precision (mAP)-50 = 65.66%, and mAP-25 = 71.36%. This model performs well in regions where individual boulders and their acoustic shadows are clearly distinguishable from one another. Model one was applied to the WRB and KB2 study sites. Model two was specifically trained for dense boulder fields using 5,027 annotations from the Plantagenet Ground site. The model achieved a mAP-50 of 51% and an average IoU of 55.31%. Pre-trained weights from a previous model were used to improve convergence, accelerating the training process and enhancing model performance. Model two was applied to areas with high boulder density, such as PTG, MB, and KB1. While this model performs significantly better in dense boulder environments, it lacks training on empty areas, resulting in an elevated number of false positives in boulder-free zones. Consequently, boulder-free zones processed with Model two required manual post-processing to remove artifacts caused by false positive detections.

2.4 Postprocessing of detection results

QGIS Version 3.34 was used for post-processing of the detection results. To make the new geospatial statistical approach comparable with the conventional 3-class approach and usable for nature conservation purposes, including related reporting obligations, we also used 50 × 50 m grid cells.

Nadir artifacts were prominent in the KB and WRB area, with many false positive detections lying along the nadir, which had to be removed. Therefore, the track lines were buffered with two–2.5 m. Detections falling within these buffered areas were deleted. The data was additionally cleaned using the boulder distribution maps from Papenmeier (2024), which used the above-described standard procedure to classify 50 × 50 m grid cells. All detections in WRB and KB1 that fell in cells defined as “no boulders” were deleted to reduce false-positive detections.

In the MB and PTG sites, there was no need for removing nadir artifacts. However, since manually created boulder distribution maps by experts were not available and the model showed poor performance in boulder-free areas, manual data cleaning was required for further analysis. Figure 2 shows examples of false positive detections. For the cleaning process, a grid of 50 $\times$ 50 m was used. Detections that fell in cells that did not show actual boulder signatures were removed. This correction was applied cell-wise instead of removing individual detections.

Figure 2

Figure 2. Examples of false positive detections (blue boxes). Backscatter mosaics from the PTG study site with inset maps showing the location of each example. (A–D) Light grayish colors represent low backscatter, and black high backscatter intensities. (A) intensity artifacts along track edges, (B) seagrass, (C) effects of water column stratification, and (D) peat ridges.

2.5 Geospatial statistical analysis

To characterize the complexity and spatial variability of boulder fields, we analyzed their distribution using multiple geospatial statistical approaches that capture different aspects of their spatial organization.

Boulder abundance was quantified by counting individual boulders within 50 $\times$ 50 m grid cells across each study site. To determine regional differences in the boulder abundance, pairwise comparisons of boulder densities between regions were conducted using the Mann-Whitney U-Test (Mann and Whitney, 1947) with Bonferroni correction for multiple comparisons $(\alpha =0.05)$ , implemented in Python using scipy. stats (Virtanen et al., 2020).

To quantify the sharpness (S) of transitions between boulder-rich and boulder-poor areas, we analyzed the spatial gradients in boulder density. As a preprocessing step, the count values per cell were exported as GeoTIFFs for practical reasons. In the next step, the data were smoothed to reduce track artifacts while preserving important structural features in the data, allowing for more accurate delineation of boulder field boundaries. This smoothing was performed using the Anisotropic Diffusion filter from the Orfeo ToolBox (OTB) integration in QGIS (Grizonnet et al., 2017). The filter was applied with default parameter values (number of iterations = 10, time step = 0.125, conductance parameter = 1.0). Transition sharpness was calculated using the roughness algorithm from the QGIS Raster Terrain Analysis toolset. For each grid cell, the algorithm examines the eight neighboring cells in a 3 $\times$ three window and computes the difference between the maximum and minimum boulder counts within this neighborhood. Higher roughness values indicate sharper transitions between areas of different boulder densities.

The spatial distribution of boulders was characterized using the Clark and Evans nearest neighbor statistic, which provides insight into whether boulder distributions follow clustered or regular patterns (Clark and Evans, 1954). This statistic is defined as the ratio of the observed nearest neighbor distance to the expected nearest neighbor distance under complete spatial randomness. Nearest neighbor distances were calculated using the complete boulder dataset by applying the “Shortest line between features” algorithm from QGIS, which performs Cartesian calculations to measure distances between points. The distance to the nearest neighbor was then extracted for each point. The Clark and Evans statistics were computed on a cell-by-cell basis. Values $<$ one indicate clustered distributions, values $\approx$ 1 suggest random distributions, and values $>$ one indicate regular patterns. Grid cells containing only a single boulder were assigned a value of 1.

