Comparing Pixel-and Object-Based Approaches for Classifying Benthic Habitats

Comparing Pixel-and Object-Based Approaches for Classifying Benthic Habitats | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Comparing Pixel-and Object-Based Approaches for Classifying Benthic Habitats Henry O. J. Simmons, Oli Dalby, Daniel Ierodiaconou, Mary A. Young This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5351238/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Context Benthic habitat mapping is crucial for effective marine spatial planning. Despite advancements in multibeam echosounder (MBES) technology, selecting appropriate classification methods to accurately map seafloor habitats remains a challenge. Objectives This study aims to provide novel comparisons of large spatial scale habitat classifications using pixel-based (PB) and object-based image analysis (OBIA) methods, applied within a hierarchical random forest framework, to classify benthic biotopes in the northern section of Apollo Marine Park and the adjacent Cape Otway coastline, Victoria, Australia. Methods We utilised high-resolution MBES-derived data, implementing a hierarchical random forests algorithm to classify benthic habitats. The PB method treated each pixel independently, allowing for high spatial detail, while the OBIA method grouped pixels into meaningful segments for classification. Prior to segmentation, backscatter data from two different MBES systems were harmonised using a bulk shift method (Misiuk et al., 2020 ) to ensure consistency across datasets. We then applied the Supercells segmentation technique (Nowosad 2022 ) to the harmonised backscatter data, forming the foundation for the OBIA-based classification. Both methods were evaluated using accuracy, F1 scores, and uncertainty maps were generated to assess classification reliability. Results Both classification methods demonstrated strong performance, with no statistically significant differences in overall accuracy. However, the complexity of the habitat maps varied: the PB approach excelled in capturing fine-scale habitat details, beneficial for management and conservation efforts requiring high detail. Conversely, the OBIA method produced more interpretable and less complex maps, suitable for general spatial analyses, though it resulted in the omission of some minority classes. Conclusion This study emphasises the importance of defining the desired level of complexity in habitat maps before analysis, ensuring that chosen methods yield maps suitable for specific applications—particularly in datasets with strong class imbalances. Future advancements in machine learning and emerging technologies have the potential to further refine habitat mapping techniques and enhance classification accuracy. Benthic Habitat Classification Remote Sensing Hierarchical Random Forest Backscatter Segmentation GIS Marine Spatial Planning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Introduction The global ocean accounts for 71% of the Earth’s surface, plays a central role in the structure and functioning of our biosphere and is critical for achieving the sustainable development of human society (Field et al., 1998 ; Mayer et al., 2018 ; von Schuckmann et al., 2020 ). Marine ecosystems provide several ecosystem services that include but are not limited to climate regulation (Reid et al., 2009 ), coastal protection (Spalding et al., 2014 ), biodiversity (Gamfeldt et al., 2015 ), fisheries and aquaculture (Beardmore et al., 1997 ; Fontoura et al., 2022 ) and help ensure food security and economic benefits for millions (The and Sumaila 2013). Additionally, oceans offer recreational and cultural value, enriching human lives and sustaining coastal communities (Martin et al., 2016 ). Without healthy marine ecosystems, these services become compromised. Vital marine ecosystems like coral reefs, kelp forests, and seagrass meadows are experiencing losses worldwide due to increased sea surface temperatures (Ramírez et al., 2017 ), ocean acidification (Barry et al., 2011 ), habitat destruction (Nikolaou and Katsanevakis 2023 ), and pollution (Buonocore et al., 2021 ) (Wernberg et al., 2019 ; McKenzie 2020; Eddy et al., 2021 ). The majority of these pressures can be linked to anthropogenic sources. Despite these challenges, there is hope. On an international scale, Goal 14 of the United Nations Sustainable Development Goals focuses on conserving and sustainably using the marine resources (Secretary-General 2017 ; Monoly et al., 2022). Whilst nationally, Australia’s Marine Science Plan has been established (National Marine Science Committee, 2015 ) with the aim to support sustainable management and conservation of marine resources by improving access and allocation of fisheries, enhancing productivity, maintaining habitat health, and addressing climate change impacts. The plan also calls for integrated efforts from government, industry, and the scientific community to support sustainable development and conservation of marine biodiversity​. Recognition of the need to protect marine ecosystems is not a recent development; it has been a priority for decades (Ward et al., 2022 ). Within Australia, the state of Victoria has a strong history of marine conservation. In 2002, Victoria implemented one of the world’s first representative systems of fully protected marine areas (Victoria 2003 ), increasing the states “no-take” Marine Protected Areas (MPAs) 100-fold to cover 5% of its coastal waters (Wescott 2006 ). In 2013, Australian national conservation efforts were expanded with the creation of the Southeast Commonwealth Marine Reserves Network, comprising 14 marine parks that allowed for regulated and sustainable commercial and recreational use (Zone and Zone 2013 ). These reserves are a part of Australia's commitment to the conservation and sustainable management of its marine environment, aiming to balance conservation goals with responsible resource use. Each reserve includes specific zoning and management plans tailored to its unique ecological and socio-economic context. However, effective marine park management requires detailed assessments of the distribution and abundance of communities within and beyond their boundaries. This can be complicated by several issues. For example, collecting in-situ data can be arduous, expensive, and time-consuming, especially in waters exposed to dangerous sea conditions. Consequently, remote sensing and modelling platforms have become essential tools for researchers tasked with mapping and monitoring marine environments and resources (Melesse et al., 2007 ; Sanders and Masri, 2016 ; Strong and Elliot, 2017; Fingas, 2019 ). By integrating remote sensing technologies with geospatial analyses and recent improvements in computing power, researchers can generate high-resolution maps of marine habitats. These maps can depict the presence, condition, and spatial arrangement of habitats present and can be repeated over time to identify trends in community composition and habitat condition. This, in turn, provides critical baseline data to inform effective marine reserve management and conservation efforts (Cogan et al., 2009 ; Brown et al., 2011 ; Buhl-Mortensen et al., 2015 ). In recent years, evolution of habitat mapping techniques has shifted from traditional pixel-based (PB) approaches to more advanced Object-Based Image Analysis (OBIA) methods (Benz et al., 2004 ; Lu and Weng 2007 ; Belgiu and Drăguţ 2016 ). PB approaches rely on individual pixel values to classify and map habitats based on spectral signatures. However, PB methods often struggle with mixed pixels and spectral confusion in complex environments, leading to a ‘salt and pepper’ effect when classifying (Turner et al., 2003 ). The development of OBIA in the late 20th century represented a paradigm shift by considering spatial relationships and context between pixels (Blaschke 2010 ). Specifically, OBIA employs segmentation algorithms to group pixels into meaningful objects or regions based on spectral, spatial, and contextual information (Hay et al., 2001 ), allowing for more precise delineation of habitat boundaries and characteristics. OBIA has since become prominent in ecological studies and habitat mapping, enhancing the accuracy and applicability of remote sensing data in environmental management and conservation efforts (Grebby et al. 2016 ; Esetlili et al. 2018 ). Despite advancements in the production and application of OBIA, several knowledge gaps remain. There has been little comparison between the application of OBIA vs PB approaches in benthic habitat classifications. In one paper, Ierodiaconou et al. ( 2018 ) compared the two methods in a relatively shallow, largely homogeneous, and spatially small cove (0.39km 2 ) in Victoria, Australia, where only five different habitat types were identified. Apart from this paper, comprehensive comparisons between OBIA and PB approaches have been sparse and typically limited to satellite data focused on shallow coastal waters, wetlands, or terrestrial areas (Whiteside et al., 2011 ; Wahidin et al., 2015 ; Berhane et al., 2017 ; Anggoro et al., 2018 ). As such, more examples comparing OBIA and PB approaches that classify acoustic data (e.g. MBES or SSS) are required to assess trade-offs between approaches. Furthermore, the application of hierarchical machine learning approaches remains unexplored. Crucially, the decision between using PB or OBIA should be informed by the specific management goals of the study. The interpretability of each method varies depending on the level of detail required for different management or conservation objectives. Previous studies have noted differences in accuracy, scalability, and practical