We propose combining these three metrics into a single index to describe the variability and complexity of boulder fields. This Boulder Field Complexity Index (BFCI) Equation 1 is defined as:

$\text{BFCI}=\frac{\sqrt[3]{\text{BC}}\times \sqrt[3]{\text{S}}}{\left(2\times \text{SDR}\right)}(1)$

where BC represents the boulder count value per 50 $\times$ 50 m cell, S the sharpness metric measuring the abruptness of transitions between boulder patches, and SDR is the spatial distribution ratio represented by the Clark and Evans statistic. Cube root transformation was applied to the boulder count and sharpness metrics to compress their dynamic range and reduce the disproportionate influence of outlier values, given that these parameters can reach values up to 550. The resulting product is divided by twice the spatial distribution ratio to scale the index to a manageable numerical range. This index was then calculated on a cell-by-cell basis. For better visualization, the BFCI was winsorized per site using the boxplot upper whisker criterion (Q3 + 1.5 $\times$ IQR) and visualized using Jenks natural breaks classification. Higher BFCI values indicate more complex boulder field structures. The index increases when more boulders are present, when transitions between boulder and non-boulder areas become sharper, and when boulders are more clustered.

To demonstrate the impact of each ratio on the index, false color RGB composite images were created with the QGIS tool gdalbuildvrt. The data was transformed identically to the index calculation $\sqrt[3]{BC},\sqrt[3]{S},(2\times SDR)$ . The transformed values were then normalized to the standard 0–255 RGB range and assigned to RGB channels (BC = red, S = green, SDR = blue).

3 Results

AI detections are shown for each study site exemplarily in Figure 3. Boulder abundance shows substantial variability across the study sites (Figure 4; Table 2). PTG exhibits the highest median boulder count (58) and the greatest range, with a maximum of 550 boulders per cell. MB demonstrates more moderate values, with a median count of 43 boulders per cell and a maximum of 219. KB presents a lower central tendency but maintains a considerable range, reaching up to 320 boulders per cell. WRB displays the lowest values across all statistical measures (median 10 and maximum 121). The results of the Mann-Whitney U-Test provide statistical evidence that the boulder count distributions differ significantly between all study sites, with all p-values $<$ 0.001.

Figure 3

Figure 3. Exemplary backscatter mosaics for each study site with AI-boulder detections. (A) Kiel Bay (KB), (B) Mecklenburg Bay (MB), (C) Plantagenet Ground (PTG), (D) Western Rönnebank (WRB). Red circles indicate detected boulders. Light grayish colors represent low backscatter, and black represents high backscatter intensities.

Figure 4

Figure 4. Boxplot of the count values per 50 × 50 m grid cell per study site.

Table 2

Table 2. Boulder count statistics per study site.

The correlation between boulder count and depth reveals a negative relationship across all study sites (Figure 5). Boulder abundance decreases consistently with increasing water depth, with the shallow-water areas showing the highest counts. The four study sites display distinct depth-count relationships that separate them from one another. The shallower areas exhibit greater variability in boulder counts, as evidenced by the wider spread of data points at lower depths compared to the more constrained distribution at greater depths. Notably, MB deviates from this trend, displaying a more constrained boulder distribution despite occurring in relatively shallow waters.

Figure 5

Figure 5. Correlation between water depth and boulder count.

The distribution of sharpness values demonstrates considerable variation in the transition of boulder-rich and boulder-poor areas across the four study sites (Figure 6). PTG exhibited the greatest spread of sharpness values with a maximum of 214 and the highest standard deviation (25), indicating pronounced transitions between boulder-rich and boulder-poor areas with considerable spatial variability. KB and MB display similar distribution patterns, both showing exponential decay from low sharpness values. KB exhibits a wider spread (median: 11.3, std dev: 14.1, max: 191) compared to MB (median: 10.3, std dev: 9.6, max: 91). Both regions are characterized by low sharpness values with occasional areas of higher variability. WRB exhibited lower sharpness values with a median of 4, a maximum of 39, and the narrowest range with a standard deviation of 3, indicating gradual transitions and homogeneous boulder distribution patterns.

Figure 6

Figure 6. Histograms representing the relative frequency of calculated sharpness per study site. (A) Kiel Bay (KB), (B) Mecklenburg Bay (MB), (C) Plantagenet Ground (PTG), (D) Western Rönnebank (WRB). The higher the sharpness value, the sharper the transition between boulder-rich and boulder-poor areas.

The distribution of nearest neighbor distances in Figure 7 shows that the peaks range from 1.9 m in PTG to 2.7 m in WRB. The skewness of the distribution varies between WRB and the other regions. The spread of boulder distances is greater in WRB. The Spatial Distribution Ratio (SDR) quantifies the spatial arrangement of boulders within each raster cell, indicating whether they exhibit clustered or regular distribution patterns. When correlating SDR with boulder count, all four study sites display similar overall patterns (Figure 8). Low boulder counts show the full range of SDR values, and there is a trend toward regular boulder distributions in cells with higher boulder counts. However, the relationship shows a weak positive correlation (Pearson r = 0.30, R ² = 0.09), indicating that boulder count explains less than 9% of the variance in spatial distribution patterns.

Figure 7

Figure 7. Relative frequency of nearest neighbor distances between boulders for each study site. Data capped at the 99th percentile (15 m).

Figure 8

Figure 8. The boulder count vs. spatial distribution ratio (SDR) for each study site. SDR $<$ 0.95 represent clusterd distribution, 0.95–1.05 random distribution, and $>$ 1.05 regular distribution.

The false-colour RGB composite image Figure 9 provides a visual representation of the individual impact of the three components on the BFCI. The color for each cell derives from the combined contribution of boulder count, sharpness, and SDR values. Panels A–C in Figure 10 show an example from the KB site representing these three metrics individually. Each metric contributes to the index by highlighting different spatial characteristics. Each study site exhibits characteristic color signatures that reflect different dominance of boulder density, transition sharpness, and spatial clustering. WRB displays purple/blue to magenta hues, indicating regularly distributed boulders combined with low to moderate boulder densities. However, some areas with sharp transitions stand out as yellow/green structures, and the boulder count increases in the central part of the WRB. In contrast, in the PTG area, red and green hues are most dominant, reflecting the region’s high boulder densities (red component) and sharp transitions between boulder patches (green component). KB shows a heterogeneous mosaic of colors, indicating variable importance of all three metrics across the region. MB presents red/magenta to purple tones, indicating moderate boulder densities with clustered spatial arrangements and smooth transitions between boulder fields. However, it does show some bright green structures indicating sharp edges of boulder-rich to boulder-poor areas.

Figure 9

Figure 9. False color RGB images for each study site showing the impact of each metric on the Boulder Field Complexity Index (BFCI). Channel red = boulder count (BC), green = sharpness (S), and blue = spatial distribution ratio (SDR).

Figure 10

Figure 10. A section from the southern part of the Kiel Bay (KB) study site showing the three metrics of the Boulder Field Complexity Index (BFCI): boulder count (A), sharpness (B), and spatial distribution ratio (C). The conventional three-class boulder count classification (D) is shown in comparison with the new BFCI (E). Panel (F) shows the false color RGB image, illustrating the impact of each metric from (A–C). Detailed color legends for (E,F) can be found in Figures 9, 12, respectively.

The BFCI shows great variability across all study sites, with values ranging from 0.14 to 14.0 (Figures 11, 12). PTG exhibits the highest median BFCI values (6.93) and the greatest variability (std = 2.63), indicating the most complex boulder field structures. In contrast, WRB displays the lowest median values (2.19) and the least variability (std = 1.02), indicating less complex and more homogeneous boulder fields. MB shows intermediate complexity (median 4.20) with moderate variability (std = 1.52), while KB presents similar median complexity (3.57) but higher variability (std = 2.21). Spatially, MB reveals a distinctive pattern divided into two boulder field areas separated by a substrate band with almost no boulders. The southern area shows higher complexity values while the northern area tends toward lower values, though two patches of elevated complexity occur in the northern section. KB depicts the most heterogeneous spatial distribution, with patches of high to very high BFCI values concentrated in the southeastern portion, transitioning to lower complexity values toward the northwest.

Figure 11

Figure 11. Boxplot for each study site showing the range of the Boulder Field Complexity Index (BFCI) calculated for each 50 × 50 m grid cell.

Figure 12

Figure 12. Maps with the calculated Boulder Field Complexity Index (BFCI) for each study site. The color scale is subdivided by the Jenks natural breaks classification method.

4 Discussion

In this study, we demonstrate that not only the number of boulders, but also their spatial arrangement, is important to describe the complexity of boulder fields. Several studies have highlighted the ecological significance of boulder characteristics in marine environments. Highly variable communities influenced by boulder properties have been described on boulder fields in intertidal and shallow subtidal zones, yet a dependence on boulder and rock parameters has not been established (Chapman, 2002a; b). Franz et al. (2021) have shown that larger boulders support higher species richness, while smaller boulders exhibit greater variability in community composition between individual boulders in coastal zones of the south-western Baltic Sea. Michaelis et al. (2019a) showed that the dominance of specific taxa is influenced by (among others) boulder size. However, a comprehensive examination of the relationship between boulder field complexities and their geogenic-biological interactions has remained limited.

The quantitative assessment of these relationships has been hindered so far by a lack of information regarding the occurrence and size distribution of individual boulders. AI-based boulder detection has improved substantially in recent years (Feldens et al., 2019; Michaelis et al., 2019b; Feldens et al., 2021; Hinz et al., 2024), paving the way for more detailed analysis of boulder distribution patterns (Feldens et al., 2023). By applying region-specific models, we achieved detection rates exceeding those of previous studies. For the PTG site, Model two identified an average of 70 additional boulders per cell — 6.7 times more than Feldens et al. (2023) shown exemplary in Figure 13. The maximum difference within 1 cell was 442 boulders. For the MB site, we achieved similar detection improvements; our model detected 7.2 times more boulders than the cited study.

Figure 13

Figure 13. The backscatter section of the Plantagenet Ground (PTG) study site shows the AI-detected boulder of this study in red, and the AI-detected boulder identified by Feldens et al. (2023) is shown in blue for comparison. Light grayish colors of the backscatter mosaic represent low backscatter, and black high backscatter intensities.