applications between the two approaches (Whiteside et al. 2011 ; Nursamsi et al. 2024 ), but little research has specifically examined how these goals influence the choice of methodology, particularly in benthic habitat classification. Exploring these aspects could provide clearer guidance on which technique might best serve conservation or resource management needs, helping to ensure that data collection is fit for purpose in informing decision-making. In this study, we aim to address these knowledge gaps by conducting a comparative analysis of OBIA and PB approaches using high-resolution bathymetric data within Apollo Bay Marine Park, VIC, Australia. Specifically, we seek to compare the accuracy, interpretability, and applicability of PB and OBIA approaches in benthic habitat mapping, using this locale as a test case. The research not only aims to contribute to the field of marine habitat mapping but also aims to provide valuable insights for effective marine spatial management and conservation strategies in marine parks. Methods Study Area The Apollo Marine Park (AMP) was established in 2007 and incorporated into the Southeast Commonwealth Marine Reserves Network in 2013 (Figure 1). Located three nautical miles south of Cape Otway in Victoria, Australia, the park covers 1184 km² of the Bass Strait continental shelf, with depths ranging from less than 50 m near the Cape to deeper waters in the Otway Depression, an ancient river valley from the Last Glacial Maximum (~21 ka) (Wass et al., 1970; Amini et al., 2004). The AMP is exposed to strong south-westerly swells and tidal flows, resulting in high geoform complexity at the southern tip of the Cape. Biological observations indicate that mesophotic reefs and non-reef forming benthic assemblages dominated by sponges, as well as complex rock habitats with sea plumes and hydroid fans, extend from the north into the park (VEAC, 2019). The AMP supports diverse marine life, including seabirds, dolphins, seals, and white sharks, and serves as a critical migration route for blue, fin, sei, and humpback whales (Gill, 2002). Despite its ecological significance, there is limited knowledge of the species and habitats beyond the water surface. Acoustic Data Acquisition and Processing Geophysical data were collected between 2006 and 2021 by the Marine Mapping Group from Deakin University, Victoria, Australia, across several multibeam surveys, using varied equipment. Historic multibeam data were acquired in November and December of 2006 and 2007 respectively as part of the ongoing Victorian Marine Habitat Mapping Project. A Reson Seabat 8101 multibeam echosounder (MBES) (operating 240 kHz, 150° angular sector coverage) was used to collect data before real-time quality control, display, navigation, and post-processing were carried out using the Starfix Suite 7.1 (Fugro proprietary software). Contemporary multibeam surveys were also conducted within the northern sector of Apollo Marine Park between 2020 and 2021 using a Kongsberg Maritime EM2040C MBES (operating frequency of 300 kHz, 60° angular sector coverage). Positioning and real time positioning was completed using an Applanix POS MV Wavemaster, and a dual-frequency Trimble Global Navigation Satellite System (GNSS) receiver. For all sets of surveys, a sound velocity profiler was used regularly to correct the sonar data for local variations in sound speed. Twenty-nine linear km of data were acquired on and outside the parks northern boundary to map subtidal components of the Otway ridge, fill data gaps, and combine it with previous data collection initiatives. MBES data from both surveys were used to produce a bathymetric and backscatter grid at 3 m resolution. Given the considerable time difference between surveys, variable multibeam technologies used, and differing technical specifications a normalisation procedure was needed to harmonize the two datasets. A mosaic was created for the bathymetry dataset, as measured depth did not vary significantly between the two systems. Backscatter data was combined using the ‘bulk shift’ method Misiuk et al., (2020) which performs relative calibration of sonar backscatter datasets by modelling the error between overlapping surveys and adjusting the values to minimise discrepancies. This ensures consistent and accurate backscatter data across different surveys, facilitating seamless seabed mapping. The two datasets were then clipped to the study area (Fig. 1). Multibeam Derived Variables (Derivatives) Thirty-two derivatives from the bathymetry data were calculated to provide information on seafloor terrain attributes (Table 1) using the R package MultiscaleDTM (version 0.8.3, Ilich et al., 2023). MultiscaleDTM was chosen because it allows for the calculation of several terrain attributes within each of the five thematic groups (slope, aspect, curvature, relative position, and roughness) using a variable window size. Various neighbourhood scales were used for each terrain attribute using the 3m resolution harmonised bathymetry data (3, 5, 9, 21, 51, and 101 cell neighbourhoods). A full outline of measures of seafloor structure calculated are included in Table 1. Table.1 . These derivatives are calculated using different functions within the MutiscaleDTM (Ilich et al. 2023) package including multiscale slope and aspect (SlpAsp), quadratic surface to a rectangular focal window (Qfit), relative position (RelPos), difference from mean value (DMV), topographic position index (TPI), bathymetric position index (BPI), vector ruggedness measure (VRM), surface area to planar area ratio (SAPA), standard deviation of depth values after removing the influence of slope (AdjSD), and roughness index-elevation (RIE). For a full overview of these functions see Ilich et al. (2023). Variable Description Slope The maximum rate of change in elevation values. Aspect The orientation of the slope in the downslope direction. Eastness The east/west component of the orientation of the slope are calculated as the sine of aspect. Thus, it ranges between -1 (due west) and 1 (due east). Northness The north/south components of the orientation of the slope are calculated as the cosine of aspect. Thus, it ranges between -1 (due South), and 1 (due North) Profile curvature The level of convexity or concavity along the direction of the maximum slope. Planform curvature The level of convexity or concavity perpendicular to the direction of the maximum slope. Twisting curvature The amount of local twisting is measured as the change in slope angle per unit distance along the direction perpendicular to the slope. Minimum curvature the minimum curvature in any plane. Maximum curvature The maximum curvature in any plane. Mean curvature The average of the minimum and maximum curvatures. Relative position An indication of whether an area is a local topographic high (positive values) or low (negative values). It can be measured in either units of the DTM or standardised based on local topography to be a unitless measure. Vector ruggedness measure The dispersion of unit vectors normal to the terrain surface. Surface area to planar area ratio The surface area divided by the planar area. Adjusted standard deviation The local standard deviation of elevation values in a focal window after removing the influence of slope by fitting a plane to the focal window using ordinary least squares and extracting the residuals. Roughness index-elevation The local standard deviation of the residual topography surface, where the residual topography surface is calculated as the DTM minus the focal mean of the DTM. Supercells Object-Based Image Analysis The R package ‘supercells’ (Nowosad 2022) was used to create ‘superpixels’ (groups of spectrally similar pixels) using the Simple Linear Iterative Clustering (SLIC) algorithm (Achanta et al., 2012). SLIC uses an adapted k-means clustering approach modified to work with arbitrary dissimilarity values to efficiently generate superpixels. Superpixels were generated using the backscatter layer as an indicator of seafloor variation using a minimum superpixel size of 3 cells. This size was chosen to prevent the amalgamation of fine-scale habitats into larger, more generalised ones, thereby retaining the heterogeneity of the input data. All other settings were left at default values. Ground Truth Data Towed Video 2006-2009 Acoustically located towed video was used to collect observational data from various habitat mapping studies conducted between 2006 and 2009. Sampling was designed to capture a range of depths and topographic and textural diversity across the region. A total of seven transects perpendicular to the coast (~19.5 linear kilometres) were collected. The system was piloted via a live-stream video at an average speed of 1-1.5 knots and at approximately 1-2m above the seabed with a 45° angle to the seabed. The resulting field of view was approximately 2 x 2 m. An ultra-short baseline transponder attached to the video unit allowed 3-dimensional positioning relative to the vessel’s dGPS antenna, resulting in a total propagated error of ± 3.6 m for video seafloor positioning (Rattray et al. 2014). Video data were classified for biotope using the Combined Classification Scheme (CBiCS; see section below and table.2 ). Baited Remote Underwater Video Stations (BRUVS) 2023 BRUVS were deployed to assess the diversity, abundance, distribution, and habitat use of marine taxa. Between the 10th and 14th of May 2023, a total of 104 BRUV deployments were conducted within Apollo Marine Park, ranging from 52 to 94 meters depth. Each BRUV array consisted of a weighted stainless-steel frame with two independent stereo-video systems positioned at 90° from the seabed facing forward. Each