Following the conventional 3-class boulder classification approach (Papenmeier et al., 2020; Heinicke et al., 2021), which categorizes 50 × 50 m cells as containing 0 boulders, 1-5 boulders, or $>$ 5 boulders, 13% of cells in the PTG area and 39% of cells in the MB area needed to be reclassified as having more than five boulders. Such reclassification has important implications for reef delineation (Heinicke et al., 2021; Darr et al., 2022; Marx et al., 2024a; Marx et al., 2024b) and marine spatial planning. The limitations of this categorical system are evident in our study sites, where 80% of cells exceeded the highest category ( $>$ 5 boulders), effectively collapsing the majority of habitat variability into a single class and not taking into account the spatial arrangements of boulders.

We therefore developed a new measure that captures the structural complexity of boulder fields. The BFCI integrates three metrics derived from individual boulder detection: boulder abundance (with complexity increasing with higher values), transition sharpness (with complexity increasing with higher values), and spatial distribution pattern (complexity decreasing with higher values). The calculated BFCI values (0.14–14.0) demonstrate differences between study sites that were previously masked by categorical approaches. This continuous index provides much finer resolution than the conventional 3-class boulder classification approach (Papenmeier et al., 2020; Heinicke et al., 2021) and prior studies that introduced simple density categories or basic heatmap visualizations (Feldens et al., 2023).

The spatial complexity patterns captured by the BFCI reflect both geological history and ongoing hydrodynamic reworking processes (von Rönn et al., 2021; Hoffmann et al., 2022) which control habitat characteristics and availability of geogenic reefs. Different geological processes create distinct boulder field signatures. The correlation between boulder count and depth demonstrates that hydrodynamic reworking of the basal till is a relevant process at all study sites. The wave base of the Baltic Sea accelerates erosion of the basal till in shallow waters and forms abrasion platforms exposing mostly randomly distributed boulders (Schrottke et al., 2006; Schwarzer et al., 2014). Since the last glacial maximum and the formation of the modern Baltic Sea predecessors, wave-based erosion has occured below the present-day wave base. This has depended on the sea level of the Baltic Sea and it predecessors Baltic Ice Lake, Yoldia Sea and Ancylus Lake which was partially more than 20 m below present sea level Lampe (2002). Sediment erosion and the appearance of new boulders is an ongoing process, as demonstrated by Bohling et al. (2009) on the western coast of the Baltic Sea. They detected more newly exposed boulders in shallow waters. An indicator of erosional processes on abrasion platforms is the formation of scours around boulders, as exemplified by observations made at depths of at least 12 m in the shallow German Baltic Sea offshore Usedom (Schwarzer et al., 2003). Recently, the impact of ship traffic on the erosion of boulders out of glacial till was pointed out by Krämer et al. (2025). This erosional effect of wave and ship-induced turbulence decreases with increasing water depth, which is evident for the deepest study site WRB. The presence of less exposed boulders results in larger distances between them (Figure 7). More clustered accumulations can indicate other geological features like glaciofluvial landforms such as drumlins or ice-marginal ridges (Feldens et al., 2013). In contrast to boulders being randomly revealed by wave erosion, glaciofluvial processes actively deposited boulders over small areas. Finally, archeological relicts (Geersen et al., 2024) represent the accumulation of boulders by human activity in the past.

False-color composite imaging is a well-established technique in marine remote sensing, enabling a more detailed representation of seafloor characteristics (Tauber, 2007; Brown et al., 2019; Fakiris et al., 2019; Tamsett et al., 2019). We adapted this technique to illustrate the spatial influence of each of the three metrics on the BFCI (Figure 9). Beyond depicting mere abundance, the RGB composites also support to figure out where, for example, abrupt or gradual transitions exist. Further, they reveal additional patterns in the spatial distribution of boulders as captured by the SDR.

Areas with fewer boulders exhibit greater variability in SDR values, indicating more heterogeneous spatial arrangements ranging from clustered to regular patterns. As boulder density increases, SDR values converge toward above one, suggesting more regular than random distributions. While we would theoretically expect convergence at exactly one for truly random distributions, the slight deviation above unity likely reflects resolution limitations for stones in the direct vicinity (Hinz et al., 2024) rather than geological processes.

The MB site deviates from the observed boulder count water depth correlation as it does not exhibit the characteristic peak in boulder count at shallower depths. The model used for the MB site was trained on PTG data, which outperformed previous models (Feldens et al., 2023) as well as Model one when applied to MB data, but may not perform as well on MB data as it does on PTG data. However, this model limitation cannot explain the observed discrepancy in boulder counts at shallow depths between the PTG and MB sites, as the model shows good overall performance across the depth range at MB. The depth-specific nature of this discrepancy—where boulder counts differ primarily in shallow waters—most likely stems from underlying geological differences between the sites. Further investigation of the geological substrate and formation processes could provide more insights into this observation.