system included a GoPro Hero7 Black action camera, with pairs of cameras fitted 0.7m apart and angled in at 8 degrees to allow for stereo imaging. BRUV frames were calibrated prior to deployment. From each deployment a still image was extracted presenting a clear view of the surrounding habitat area. Benthic habitats were classified for each cell within a 20-cell grid (4 rows x 5 columns) overlaid on the still image, resulting in a single habitat classification for each site using the CBiCS classification. These habitats were annotated using TransectMeasure (Version 4.21, ® SeaGIS). Autonomous Underwater Vehicle (AUV) 2023 The AUV Nimbus was deployed in September and October 2023 to collect fine-scale imagery of the seabed over representative but spatially discrete areas. AUV deployment sites were selected using a feature space representation of bathymetric data combined with robotic information gathering to plan surveys that comprehensively explored the feature space (Shields et al., 2021). This process created a survey plan comprising 20 missions throughout the AMP. Each mission included three transects of 500 m, split by a 200 m parallel connecting transect. Managed by the Integrated Marine Observing System (IMOS) Autonomous Underwater Vehicle facility at the University of Sydney, the AUV navigated a predetermined course at 2 metres above the seabed and took continuous stereo photographs at 1 photo per second. Each stereo image covered an area of 1.8 x 1.5 m at a resolution of 1360x1024 pixels. More than 150,000 images were taken across the 20 missions. Habitat types were classified for every 50th AUV image post-processed using SQUIDLE+ (Proctor et al., 2018), using 25 randomised point counts for habitat and morphotype categories (Perkins et al. 2016) and later translated into CBiCS to match the other habitat data. Habitat Classifications Scheme The Combined Biotope Classification Scheme (CBiCS; Edmunds and Flynn 2018) is a hierarchical system used to classify and categorise marine habitats based on environmental and biological characteristics. It integrates components from the Joint Nature Conservation Committee-European Nature Information System (JNCC-EUNIS) and the United States Coastal and Marine Ecological Classification Standard (CMECS), emphasising factors such as biota, substrate, geoform, exposure, and biogeographic area. Beyond habitat types, CBiCS offers crucial ecological context by describing the communities within these habitats, thereby facilitating standardised habitat classification for ecological studies and conservation efforts. Data Preparation To prepare the ground truth data for modelling, standard point data extraction methods were utilised for both PB and OBIA approaches. Prior to modelling raster layer values were extracted at ground truthing locales. Within PB approaches this was completed using the pixel value whilst for the OBIA approach the superpixel value was used. If a supercell contained more than one biotope label, the most frequent label was selected, or, in case of a tie, the least represented label in the full dataset was chosen. This resulted in a significant reduction in the number of data points for the OBIA models compared to the PB approach. Sample sizes for either approach is shown in Table 2. Once prepared, data were split into training and testing sets with an 70-30 ratio. Model Approach Hierarchical random forests models were used to predict class membership. Random Forest (RF) is an ensemble machine learning method that combines tree-based classifiers with bootstrap aggregation (Breiman 2001) and mitigates the tendency of individual decision trees to overfit by aggregating results from multiple trees trained on random subsets of the data (Cutler et al., 2007). RF minimises tree correlation through random variable selection at each split, reducing biases associated with the training data (Rigatti 2017). RF models have proven effective in mapping marine substrates (Misiuk et al., 2019; Gregr et al., 2021; Dalby et al. 2024) and identifying dominant biological communities (Rubidge et al., 2016; Martinez-Santos et al., 2021). Hierarchical random forests (HRF) use predictions from previous models as predictor variables in subsequent models. RFs were created using the 'ranger' package in R (version 0.16.0). Using HRFs allowed the nested nature of habitat classifications to be modelled explicitly. This approach can capture dependencies between habitat types and their subcategories, providing a more accurate and nuanced prediction. For example, High energy circalittoral rock with seabed covering sponges can only occur on Circalittoral rock and other hard substrata , and so knowing the distribution of circalittoral rock aids predictions of the distribution of sponge reefs. To optimise the RF model, we tuned the algorithm by varying the following hyperparameters (1) the number of trees (50, 100, 500, 600, 1000), (2) the method used to select the number of features at each split (Square root and Log functions), and (3) the impurity functions (Gini coefficient and Entropy). The best combination of hyperparameters (those that maximised the F1 score) were then retained as the chosen model for that level in the HRF. Variable Importance Variable importance in Random Forests is quantified as the mean decrease in overall model accuracy when the values of a focal variable are permuted, and all other variables are held constant (Archer and Kimes 2008). This helps identify predictors with the greatest impact on the model performance. In the context of HRF, variable importance is assessed for both approaches, providing a more nuanced understanding of why certain variables are more critical than others (Fig.S1 & S2.). During analyses, 48 predictors that demonstrated less than 5% variable importance were removed as including too many variables, especially those with minimal importance, can introduce noise into the model, potentially leading to overfitting and reducing interpretability (Galelli et al., 2014; Gregorutti et al., 2017). Accuracy Assessments and Uncertainty Accuracy for both PB and OBIA models was assessed using confusion matrices, overall accuracy, precision, recall, and F1 scores. Additionally, the classifications were visually compared as habitat maps to emphasise interpretability, which is a crucial factor in evaluating the effectiveness of classifiers. The class-based confusion matrix along with uncertainty maps were also used to identify class accuracy imbalances and where these imbalances appear in space. This dual approach ensured a comprehensive assessment of model performance, balancing quantitative metrics with practical visual analysis to gauge the accuracy and reliability of habitat classifications. By analysing the results from both models, we aimed to gain deeper insights into their practical effectiveness, identify potential areas for improvement, investigate the effects of data loss in our methodologies, and explore how each method contributes to the overall goal of accurate and comprehensive benthic habitat mapping. In a classification task, Random Forests provide not only class predictions but also prediction probabilities for each class. This capability allows us to calculate and visualise the uncertainty associated with each prediction. By plotting these measures spatially, we can create uncertainty or confidence maps that highlight areas of high and low accuracy. Lower uncertainty typically corresponds to higher accuracy, making these maps useful for identifying regions where predictions are reliable or not (Mitchell et al. 2018). By examining these maps, researchers can better understand the reliability of the classification and identify where additional validation or refinement may be needed (Wager et al. 2014). Results Mapping of biotopes revealed that Apollo Marine Park and the adjacent Otway coastline was dominated by sand substrata, with sponge and macroalgae communities to the north and northwest (Fig.5). Increased biodiversity in north and northwestern areas coincided with the shallower depths and the presence of a hard seafloor. Paired t-tests showed no significant differences in precision, recall, or F1-score between the PB and OBIA models across BC levels and found only negligible differences in accuracy (See Table 2). Table 2 The results of the paired t-tests, and a comparison of accuracy, for PB and OBIA approaches at each hierarchical level, where a = 0.05. BC2 BC3 BC4 Precision t2 = -0.99, p = 0.42 t2 = -1.12, p = 0.34 t2 = 1.37, p = 0.21 Recall t2 = -0.15, p = 0.89 t2 = 0.66, p = 0.60 t2 = 1.07, p = 0.32 F1-score t2 = -1.10, p = 0.39 t2 = 0.49, p = 0.66 t2 = 1.7, p = 0.14 Accuracy Difference 0.015 0.014 0.001 Comparison of PB BC2 and OBIA BC2 Models Table 3 Comparison of Precision, Recall, F1-Score, and Accuracy between Object-Based Image Analysis (OBIA) and Pixel-Based (PB) random forest models for biotope classification at the BC2 level of hierarchy. OBIA PB Precision 0.76 0.88 Recall 0.88 0.89 F1-score 0.79 0.86 Accuracy 84% 85% When compared at the BC2 level, model accuracy was nearly identical (Table 3). Both models predict three main habitat classes: Circalittoral rock, Infralittoral rock, and Sublittoral sediment . The spatial extent of these classes in the PB map appears more fragmented, with smaller patches, indicative of the PB model’s tendency to produce more granular “salt-and-pepper” patterns, especially in transitional zones between classes. In contrast, the OBIA model predicted classes in broader, more continuous patches, reflecting a more generalized and smoother representation of biotopes present. Supplementary Figure 3 visualised differences in areal extents predicted for each biotope. Although both the PB and OBIA models predicted similar spatial extents of key biotopes, differences were present when assessed at the fine scale. The OBIA model predicted slightly larger extents for Circalittoral rock and Infralittoral rock, while the PB model predicted a slightly greater coverage of Sublittoral sediment. Further, the PB model had a prediction ranging from 0.33 to 1, with a mean of 0.80. This was marginally more uncertain than the equivalent OBIA model (rand of 0.35 to 1, mean of 0.86). Comparison of PB BC3 and OBIA BC3 ModelS Table 4 Comparison of Precision, Recall, F1-Score, and Accuracy between Object-Based Image Analysis (OBIA) and Pixel-Based (PB) random forest models for biotope classification at the BC3 level of hierarchy. OBIA PB Precision 0.93 0.79 Recall 0.87 0.92 F1-score 0.88 0.82 Accuracy 93% 92% As with BC2 predictions, the accuracy of PB and OBIA BC3 model was largely comparable (Table 4). Although, marginal difference was found in precision between the model types. Both models predict five habitat classes: Sublittoral course sediment, High energy open coast circalittoral rock, High energy infralittoral rock, Sublittoral sand and muddy sand, Sublittoral seaweed on sediment. Mirroring the results from the BC2 level of hierarchy; the PB model predicted more fragmented biotope classes with stronger “salt-and-pepper” effects that were particularly noticeable in transitional zones between biotopes. In contrast, the OBIA model presents even more pronounced generalisation, with broader and more continuous patches. This generalised representation is particularly evident in the shallow, diverse habitats, where the OBIA model predicts a much less complex arrangement of biotopes (Fig.7.). At the BC3 hierarchical level, the pixel-based (PB) model shows an uncertainty range of 0.20 to 1, with a mean of 0.91, indicating a high level of confidence in its predictions. The OBIA model, at this level, has an uncertainty range of 0.26 to 1 and a mean of 0.88, reflecting slightly lower but still high confidence. Supplementary Figure 4 illustrates that both models maintain similar spatial patterns of confidence. Compared to the previous level, this higher hierarchical level shows increased overall confidence while preserving the same general patterns across both approaches. Comparison of Pixel-based BC4 and OBIA BC4 Models Table 5 Comparison of Precision, Recall, F1-Score, and Accuracy between Object-Based Image Analysis (OBIA) and Pixel-Based (PB) random forest models for biotope classification at the BC3 level of hierarchy. OBIA PB Precision 0.69479 0.74242 Recall 0.67942 0.72448 F1-score 0.68169 0.72957 Accuracy 82.6% 84.4% As observed at previous hierarchical levels, the accuracy of the PB and OBIA BC4 models is largely comparable (Table 5). Both models predicted the following nine habitat classes: Circalittoral fine sand, High-energy circalittoral rock with seabed covering sponges, High-energy Durvillaea communities, Low-complexity circalittoral rock with non-erect sponges, High-energy lower infralittoral zone, Moderate-complexity circalittoral rock with prominent sea plumes, sea tulips, and hydroid fans, thallose red algae on sand with ground surge, sandy low-profile reef wave surge communities, and infralittoral shell and sand mixes. Notably, the OBIA model fails to identify High-energy Durvillaea communities , which are captured by the PB model. The PB model predicted more fragmented biotope classes with a stronger “salt-and-pepper” effect, particularly noticeable in transitional zones between biotopes. In contrast, the OBIA model showed more pronounced generalization, resulting in broader and more continuous patches. These differences were particularly evident in shallow, diverse habitats, where the OBIA model predicted a much less complex arrangement of biotopes (Fig. 9). At the final hierarchical level, the PB model displayed an uncertainty range of 0.17 to 1, with a mean of 0.89, indicating a high level of confidence in its predictions. Similarly, the OBIA model exhibited an uncertainty range of 0.20 to 1, with the same mean of 0.89, reflecting comparable overall confidence. Supplementary Figure 5 shows that both models continue to display similar spatial patterns of confidence. This level maintains high confidence, consistent with previous hierarchical levels, while preserving general patterns across both approaches. Visual comparisons of biotope classification between the PB and OBIA models reveal differences at both broad and fine scales. The highlighted areas illustrate how the models vary in their predictions, particularly in regions with higher heterogeneity. The OBIA model tends to capture broader biotope categories, while the PB model offers more detailed and nuanced classifications, leading to discrepancies in area coverage for certain biotopes. These differences are evident in the zoomed-in sections, which demonstrate how each model performs in predicting biotope distributions. This comparison provides insights into the varying levels of detail and coverage between the models. Discussion In this study, we developed benthic biotope maps for a northern section of Apollo Marine Park and the adjacent Cape Otway coastline, VIC, Australia. Maps were created using habitat/environment relationships derived from hierarchical random forest models with either PB or OBIA sets of MBES-derived input features. Our findings revealed that the study site is dominated by shell and sand mixes, with increased biodiversity in sponge and macroalgae communities towards the north and northwest, corresponding with geoform complexities and hard-bottom seafloors. We compared various metrics such as the accuracy and interpretability of the hierarchical random forest classifications. While the PB approach achieved a higher overall accuracy than the OBIA approaches, the differences were not statistically significant. The major differences between the models' classifications are evident when examining specific areas of interest (see Fig. 10 ). Notably, the ‘salt and pepper’ effect in the PB map contrasts with the more homogenous outputs of the OBIA map, which provides more clearly defined habitat boundaries. Analysing these results and their underlying causes highlight the weaknesses and strengths of both the PB and OBIA approaches. Across the three classification levels presented in this study, generally, the PB approach demonstrates higher average precision and recall, leading to better overall performance metrics, particularly in achieving a more refined and detailed classification of biotopes. This can be attributed to their fine-grain approach, evaluating each pixel independently and capturing subtle variations in habitat characteristics, making them effective for detecting and correctly classifying a broader range of biotopes on a smaller scale. Although the majority of studies comparing OBIA and pixel-based (PB) classification tend to favour OBIA due to its ability to group pixels into meaningful objects, there are some studies that have reported PB methods outperforming OBIA under specific conditions. For example, research from Universiti Teknologi MARA (Othman et al., 2023 ) found that PB methods outperformed OBIA for coral reef mapping off Mabul Island, Malaysia, using Sentinel-2B and SPOT-7 imagery. The PB approach yielded higher accuracy (97.5% for Sentinel-2 and 90% for SPOT-7) compared to OBIA (82.81% and 87.05%, respectively). Othman explained that this is likely due to PB’s ability to capture fine-scale habitat details in highly complex and heterogeneous environments. Additionally, Blaschke ( 2010 ) emphasized that while object-based image analysis (OBIA) has its advantages in some situations, pixel-based methods are essential for accurately capturing detailed spatial patterns, especially in high-resolution environments.". This study underscored the advantages of PB methods in achieving refined habitat boundaries and detecting subtle habitat features. These findings highlight the robustness of PB methods in delivering high-resolution habitat maps, which is critical for effective marine spatial planning and management. In contrast, OBIA models offer more streamlined classifications with a tendency towards broader habitat categories. This understanding of OBIA models is mirrored in Fig. 10 within the results section. Although there are no significant differences identified in the PB and OBIA approach’s performance metrics, there are clear differences in their classification approaches, whereby the OBIA model classifies with a far more general, or broader, brush than the PB approach. OBIA’s strengths lie in object-based segmentation, which groups adjacent pixels into meaningful objects and leverages spatial relationships for classification. Previous literature has shown that OBIA methods outperform PB approaches in various ecological and remote sensing applications (Mishra and Crews 2014 ; Wallace et al., 2019 ; Nababan et al., 2021 ). For example, Cui et al., ( 2023 ) found that OBIA significantly improved classification accuracy for benthic habitats due to its ability to incorporate spatial context and shape information. Similarly, Zheng et al., ( 2022 ) demonstrated that OBIA was more effective in classifying coastal wetland vegetation, highlighting the method’s advantage in utilising both spectral and spatial data for better delineation of habitat boundaries and diverse vegetation types. However, this trend was not observed in our study. The OBIA model's tendency to generalise segments may have reduced the predictive utility of the data, especially given the heterogeneous nature of the habitat area in our research compared to the more homogenous