While the BFCI offers a novel and efficient approach for analyzing and characterizing large areas, certain limitations remain. A key limitation of the current implementation, based on side-scan sonar backscatter mosaics combined with YOLO object detection models, is its inability to capture the morphological complexity of hard substrates, specifically the size of individual boulders. First attempts to automatically measure boulder size and segment boulders based on multibeam echo sounder (MBES) surveys have been made by Hinz et al. (2024) and von Rönn (2021). Addressing this gap through integrated biological and geophysical surveys would substantially enhance our understanding of habitat-function relationships in geogenic reefs. Another area for further research is to gain a better understanding of how regional geology and landscape evolution affect boulder field characteristics as imaged by the BFCI.

5 Conclusion

This study has shown that the detection of individual boulders is essential for characterizing the complex distribution of hard substrates within soft-sediment seafloor environments. Recent advancements in AI-driven boulder detection enable the accurate identification of boulders across extensive seafloor areas in side-scan sonar imagery, facilitating large-scale statistical point pattern analysis. The BFCI describes spatial patterns of boulder fields by integrating boulder count, transition sharpness, and spatial distribution. The observed variability in boulder field complexity across sites underscores the need for continuous, quantitative metrics, as traditional categorical classifications fail to capture the spatial heterogeneity of these environments. We recommend that management and monitoring strategies incorporate the observed variability in boulder field complexity and use this knowledge as a foundation for future ecological assessments. However, the ecological responses associated with different boulder field complexities remain largely unexplored. Future research should therefore aim to investigate how benthic community composition changes with BFCI values across. For example, boulder fields with high BFCI values may support distinct community assemblages, while those with intermediate or low BFCI values may favor different communities. Edge zones with high BFCI values may act as ecotones, but whether these transitional areas host unique communities or simply represent species overlap needs further study. The ecological roles of patches with high BFCI values or of dispersed boulders as ‘stepping stones’ for species movement also require investigation. Similarly, the interaction between mobile species and varying boulder field complexities remains unexplored.

Data availability statement

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

Author contributions

AN: Conceptualization, Data curation, Formal Analysis, Investigation, Methodology, Visualization, Writing – original draft, Writing – review and editing. PF: Conceptualization, Investigation, Methodology, Resources, Supervision, Writing – review and editing. MZ: Conceptualization, Supervision, Writing – review and editing. SP: Conceptualization, Data curation, Formal Analysis, Investigation, Methodology, Project administration, Resources, Supervision, Visualization, Writing – review and editing.

Funding

The author(s) declared that financial support was received for this work and/or its publication. The data was provided from two projects: ATLAS “Mapping of habitats (biotopes) and their communities on the seabed in Mecklenburg-Western Pomerania” was funded by the LUNG MV, (grant number 220-01-Sc-19), and the project SedAWZ/AWZ-Project 6 “Area-wide high-resolution sediment mapping of the Exclusive Economic Zone (EEZ) of the North Sea and Baltic Sea” financed by the BSH and BfN (grant number 3522520500).

Acknowledgements

We would like to thank all captains and crew members of the multiple cruises on which the data used in this study were acquired. Great acknowledgment goes to Agata Feldens and Matthias Hinz for training the YOLO models we used for the boulder detection. Furthermore, we would like to thank our volunteer Svea Morawiak for her assistance with data processing.

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

Arosio, R., Hobley, B., Wheeler, A. J., Sacchetti, F., Conti, L. A., Furey, T., et al. (2023). Fully convolutional neural networks applied to large-scale marine morphology mapping. Front. Mar. Sci. 10, 1228867. doi:10.3389/fmars.2023.1228867

CrossRef Full Text | Google Scholar

Bohling, B., May, H., Mosch, T., and Schwarzer, K. (2009). Regeneration of submarine hard-bottom substrate by natural abrasion in the Western Baltic sea. Marbg. Geogr. Schriften 145, 66–79.

Google Scholar

Bressau, S. (1957). Abrasion, transport und sedimentation in der beltsee. Die Küste 6, 64–102.

Google Scholar

Brown, C. J., Smith, S. J., Lawton, P., and Anderson, J. T. (2011). Benthic habitat mapping: a review of progress towards improved understanding of the spatial ecology of the seafloor using Acoustic techniques. Estuar. Coast. Shelf Sci. 92, 502–520. doi:10.1016/j.ecss.2011.02.007

CrossRef Full Text | Google Scholar

Brown, C. J., Beaudoin, J., Brissette, M., and Gazzola, V. (2019). Multispectral multibeam echo sounder backscatter as a tool for improved seafloor characterization. Geosciences 9, 126. doi:10.3390/geosciences9030126

CrossRef Full Text | Google Scholar

Casabona, E., Wilms, T., Moltesen, M., Bertelsen, J. L., Kruse, B. M., Flávio, H., et al. (2024). Cobble reef restoration in the Baltic sea: implications for life below water. Aquatic Conservation Mar. Freshw. Ecosyst. 34, e4216. doi:10.1002/aqc.4216

CrossRef Full Text | Google Scholar

Chapman, M. G. (2002a). Patterns of spatial and temporal variation of macrofauna under boulders in a sheltered Boulder field. Austral Ecol. 27, 211–228. doi:10.1046/j.1442-9993.2002.01172.x