environments in the cited studies. While the OBIA approach generally results in higher accuracy for broader biotope categories and improved specificity, it may overlook finer details and lead to misclassification. This is especially relevant for rarer or less represented biotopes. For instance, Othman et al., (2013) demonstrated that while object-based image analysis (OBIA) methods are generally effective, they can overlook finer details compared to pixel-based classifications, particularly in complex benthic habitats. This limitation was evident in the OBIA model’s performance from this study when classifying “Thallose red algae on sand with ground surge”, “Ecklonia-Phyllospora communities”, and the “High energy infralittoral zone”. Markedly, the OBIA model failed to learn to classify “Durvillaea communities”, as the pathway of the OBIA approach significantly reduced the number of sample points, which adversely influenced the model's performance and consequently lowered its overall performance metrics. This limitation meant that the OBIA approach suffered in comparison to the data-rich PB approach, which was better equipped to capture finer details and provide more accurate classifications for a broader range of biotopes. Trade-offs between specificity and map complexity have considerable implications for ecological management (Lecours et al., 2015 ) and it is likely that a middle ground is needed. However, it is probable that the preferred level of complexity is situation and application specific based on the needs of managers and available data. As such, authors suggest that trade-offs between PB and OBIA approaches be assessed more commonly in remote sensing projects and the differences clearly communicated to managers to ensure that informing conservation efforts remains centre stage. Analysing both approaches' variable importance (Fig. S1 .) and their density plots (Fig. S2.) within the supplementary materials sheds light on the underlying mechanisms driving the two models’ workings. The density plot shows that the OBIA model distributes importance across a broader range of variables, reflecting its tendency to utilise spatial relationships and context in its classifications (Hossain and Chen 2019 ). Gao and Mas ( 2008 ) found that OBIA methods excel in leveraging multiple scales of information, as evidenced here by the OBIA model’s top ten most important variables, which include various scales of elevation and landscape features. This aligns with Wahidin et al. ( 2015 ), who demonstrated that OBIA techniques can effectively classify diverse coral reef habitats by incorporating contextual information and spatial relationships—an ability mirrored in the broader variable importance observed in the OBIA model. By comparison, the PB model displays a more concentrated distribution of variable importance, with key variables dominating the classification process. Sibarddin et al., (2018) showed that PB methods rely heavily on specific, detailed features to achieve high classification accuracy, which likely contributes to the PB approach’s higher precision and recall by capturing fine-scale variations in habitat characteristics. This is supported by Anggoro et al., ( 2018 ), who found that pixel-based methods performed better in classifying detailed benthic habitats due to their fine-scale feature extraction capabilities. Furthermore, Whiteside et al., ( 2011 ) compared OBIA and PB methods for savanna mapping and found that while OBIA provided a more generalised classification, PB methods achieved greater detail and accuracy in distinguishing between different vegetation types. The concentrated variable importance observed in the PB model, as also reported by Othman et al., ( 2023 ), highlights its effectiveness in leveraging specific features to enhance precision and recall in classification tasks (Baker et al., 2013 ). Figure S3, S4, & S5 in the supplementary materials illustrates the model certainty across all biotope hierarchy classifications. Notably, regions with the highest heterogeneity, where biotope types are likely to overlap, also show the greatest uncertainty in both the PB and OBIA models. This pattern highlights the difficulties in distinguishing closely related biotope types in complex or transitional environments. These overlapping biotopes are often those that were least represented or under-sampled in the ground truth data (see Table.3.). While the PB approach captures these areas with higher resolution, providing finer-scale detail, both models reveal that these high-heterogeneity regions are prone to increased ambiguity. This suggests that targeted data collection in these under-sampled, high-uncertainty areas could substantially enhance model performance and improve biotope classification accuracy for both approaches. The implications of these findings are significant for marine spatial planning and management. The higher precision and recall of the PB models suggest they are more adept at capturing the fine-grained details necessary for detailed biotope classification. This can be particularly valuable in conservation planning, where understanding the specific distribution of various biotopes can inform more targeted and effective management strategies (Rubbens et al. 2023 ). On the other hand, the OBIA approach’s ability to provide clearer, more homogeneous classification makes it an excellent tool for broader scale mapping efforts where the focus may be on general habitat distribution rather than detailed biotope distinctions. This duality of PB and OBIA methods underscores the importance of selecting the appropriate classification technique based on the specific objectives of a given marine spatial planning project. The integration of PB and OBIA methods, as suggested by Ierodiaconou et al., ( 2018 ) and mirrored here, has the potential to leverage the strengths of both approaches, resulting in improved biotope classification accuracy. This approach could provide detailed and precise classifications while maintaining clear and coherent habitat boundaries, offering a more accurate and interpretable representation of marine biotopes. Future improvements in classification techniques could significantly enhance the accuracy and utility of both PB and OBIA methods. Advances in machine learning, such as deep learning techniques, have the potential to refine PB methods by capturing more complex patterns and relationships in the data (Misiuk et al., 2024 ). For OBIA methods, developments in spatial statistics and the incorporation of higher-resolution data from emerging technologies like drones or autonomous vehicles could provide more detailed inputs and improve segmentation and classification performance. Exploring these technological advancements and their applications in marine habitat mapping can drive further innovation and effectiveness. Conclusion In this study, we compared the performance of hierarchical random forests built using pixel-based (PB) or object-based image analysis (OBIA) methods in the classification of benthic biotopes for the northern section of Apollo Marine Park and the adjacent Cape Otway coastline, Victoria, Australia. The findings reveal that there was little to separate the two approaches’ accuracy and performance metrics. However, when visualised, their large-scale predictions exhibited clear differences. These results underscore the importance of selecting the appropriate sampling and classification techniques based on specific project objectives. PB methods are well-suited for detailed biotope classification, which is valuable for targeted conservation efforts, whereas OBIA methods are effective for broader-scale habitat mapping. Future research should explore integrating PB and OBIA methods to combine their strengths, potentially enhanced by advancements in machine learning and emerging technologies such as deep learning and higher resolution bathymetric data. This study highlights the potential for future innovations to improve both the accuracy and utility of habitat classification techniques for marine spatial planning and management. Declarations Funding For this project, Deakin University secured the Marine Parks grant (4-HAZCNTR) to provide funding for the collection of the data used in the Apollo Marine Park and along the Cape Otway coastline. Competing Interests The authors have no relevant financial or non-financial interests to disclose. Author Contributions All authors contributed to the study conception and design. Material preparation and data collection were performed by Daniel Ierodiaconou and Henry Simmons. Analysis was performed by Henry Simmons. The first draft of the manuscript was written by Henry Simmons and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript. Data Availability Multibeam Echosounder (MBES) Data: Available through the AusSeabed initiative, a national repository for seabed mapping data in Australian waters (https://www.ausseabed.gov.au/). Baited Remote Underwater Video Systems (BRUVS) Data: Accessible via GlobalArchive , an online platform for managing, storing, and sharing underwater imagery and associated metadata collected from BRUVS surveys worldwide (https://globalarchive.org/). Towed Video Data: Hosted on CoastKit , a centralized database that provides access to towed video imagery collected along Australia’s coastline (https://mapshare.vic.gov.au/coastkit/). Autonomous Underwater Vehicle (AUV) Data: Available through SQUIDLE+ , a platform designed for the annotation and analysis of underwater imagery, including data collected by AUVs. SQUIDLE+ facilitates the classification of marine species and habitats using advanced tools and AI-driven workflows (https://squidle.org/). Acknowledgement The authors respectfully acknowledge the Eastern Maar Aboriginal Corporation, the Traditional Owners of the land, sea, and sky country where this work was conducted. We extend our gratitude to Parks Australia for their funding through the Our Marine Parks Grant Program, which supported multibeam sonar data collection, as well as the deployment of stereo-baited remote underwater video and autonomous underwater video (AUV) in the Apollo Marine Park.We also thank the Department of Energy, Environment, and Climate Action for funding the multibeam data and towed video collection in State waters through the Victorian Coastal Monitoring Program. Further appreciation is extended to the IMOS AUV Facility for supporting the AUV deployments, and to the IMOS Understanding of Marine Imagery (UMI) Sub-Facility for image scoring and archiving via the Squidle+ platform.Data were sourced from Australia’s Integrated Marine Observing System (IMOS) – IMOS is enabled by the National Collaborative Research Infrastructure Strategy (NCRIS). We are grateful to Dr Ben Misiuk for his expertise in backscatter harmonization processes, and to Dr Sasha Whitmarsh, Dr Jacquomo Monk, and Dr Ariell Friedman for their training and assistance in habitat classification. References Achanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P. and Süsstrunk, S., 2012. SLIC superpixels compared to state-of-the-art superpixel methods. IEEE Trans. Pattern Anal. Mach., 34 (11), pp.2274-2282. https://doi.org/10.1109/TPAMI.2012.120 Amini, Z.Z., Adabi, M.H., Burrett, C.F. and Quilty, P.G., 2004. Bryozoan distribution and growth form associations as a tool in environmental interpretation, Tasmania, Australia. Sediment. Geol. , 167 (1-2), pp.1-15. https://doi.org/10.1016/j.sedgeo.2004.01.010 Anggoro, A., Sumartono, E., Siregar, V.P., Agus, S.B., Purnama, D., Puspitosari, D.A., Listyorini, T. and Sulistyo, B., 2018, November. Comparing object-based and pixel-based classifications for benthic habitats mapping in Pari Islands. In J. Phys. Conf. Ser. (Vol. 1114, No. 1, p. 012049). https://doi.org/10.1088/1742-6596/1114/1/012049 Archer, K.J. and Kimes, R.V., 2008. Empirical characterization of random forest variable importance measures. Comput. Stat. Data Anal. , 52 (4), pp.2249-2260. https://doi.org/10.1016/j.csda.2007.08.015 Baker, B.A., Warner, T.A., Conley, J.F. and McNeil, B.E., 2013. Does spatial resolution matter? A multi-scale comparison of object-based and pixel-based methods for detecting change associated with gas well drilling operations. Int. J. Remote Sens. , 34 (5), pp.1633-1651. https://doi.org/10.1080/01431161.2012.724540 Barry, J.P., Widdicombe, S. and Hall-Spencer, J.M., 2011. Effects of ocean acidification on marine biodiversity and ecosystem function. Ocean acidification , pp.192-209. https://doi.org/10.1093/oso/9780199591091.003.0015 Beardmore, J.A., Mair, G.C. and Lewis, R.I., 1997. Biodiversity in aquatic systems in relation to aquaculture. Aquac. Res. , 28 (10), pp.829-839. https://doi.org/10.1046/j.1365-2109.1997.00947.x Belgiu, M. and Drăguţ, L., 2016. Random forest in remote sensing: A review of applications and future directions. ISPRS J. Photogramm. Remote Sens , 114 , pp.24-31. https://doi.org/10.1016/j.isprsjprs.2016.01.011 Benz, U.C., Hofmann, P., Willhauck, G., Lingenfelder, I. and Heynen, M., 2004. Multi-resolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready information. ISPRS J. Photogramm. Remote Sens , 58 (3-4), pp.239-258. https://doi.org/10.1016/j.isprsjprs.2003.10.002 Berhane, T.M., Lane, C.R., Wu, Q., Anenkhonov, O.A., Chepinoga, V.V., Autrey, B.C. and Liu, H., 2017. Comparing pixel-and object-based approaches in effectively classifying wetland-dominated landscapes. J. Remote Sens. , 10 (1), p.46. https://doi.org/10.3390/rs10010046 Blaschke, T., 2010. Object based image analysis for remote sensing. ISPRS J. Photogramm. Remote Sens , 65 (1), pp.2-16. https://doi.org/10.1016/j.isprsjprs.2009.06.004 Breiman, L., 2001. Random forests. Mach. Learn. , 45 , pp.5-32. http://dx.doi.org/10.1023/A:1010933404324 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 (3), pp.502-520. https://doi.org/10.1016/j.ecss.2011.02.007 Buhl-Mortensen, L., Buhl-Mortensen, P., Dolan, M.J.F. and Gonzalez-Mirelis, G., 2015. Habitat mapping as a tool for conservation and sustainable use of marine resources: Some perspectives from the MAREANO Programme, Norway. J. Sea Res. , 100 , pp.46-61. https://doi.org/10.1016/j.seares.2014.10.014 Buonocore, E., Grande, U., Franzese, P.P. and Russo, G.F., 2021. Trends and evolution in the concept of marine ecosystem services: an overview. Water , 13 (15), p.2060. https://doi.org/10.3390/w13152060 Cogan, C.B., Todd, B.J., Lawton, P. and Noji, T.T., 2009. The role of marine habitat mapping in ecosystem-based management. ICES Mar. Sci. , 66 (9), pp.2033-2042. https://doi.org/10.1093/icesjms/fsp214 Cui, Q., Pan, H., Zhang, K., Li, X. and Sun, H., 2023. Multiscale and multisubgraph-based segmentation method for ocean remote sensing images. IEEE Trans. Geosci. Remote Sens , 61 , pp.1-20. https://doi.org/10.1109/TGRS.2023.3247697 Cutler, D.R., Edwards Jr, T.C., Beard, K.H., Cutler, A., Hess, K.T., Gibson, J. and Lawler, J.J., 2007. Random forests for classification in ecology. Ecol. , 88 (11), pp.2783-2792. https://doi.org/10.1890/07-0539.1 Dalby, O., Kennedy, D.M., McCarroll, R.J., Young, M. and Ierodiaconou, D., 2024. Mapping surface sediment characteristics in enclosed shallow‐marine environments using spatially balanced designs and the random forest algorithm. Earth Surf. Process. Landf. https://doi.org/10.1002/esp.5864 Eddy, T.D., Lam, V.W., Reygondeau, G., Cisneros-Montemayor, A.M., Greer, K., Palomares, M.L.D., Bruno, J.F., Ota, Y. and Cheung, W.W., 2021. Global decline in capacity of coral reefs to provide ecosystem services. One Earth , 4 (9), pp.1278-1285. https://doi.org/10.1016/j. oneear.2021.08.016 Edmunds, M. and Flynn, A., 2018. CBiCS classification of Victorian biotopes. https://doi.org/10.13140/RG.2.2.16919.29603 Esetlili, M.T., Balcik, F.B., Sanli, F.B., Kalkan, K., Ustuner, M., Goksel, C., Gazioğlu, C. and Kurucu, Y., 2018. Comparison of object and pixel-based classifications for mapping crops using Rapideye imagery: a case study of Menemen Plain, Turkey. Int. j. environ. geoinformatics , 5(2), pp.231-243. https://doi.org/10.30897/ijegeo.442002 Field, C.B., Behrenfeld, M.J., Randerson, J.T. and Falkowski, P., 1998. Primary production of the biosphere: integrating terrestrial and oceanic components. J. Sci , 281 (5374), pp.237-240. https://doi.org/10.1126/science.281.5374.237 Fingas, M., 2019. Remote sensing for marine management. In World seas: An environmental evaluation (pp. 103-119). Academic Press. https://doi.org/10.1016/B978-0-12-805052-1.00005-X Fontoura, L., D’agata, S., Gamoyo, M., Barneche, D.R., Luiz, O.J., Madin, E.M., Eggertsen, L. and Maina, J.M., 2022. Protecting connectivity promotes successful biodiversity and fisheries conservation. J. Sci. , 375 (6578), pp.336-340. https://doi.org/10.1126/science.abg4351 Galelli, S., Humphrey, G.B., Maier, H.R., Castelletti, A., Dandy, G.C. and Gibbs, M.S., 2014. An evaluation framework for input variable selection algorithms for environmental data-driven models. Environ. Model. Softw. , 62 , pp.33-51. https://doi.org/10.1016/j.envsoft.2014.08.015 Gamfeldt, L., Lefcheck, J.S., Byrnes, J.E., Cardinale, B.J., Duffy, J.E. and Griffin, J.N., 2015. Marine biodiversity and ecosystem functioning: what's known and what's next?. Oikos , 124 (3), pp.252-265. https://doi.org/10.1111/oik.01549 Gao, Y. and Mas, J.F., 2008. A comparison of the performance of pixel-based and object-based classifications over images with various spatial resolutions. J. Earth Sci. , 2 (1), pp.27-35. https://doi.org/10.30897/ijegeo.442002 Gill, P.C. (2002). A blue whale (Balaenoptera musculus) feeding ground in a southern Australian coastal upwelling zone. J. Cetac. Res. Manage., 4(2): 179-184. https://doi.org/10.47536/jcrm.v4i2.854 Grebby, S., Field, E. and Tansey, K., 2016. Evaluating the use of an object-based approach to lithological mapping in vegetated terrain. Remote Sens., 8(10), p.843. https://doi.org/10.3390/rs8100843 Gregorutti, B., Michel, B. and Saint-Pierre, P., 2017. Correlation and variable importance in random forests. Stat. Comput., 27 , pp.659-678. https://doi.org/10.1007/s11222-016-9646-1 Gregr, E.J., Haggarty, D.R., Davies, S.C., Fields, C. and Lessard, J., 2021. Comprehensive marine substrate classification applied to Canada’s Pacific shelf. PloS one , 16 (10), p.e0259156. https://doi.org/10.1371/journal.pone.0259156 Hasan, R. C., Ierodiaconou, D., Laurenson, L. and Schimel, A., 2014. Integrating multibeam backscatter angular response, mosaic and bathymetry data for benthic habitat mapping. Plos one , 9 (5), p.e97339. https://doi.org/10.1371/journal.pone.0097339 Hasan, R.C., Ierodiaconou, D. and Laurenson, L., 2012. Combining angular response classification and backscatter imagery segmentation for benthic biological habitat mapping. Estuar. Coast. Shelf Sci. , 97 , pp.1-9. https://doi.org/10.1016/j.ecss.2011.10.004 Hay, G.J., Marceau, D.J., Dube, P. and Bouchard, A., 2001. A multiscale framework for landscape analysis: object-specific analysis and upscaling. Landsc. Ecol. , 16 , pp.471-490. https://doi.org/10.1023/A:1013101931793 Hossain, M.D. and Chen, D., 2019. Segmentation for Object-Based Image Analysis (OBIA): A review of algorithms and challenges from remote sensing perspective. ISPRS J. Photogramm. Remote Sens , 150 , pp.115-134. https://doi.org/10.1016/j.isprsjprs.2019.02.009 Ierodiaconou, D., Schimel, A.C., Kennedy, D., Monk, J., Gaylard, G., Young, M., Diesing, M. and Rattray, A., 2018. Combining pixel