CrossRef Full Text | Google Scholar

Chapman, M. G. (2002b). Early colonization of shallow subtidal boulders in two habitats. J. Exp. Mar. Biol. Ecol. 275, 95–116. doi:10.1016/S0022-0981(02)00134-X

CrossRef Full Text | Google Scholar

Clark, P. J., and Evans, F. C. (1954). Distance to nearest neighbor as a measure of spatial relationships in populations. Ecology 35, 445–453. doi:10.2307/1931034

CrossRef Full Text | Google Scholar

Coolen, J. W. P., Bos, O. G., Glorius, S., Lengkeek, W., Cuperus, J., van der Weide, B., et al. (2015). Reefs, sand and reef-like sand: a comparison of the benthic biodiversity of habitats in the Dutch borkum reef grounds. J. Sea Res. 103, 84–92. doi:10.1016/j.seares.2015.06.010

CrossRef Full Text | Google Scholar

Darr, A., Heinicke, K., Meier, F., Papenmeier, S., Richter, P., Schwarzer, K., et al. (2022). Die Biotope des Meeresbodens im Naturschutzgebiet Fehmarnbelt. No. 636 in BfN-Schriften (632.2022 -) (Bundesamt für Naturschutz). doi:10.19217/skr636

CrossRef Full Text | Google Scholar

European Commission (1992). Council directive 92/43/EEC on the conservation of natural habitats and of wild fauna and flora (habitats directive), annex I, habitat type 1170: reefs. Tech. Rep. EU.

Google Scholar

European Commission (2008). Directive 2008/56/ec of the European parliament and of the council establishing a framework for community action in the field of marine environmental policy (marine strategy framework directive). Available online at: https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32008L0056 (Accessed April 11, 2025).

Google Scholar

European Environment Agency (2025). Eunis habitat mb13: rocky reefs and boulders. Available online at: https://eunis.eea.europa.eu/habitats/30537 (Accessed April 11, 2025).

Google Scholar

Fakiris, E., Blondel, P., Papatheodorou, G., Christodoulou, D., Dimas, X., Georgiou, N., et al. (2019). Multi-frequency, multi-sonar mapping of shallow habitats—efficacy and management implications in the national marine park of zakynthos, Greece. Remote Sens. 11, 461. doi:10.3390/rs11040461

CrossRef Full Text | Google Scholar

Feldens, P. (2020). Super resolution by deep learning improves Boulder detection in side scan sonar backscatter mosaics. Remote Sens. 12, 2284. doi:10.3390/rs12142284

CrossRef Full Text | Google Scholar

Feldens, P., Diesing, M., Wilken, D., and Schwarzer, K. (2013). Submarine eskers preserved on adler grund, south-western Baltic sea. Baltica 26, 137–144. doi:10.5200/baltica.2013.26.14

CrossRef Full Text | Google Scholar

Feldens, P., Darr, A., Feldens, A., and Tauber, F. (2019). Detection of boulders in side scan sonar mosaics by a neural network. Geosciences 9, 159. doi:10.3390/geosciences9040159

CrossRef Full Text | Google Scholar

Feldens, P., Westfeld, P., Valerius, J., Feldens, A., and Papenmeier, S. (2021). Automatic detection of boulders by neural networks: a comparison of multibeam echo sounder and side-scan sonar performance. HENRY Hydrogr. Nachrichten 119, 6–17. doi:10.23784/HN119-01

CrossRef Full Text | Google Scholar

Feldens, A., Marx, D., Herbst, A., Darr, A., Papenmeier, S., Hinz, M., et al. (2023). Distribution of boulders in coastal waters of Western pomerania, German Baltic sea. Front. Earth Sci. 11, 1155765. doi:10.3389/feart.2023.1155765

CrossRef Full Text | Google Scholar

Franz, M., von Rönn, G. A., Barboza, F. R., Karez, R., Reimers, H.-C., Schwarzer, K., et al. (2021). How do geological structure and biological diversity relate? Benthic communities in Boulder fields of the Southwestern Baltic sea. Estuaries Coasts 44, 1994–2009. doi:10.1007/s12237-020-00877-z

CrossRef Full Text | Google Scholar

Garone, R. V., Birkenes Lønmo, T. I., Schimel, A. C. G., Diesing, M., Thorsnes, T., and Løvstakken, L. (2023). Seabed classification of multibeam echosounder data into bedrock/non-bedrock using deep learning. Front. Earth Sci. 11. doi:10.3389/feart.2023.1285368

CrossRef Full Text | Google Scholar

Geersen, J., Bradtmöller, M., Schneider von Deimling, J., Feldens, P., Auer, J., Held, P., et al. (2024). A submerged stone age hunting architecture from the Western Baltic sea. Proc. Natl. Acad. Sci. 121, e2312008121. doi:10.1073/pnas.2312008121