and object based image analysis of ultra-high resolution multibeam bathymetry and backscatter for habitat mapping in shallow marine waters. Mar. Geophys. Res. , 39 , pp.271-288. https://doi.org/10.1007/s11001-017-9338-z Ilich, A.R., Misiuk, B., Lecours, V. and Murawski, S.A., 2023. MultiscaleDTM: An open‐source R package for multiscale geomorphometric analysis. Trans. GIS. , 27 (4), pp.1164-1204. https://doi.org/10.1111/tgis.13067 Intergovernmental Oceanographic Commission, 2018. United Nations Decade of Ocean Science for Sustainable Development (2021–2030): call for nomination to the Executive Planning Group (EPG). Lecours, V., Devillers, R., Schneider, D.C., Lucieer, V.L., Brown, C.J. and Edinger, E.N., 2015. Spatial scale and geographic context in benthic habitat mapping: review and future directions. Mar. Ecol. Prog. Ser. , 535 , pp.259-284. https://doi.org/10.3354/meps11378 Lu, D. and Weng, Q., 2007. A survey of image classification methods and techniques for improving classification performance. Int. J. Remote Sens. , 28 (5), pp.823-870. https://doi.org/10.1080/01431160600746456 Martin, C.L., Momtaz, S., Gaston, T. and Moltschaniwskyj, N.A., 2016. A systematic quantitative review of coastal and marine cultural ecosystem services: current status and future research. Mar. Policy , 74 , pp.25-32. https://doi.org/10.1016/j.marpol.2016.09.004 Martínez-Santos, P., Aristizábal, H.F., Díaz-Alcaide, S. and Gómez-Escalonilla, V., 2021. Predictive mapping of aquatic ecosystems by means of support vector machines and random forests. J. Hydrol. , 595 , p.126026. https://doi.org/10.1016/j.jhydrol.2021.126026 Mayer, L., Jakobsson, M., Allen, G., Dorschel, B., Falconer, R., Ferrini, V., Lamarche, G., Snaith, H. and Weatherall, P., 2018. The Nippon Foundation—GEBCO seabed 2030 project: The quest to see the world’s oceans completely mapped by 2030. Geosciences , 8 (2), p.63. https://doi.org/10.3390/geosciences8020063 McKenzie, L.J., Nordlund, L.M., Jones, B.L., Cullen-Unsworth, L.C., Roelfsema, C. and Unsworth, R.K., 2020. The global distribution of seagrass meadows. Environ. Res. Lett. , 15 (7), p.074041. https://doi.org/10.1088/1748-9326/ab7d06 Melesse, A.M., Weng, Q., Thenkabail, P.S. and Senay, G.B., 2007. Remote sensing sensors and applications in environmental resources mapping and modelling. Sens. J. , 7 (12), pp.3209-3241. https://doi.org/10.3390/s7123209 Mishra, N.B. and Crews, K.A., 2014. Mapping vegetation morphology types in a dry savanna ecosystem: Integrating hierarchical object-based image analysis with Random Forest. Int. J. Remote Sens., 35 (3), pp.1175-1198. https://doi.org/10.3390/land4010197 Mitchell, P.J., Downie, A.-L. & Diesing, M. (2018) How good is my map? A tool for semi-automated thematic mapping and spatially explicit confidence assessment. Environ. Model. Softw., 108, 111–122. https://doi.org/10.1016/j.envsoft.2018.07.014 Misiuk, B., Brown, C.J., Robert, K. and Lacharité, M., 2020. Harmonizing multi-source sonar backscatter datasets for seabed mapping using bulk shift approaches. Remote Sens. , 12 (4), p.601. https://doi.org/10.3390/rs12040601 Misiuk, B., Diesing, M., Aitken, A., Brown, C.J., Edinger, E.N. and Bell, T., 2019. A spatially explicit comparison of quantitative and categorical modelling approaches for mapping seabed sediments using random forest. Geosci. , 9 (6), p.254. https://doi.org/10.3390/geosciences9060254 Misiuk, B., Tan, Y.L., Li, M.Z., Trappenberg, T., Alleosfour, A., Church, I.W., Ferrini, V. and Brown, C.J., 2024. Multivariate mapping of seabed grain size parameters in the Bay of Fundy using convolutional neural networks. Mar. Geol. , 472 , p.107299. https://doi.org/10.1016/j.margeo.2024.107299 Nababan, B., Mastu, L.O.K., Idris, N.H. and Panjaitan, J.P., 2021. Shallow-water benthic habitat mapping using drone with object based image analyses. Remote sens. , 13 (21), p.4452. https://doi.org/10.3390/rs13214452 National Marine Science Committee, 2015. National marine science plan 2015-2025: driving the development of Australia's blue economy . National Marine Science Committee. National Research Council, Board on Environmental Change and Committee on the Human Dimensions of Global Change, 1998. People and pixels: Linking remote sensing and social science . National Academies Press. Nikolaou, A. and Katsanevakis, S., 2023. Marine extinctions and their drivers. Reg. Environ. Change , 23 (3), p.88. https://doi.org/10.1007/s10113-023-02081-8 Nowosad J, Stepinski T (2022). “Extended SLIC superpixels algorithm for applications to non-imagery geospatial rasters.” Int. J. Appl. Earth Obs. Geoinf. , pp.112. https://doi.org/10.1016/j.jag.2022.102935 Nursamsi, I., Phinn, S.R., Levin, N., Luskin, M.S. and Sonter, L.J., 2024. Remote sensing of artisanal and small-scale mining: A review of scalable mapping approaches. Sci. Total Environ ., p.175761. https://doi.org/10.1016/j.scitotenv.2024.175761 Othman, A.N., Hashim, N., Salim, P.M. and Zaidi, P.N.S.M., 2023. Comparative Study of Pixel-Based and Object-Based Classifications in Benthic Mapping. J. Adv. Geosp. Sci. Technol. 3 (2), pp.51-62. https://doi.org/10.11113/jagst.v3n2.69 Perkins, N.R., Foster, S.D., Hill, N.A. and Barrett, N.S., 2016. Image subsampling and point scoring approaches for large-scale marine benthic monitoring programs. Estuar. Coast. Shelf Sci. , 176 , pp.36-46. https://doi.org/10.1016/j.ecss.2023.108438 Porskamp, P., Rattray, A., Young, M. and Ierodiaconou, D., 2018. Multiscale and hierarchical classification for benthic habitat mapping. Geosci. , 8 (4), p.119. https://doi.org/10.3390/geosciences8040119 Proctor, R., Langlois, T., Friedman, A., Mancini, S., Hoenner, X. and Davey, B., 2018. Cloud-based national on-line services to annotate and analyse underwater imagery. Ramírez, F., Afán, I., Davis, L.S. and Chiaradia, A., 2017. Climate impacts on global hot spots of marine biodiversity. Sci. Adv. , 3 (2), p.e1601198. https://doi.org/10.1126/sciadv.1601198 Rattray, A., Ierodiaconou, D., Monk, J., Laurenson, L.J.B. and Kennedy, P., 2014. Quantification of spatial and thematic uncertainty in the application of underwater video for benthic habitat mapping. Mar. Geod. , 37(3), pp.315-336. Reid, P.C., Fischer, A.C., Lewis-Brown, E., Meredith, M.P., Sparrow, M., Andersson, A.J., Antia, A., Bates, N.R., Bathmann, U., Beaugrand, G. and Brix, H., 2009. Impacts of the oceans on climate change. Adv. Mar. Biol , 56 , pp.1-150. https://doi.org/10.1016/S0065-2881(09)56001-4 Rigatti, S.J., 2017. Random forest. J Insur. Med , 47 (1), pp.31-39. https://doi.org/10.17849/insm-47-01-31-39.1 Rubbens, P., Brodie, S., Cordier, T., Destro Barcellos, D., Devos, P., Fernandes-Salvador, J.A., Fincham, J.I., Gomes, A., Handegard, N.O., Howell, K. and Jamet, C., 2023. Machine learning in marine ecology: an overview of techniques and applications. ICES J. Mar. Sci., 80(7), pp.1829-1853. https://doi.org/10.1093/icesjms/fsad100 Rubidge, E.M., Gale, K.S. and Curtis, J.M., 2016. Community ecological modelling as an alternative to physiographic classifications for marine conservation planning. Biodivers. Conserv. , 25 , pp.1899-1920. https://doi.org/10.1016/j.ecolind.2021.107849 Sanders, K.T. and Masri, S.F., 2016. The energy-water agriculture nexus: the past, present and future of holistic resource management via remote sensing technologies. J. Clean. Prod. , 117, pp.73-88. https://doi.org/10.1016/j.heliyon.2023.e17016 Secretary-General, U.N., 2017. Progress towards the Sustainable Development Goals: report of the Secretary-General. Shapiro, S. S., and Wilk, M. B. (1965). An Analysis of Variance Test for Normality (Complete Samples). Biometrika , 52(3/4), 591–611. https://doi.org/10.2307/2333709 Shields, J., Pizarro, O. and Williams, S.B., 2021. Feature space exploration for planning initial benthic AUV surveys. arXiv preprint at https://doi.org/10.48550/arXiv.2105.11598 Sibaruddin, H.I., Shafri, H.Z.M., Pradhan, B. and Haron, N.A., 2018, June. Comparison of pixel-based and object-based image classification techniques in extracting information from UAV imagery data. IOP Conf. Ser.: Earth Environ. Sci. (Vol. 169, No. 1, p. 012098). IOP Publishing. https://doi.org/10.1088/1755-1315/169/1/012098 Spalding, M.D., Ruffo, S., Lacambra, C., Meliane, I., Hale, L.Z., Shepard, C.C. and Beck, M.W., 2014. The role of ecosystems in coastal protection: Adapting to climate change and coastal hazards. Ocean Coast. Manag., 90 , pp.50-57. https://doi.org/10.1016/j.ocecoaman.2013.09.007 Strong, J.A. and Elliott, M., 2017. The value of remote sensing techniques in supporting effective extrapolation across multiple marine spatial scales. Mar. Pollut. Bull. , 116 (1-2), pp.405-419. https://doi.org/10.1016/j.marpolbul.2017.01.028 Turner, W., Spector, S., Gardiner, N., Fladeland, M., Sterling, E. and Steininger, M., 2003. Remote sensing for biodiversity science and conservation. rends Ecol. Evol., 18 (6), pp.306-314. https://doi.org/10.1016/S0169-5347(03)00070-3 Victoria, P., 2003. Victoria’s System of Marine National Parks and Marine Sanctuaries. Management Strategy 2003–2010 . Von Schuckmann, K., Holland, E., Haugan, P. and Thomson, P., 2020. Ocean science, data, and services for the UN 2030 Sustainable Development Goals. Mar. Policy , 121 , p.104154. https://doi.org/10.1016/j.marpol.2020.104154 Wager, S., Hastie, T. & Efron, B. (2014) Confidence intervals for random forests: the jackknife and the infinitesimal jackknife. J. Mach. Learn. Res., 15(1), 1625–1651. Wahidin, N., Siregar, V.P., Nababan, B., Jaya, I. and Wouthuyzen, S., 2015. Object-based image analysis for coral reef benthic habitat mapping with several classification algorithms. Procedia Environ. Sci. , 24 , pp.222-227. https://doi.org/10.1016/j.proenv.2015.03.029 