PubMed Abstract | CrossRef Full Text | Google Scholar

Grizonnet, M., Michel, J., Poughon, V., Inglada, J., Savinaud, M., and Cresson, R. (2017). Orfeo toolbox: open source processing of remote sensing images. Open Geospatial Data, Softw. Stand. 2, 15. doi:10.1186/s40965-017-0031-6

CrossRef Full Text | Google Scholar

Halpern, B. S., Walbridge, S., Selkoe, K. A., Kappel, C. V., Micheli, F., D’Agrosa, C., et al. (2008). A global map of human impact on marine ecosystems. Science 319, 948–952. doi:10.1126/science.1149345

PubMed Abstract | CrossRef Full Text | Google Scholar

Heinicke, K., Bildstein, T., Reimers, H.-C., and Boedeker, D. (2021). Leitf. zur großflächigen Abgrenzung Kart. Leb. ”Riffe”in Dtsch. Ostsee, 612. doi:10.19217/skr612

CrossRef Full Text | Google Scholar

HELCOM (2013). “Technical report on the HELCOM underwater biotope and habitat classification,”. Helsinki Commission.Tech. Rep. Balt. Mar. Environmen Balt. Sea Environ. Proc. No. 139

Google Scholar

Hinz, M., Westfeld, P., Feldens, P., Feldens, A., Themann, S., and Papenmeier, S. (2024). AI-based Boulder detection in sonar data – bridging the gap from experimentation to application. Int. Hydrogr. Rev. 30, 78–98. doi:10.58440/ihr-30-1-a08

CrossRef Full Text | Google Scholar

Hoffmann, J. J. L., Michaelis, R., Mielck, F., Bartholomä, A., and Sander, L. (2022). Multiannual seafloor dynamics around a subtidal rocky reef habitat in the north sea. Remote Sens. 14, 2069. doi:10.3390/rs14092069

CrossRef Full Text | Google Scholar

Krämer, K., Becker, M., and Winter, C. (2025). Ship-induced morphology along a maritime waterway in a micro-tidal environment. Geomorphology 486, 109882. doi:10.1016/j.geomorph.2025.109882

CrossRef Full Text | Google Scholar

Kulha, N., Ruha, L., Väkevä, S., Koponen, S., Viitasalo, M., and Virtanen, E. A. (2024). Satellite bathymetry estimation in the optically complex northern Baltic sea. Estuar. Coast. Shelf Sci. 298, 108634. doi:10.1016/j.ecss.2024.108634

CrossRef Full Text | Google Scholar

Lampe, R. (2002). Post-glacial water-level variability along the south Baltic Coast–A short overview. Greifswald. Geogr. Arb. 27, 13–19.

Google Scholar

Lundine, M. A., Brothers, L. L., and Trembanis, A. C. (2023). Deep learning for pockmark detection: implications for quantitative seafloor characterization. Geomorphology 421, 108524. doi:10.1016/j.geomorph.2022.108524

CrossRef Full Text | Google Scholar

Mann, H. B., and Whitney, D. R. (1947). On a test of whether one of two random variables is stochastically larger than the other. Ann. Math. Statistics 18, 50–60. doi:10.1214/aoms/1177730491

CrossRef Full Text | Google Scholar

Marx, D., Feldens, A., Papenmeier, S., Feldens, P., Darr, A., Zettler, M. L., et al. (2024a). Habitats and biotopes in the German Baltic sea. Biology 13, 6. doi:10.3390/biology13010006

PubMed Abstract | CrossRef Full Text | Google Scholar

Marx, D., Romoth, K., Papenmeier, S., Valerius, J., Eisenbarth, S., and Heinicke, K. (2024b). Die Biotope des Meeresbodens im Naturschutzgebiet “Kadetrinne”. No. Version 1.0, Stand: 05.04.2024 in BfN-Schriften. doi:10.19217/skr690

CrossRef Full Text | Google Scholar

Matthäus, W. (2019). Die Anfänge der meeresgeologischen und Küstenforschung in Warnemünde (1950 – 1970). Tech. rep., Meereswiss. Berichte No 111 2019 - Mar. Sci. Rep. No 111 2019 Leibniz Inst. Baltic Sea Res. (IOW). doi:10.12754/MSR-2019-0111

CrossRef Full Text | Google Scholar

Michaelis, R., Hass, H. C., Mielck, F., Papenmeier, S., Sander, L., Gutow, L., et al. (2019a). Epibenthic assemblages of hard-substrate habitats in the German bight (south-eastern north sea) described using drift videos. Cont. Shelf Res. 175, 30–41. doi:10.1016/j.csr.2019.01.011

CrossRef Full Text | Google Scholar

Michaelis, R., Hass, H. C., Papenmeier, S., and Wiltshire, K. H. (2019b). Automated stone detection on side-scan sonar mosaics using haar-like features. Geosciences 9, 216. doi:10.3390/geosciences9050216

CrossRef Full Text | Google Scholar

Papenmeier, S. (2024). Map of Boulder distribution in the German eez. Available online at: https://gdi.bsh.de/mapapps/resources/apps/sedimentverteilung_auf_dem_meeresboden/index.html?lang=de (Accessed July 15, 2025).

Google Scholar

Papenmeier, S., and Hass, H. C. (2018). Detection of stones in marine habitats combining simultaneous hydroacoustic surveys. Geosciences 8, 279. doi:10.3390/geosciences8080279

CrossRef Full Text | Google Scholar

Papenmeier, S., Darr, A., Feldens, P., and Michaelis, R. (2020). Hydroacoustic mapping of geogenic hard substrates: challenges and review of German approaches. Geosci. 10. (3). doi:10.3390/geosciences10030100

CrossRef Full Text | Google Scholar

Rajani, H., Gracias, N., and Garcia, R. (2023). A convolutional vision transformer for semantic segmentation of side-scan sonar data. Ocean. Eng. 286, 115647. doi:10.1016/j.oceaneng.2023.115647

CrossRef Full Text | Google Scholar

Romoth, K., Darr, A., Papenmeier, S., Zettler, M. L., and Gogina, M. (2023). Substrate heterogeneity as a trigger for species diversity in marine benthic assemblages. Bio. 12. (6). doi:10.3390/biology12060825

CrossRef Full Text | Google Scholar

Schrottke, K., Schwarzer, K., and Fröhle, P. (2006). Mobility and transport directions of residual sediments on abrasion platforms in front of active cliffs (southwestern Baltic sea). J. Coast. Res., 459–464.

Google Scholar

Schwarzer, K., and Bohling, B. (2008). “Regeneration subaquatischer Steinfelder in der westlichen Ostsee,”Tech. Rep. Universität Kiel. doi:10.13140/2.1.1782.8641

CrossRef Full Text | Google Scholar

Schwarzer, K., Diesing, M., Larson, M., Niedermeyer, R. O., Schumacher, W., and Furmanczyk, K. (2003). Coastline evolution at different time scales – examples from the Pomeranian bight, southern Baltic sea. Mar. Geol. 194, 79–101. doi:10.1016/S0025-3227(02)00700-4

CrossRef Full Text | Google Scholar

Schwarzer, K., Bohling, B., and Heinrich, C. (2014). Submarine hard-bottom substrates in the Western Baltic sea – human impact versus natural development. J. Coast. Res. 70, 145–150. doi:10.2112/SI70-025.1

CrossRef Full Text | Google Scholar

Schwarzer, K., Ricklefs, K., Lohrberg, A., and Valerius, J. (2019). Die geologische entwicklung von nord-und ostsee. HENRY Hydraul. Engeneering Repos. doi:10.18171/1.087114

CrossRef Full Text | Google Scholar

Tamsett, D., McIlvenny, J., Baxter, J., Gois, P., and Williamson, B. (2019). On the information advantage of sidescan sonar three-frequency colour over greyscale imagery. J. Mar. Sci. Eng. 7, 276. doi:10.3390/jmse7080276

CrossRef Full Text | Google Scholar

Tauber, F. (2007). Seafloor exploration with sidescan sonar for geo-archaeological investigations. Ber. Römisch-Germanischen Komm. 88, 67–79.

Google Scholar

Virtanen, P., Gommers, R., Oliphant, T. E., Haberland, M., Reddy, T., Cournapeau, D., et al. (2020). Scipy 1.0: fundamental algorithms for scientific computing in python. Nat. Methods 17, 261–272. doi:10.1038/s41592-019-0686-2

PubMed Abstract | CrossRef Full Text | Google Scholar

von Rönn, G. A. (2021). Geological structure and natural dynamics of stone habitats in shallow waters of the Southwestern Baltic sea. Dissertation zur Erlangung des Doktorgrades an der Mathematisch-Naturwissenschaftlichen Fakultät. Dissertation, Kiel, Germany: Christian-Albrechts-Universität zu Kiel.

Google Scholar

von Rönn, G. A., Schwarzer, K., Reimers, H.-C., and Winter, C. (2019). Limitations of Boulder detection in shallow water habitats using high-resolution sidescan sonar images. Geosciences 9, 390. doi:10.3390/geosciences9090390

CrossRef Full Text | Google Scholar

von Rönn, G. A., Krämer, K., Franz, M., Schwarzer, K., Reimers, H.-C., and Winter, C. (2021). Dynamics of stone habitats in coastal waters of the Southwestern Baltic sea (hohwacht Bay). Geosciences 11, 171. doi:10.3390/geosciences11040171

CrossRef Full Text | Google Scholar

Wentworth, C. K. (1922). A scale of grade and class terms for clastic sediments. J. Geol. 30, 377–392. doi:10.1086/622910

CrossRef Full Text | Google Scholar

Wilms, T. J. G., Nordfoss, P. H., Baktoft, H., Støttrup, J. G., Kruse, B. M., and Svendsen, J. C. (2021). Restoring marine ecosystems: spatial reef configuration triggers taxon-specific responses among early colonizers. J. Appl. Ecol. 58, 2936–2950. doi:10.1111/1365-2664.14014

CrossRef Full Text | Google Scholar