Wallace, L., Saldias, D.S., Reinke, K., Hillman, S., Hally, B. and Jones, S., 2019. Using orthoimages generated from oblique terrestrial photography to estimate and monitor vegetation cover. Ecol. Indic. , 101 , pp.91-101. https://doi.org/10.1080/15481603.2020.1859264 Ward, D., Melbourne-Thomas, J., Pecl, G.T., Evans, K., Green, M., McCormack, P.C., Novaglio, C., Trebilco, R., Bax, N., Brasier, M.J. and Cavan, E.L., 2022. Safeguarding marine life: conservation of biodiversity and ecosystems. Rev. Fish Biol. Fish.. , 32 (1), pp.65-100. https://doi.org/10.1007/s11160-022-09700-3 Wass, R.E., Conolly, J.R. and MacIntyre, R.J., 1970. Bryozoan carbonate sand continuous along southern Australia. Mar. Geol. , 9 (1), pp.63-73. https://doi.org/10.1016/0025-3227(70)90080-0 Wernberg, T., Krumhansl, K., Filbee-Dexter, K. and Pedersen, M.F., 2019. Status and trends for the world’s kelp forests. In World seas: An environmental evaluation (pp. 57-78). Academic Press. https://doi.org/10.1016/B978-0-12-805052-1.00003-6 Wescott, G., 2006. The long and winding road: the development of a comprehensive, adequate and representative system of highly protected marine protected areas in Victoria, Australia. Ocean Coast. Manag., 49 (12), pp.905-922. https://doi.org/10.1016/j.ocecoaman.2006.08.001 Whiteside, T.G., Boggs, G.S. and Maier, S.W., 2011. Comparing object-based and pixel-based classifications for mapping savannas. J. Appl. Earth Obs. Geoinformation. , 13 (6), pp.884-893. https://doi.org/10.1016/j.jag.2011.06.008 Zheng, J.Y., Hao, Y.Y., Wang, Y.C., Zhou, S.Q., Wu, W.B., Yuan, Q., Gao, Y., Guo, H.Q., Cai, X.X. and Zhao, B., 2022. Coastal wetland vegetation classification using pixel-based, object-based and deep learning methods based on RGB-UAV. Land , 11 (11), p.2039. https://doi.org/10.3390/land11112039 Zone, S. and Zone, H.P., 2013. South-east Commonwealth Marine Reserves Network. Additional Declarations No competing interests reported. Supplementary Files SimmonsetalPBVSOBIASupplementary20241030.docx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-5351238","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":374621225,"identity":"fe382a9f-b335-4a67-8f32-7dfa0c89d93b","order_by":0,"name":"Henry O. J. Simmons","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABFklEQVRIiWNgGAWjYBACAyjNA6EqmIFMxgYgi5lYLWdI0AIBjG3MMN24tZizn332mYehTkZ32uFjD3/Os5bj7znc9oChwjqxgf2MATYtlj3pxrN5GA7zmN1OSzfm3ZZuLHG2sd2A4Ux6YgNPDlYtBgfSmJl5GA4AteSYSTNuO5zYcJ6xTYKxDchgwKHl/DOQljqwFsmfcw7Xzwdr+QfUwv8Gu5YbYFuYwVokeBsOJxicbQRqaQBqkcBhy41nzIxzDMB+SZPmOZZuuPHMwTaJhGPpxm0SzwqwOyyNmeFNRZ292e3kY5I/aqzl5c6kP5P4UGMt28+fvAF7MDMwMPFgOCABiNlwqQcCxh94JEfBKBgFo2AUMAAAF5la5AQmfgIAAAAASUVORK5CYII=","orcid":"","institution":"Deakin University","correspondingAuthor":true,"prefix":"","firstName":"Henry","middleName":"O. J.","lastName":"Simmons","suffix":""},{"id":374621226,"identity":"d98ac20e-105d-453d-bf8b-760f80d2ddd0","order_by":1,"name":"Oli Dalby","email":"","orcid":"","institution":"Deakin University","correspondingAuthor":false,"prefix":"","firstName":"Oli","middleName":"","lastName":"Dalby","suffix":""},{"id":374621227,"identity":"764071dd-ed97-40d3-bf19-38a07bd76e16","order_by":2,"name":"Daniel Ierodiaconou","email":"","orcid":"","institution":"Deakin University","correspondingAuthor":false,"prefix":"","firstName":"Daniel","middleName":"","lastName":"Ierodiaconou","suffix":""},{"id":374621228,"identity":"1d3a710c-21ff-44ac-ab04-bd7fcd3529f9","order_by":3,"name":"Mary A. Young","email":"","orcid":"","institution":"Deakin University","correspondingAuthor":false,"prefix":"","firstName":"Mary","middleName":"A.","lastName":"Young","suffix":""}],"badges":[],"createdAt":"2024-10-29 05:23:30","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5351238/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5351238/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":69437124,"identity":"eed4a772-6de5-48c3-8883-393587f40346","added_by":"auto","created_at":"2024-11-20 10:44:17","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":225385,"visible":true,"origin":"","legend":"\u003cp\u003eLocation of Apollo Marine Park within the waters of Bass strait, Victoria, Australia and the site MBES bathymetry and backscatter.\u003c/p\u003e","description":"","filename":"image1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-5351238/v1/33931b568542a9b2398ac82f.jpeg"},{"id":69436137,"identity":"03f9f33f-eb09-4c2a-b5f7-9a890e953562","added_by":"auto","created_at":"2024-11-20 10:36:18","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":535643,"visible":true,"origin":"","legend":"\u003cp\u003eSupercells segmentation of MBES backscatter. Panels A, B, and C show a closer look at the segmentation with the superpixels outlined in white.\u003c/p\u003e","description":"","filename":"image3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-5351238/v1/fb4d751a015734e559bff83a.jpeg"},{"id":69436134,"identity":"72da269f-9a18-4aeb-9a06-d2678b2d356c","added_by":"auto","created_at":"2024-11-20 10:36:17","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":205900,"visible":true,"origin":"","legend":"\u003cp\u003eGround-truth data for this study. A) represents the towed video. B) represents the AUV transect imagery. C) represents the BRUV drop images.\u003c/p\u003e","description":"","filename":"image5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-5351238/v1/975cad1901e500d93280825b.jpeg"},{"id":69436133,"identity":"e63f8938-d557-4650-b3de-e05aed0b4a26","added_by":"auto","created_at":"2024-11-20 10:36:17","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":19677,"visible":true,"origin":"","legend":"\u003cp\u003eBiotope coverage comparison between Object-Based Image Analysis (OBIA) and Pixel-Based (PB) hierarchical random forest models at the BC2 hierarchical level of the Combined Biotope Classification Scheme.\u003c/p\u003e","description":"","filename":"image7.png","url":"https://assets-eu.researchsquare.com/files/rs-5351238/v1/bbd54f62469992fdec1c6e2d.png"},{"id":69436136,"identity":"b2dce608-a34a-4bb0-b3fd-00f6544c090c","added_by":"auto","created_at":"2024-11-20 10:36:17","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":1316310,"visible":true,"origin":"","legend":"\u003cp\u003eThe classification result of pixel-base method (A) and object-based method (B) for the BC2 hierarchical level of the Combined Biotope Classification Scheme.\u003c/p\u003e","description":"","filename":"image9.png","url":"https://assets-eu.researchsquare.com/files/rs-5351238/v1/542633fd3d8edbd80b02ff4b.png"},{"id":69437125,"identity":"64ca1bc6-ac0d-4cd6-8075-c8024f7afa4e","added_by":"auto","created_at":"2024-11-20 10:44:18","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":18648,"visible":true,"origin":"","legend":"\u003cp\u003eBiotope coverage comparison between Object-Based Image Analysis (OBIA) and Pixel-Based (PB) hierarchical random forest models at the BC3 hierarchical level of the Combined Biotope Classification Scheme.\u003c/p\u003e","description":"","filename":"image11.png","url":"https://assets-eu.researchsquare.com/files/rs-5351238/v1/7445250d98b5934f5059af3b.png"},{"id":69437883,"identity":"979cecbd-3e5f-4b1e-a3ac-832841aa388b","added_by":"auto","created_at":"2024-11-20 10:52:18","extension":"jpeg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":327505,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eClassification result of pixel-base method (A) and object-based method (B) for the BC3 hierarchical level of the Combined Biotope Classification Scheme.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"image13.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-5351238/v1/055047a78caf517aafc2d3ac.jpeg"},{"id":69436141,"identity":"5a27c0d2-beab-4573-ae9b-7866a688102c","added_by":"auto","created_at":"2024-11-20 10:36:18","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":32755,"visible":true,"origin":"","legend":"\u003cp\u003eBiotope coverage comparison between Object-Based Image Analysis (OBIA) and Pixel-Based (PB) hierarchical random forest models at the BC4 hierarchical level of the Combined Biotope Classification Scheme.\u003c/p\u003e","description":"","filename":"image15.png","url":"https://assets-eu.researchsquare.com/files/rs-5351238/v1/1d5894a54c49360116fac145.png"},{"id":69436142,"identity":"1242a04e-c3af-40e1-ac43-f591f8edff82","added_by":"auto","created_at":"2024-11-20 10:36:18","extension":"jpeg","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":423158,"visible":true,"origin":"","legend":"\u003cp\u003eClassification result of pixel-base method (A) and object-based method (B) for the BC4 hierarchical level of the Combined Biotope Classification Scheme.\u003c/p\u003e","description":"","filename":"image17.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-5351238/v1/edfefe178b45c3f0d4546f52.jpeg"},{"id":69436138,"identity":"b36f6c89-17d3-43a4-82ef-cdf9233e0a6f","added_by":"auto","created_at":"2024-11-20 10:36:18","extension":"jpeg","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":740742,"visible":true,"origin":"","legend":"\u003cp\u003eA closer look at some areas of interest for the classification result of pixel-base method and object-based method for the BC4 hierarchical level of the Combined Biotope Classification Scheme.\u003c/p\u003e","description":"","filename":"image19.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-5351238/v1/62fea574860693ea79e1d761.jpeg"},{"id":80370490,"identity":"5a7ce08f-7fa7-4bb5-9770-4565d6335169","added_by":"auto","created_at":"2025-04-11 06:38:50","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4571878,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5351238/v1/7a55fc21-7679-48d8-a4bc-ac36f1e15626.pdf"},{"id":69436143,"identity":"4ec18fc6-ce0a-4cbc-be8c-462fea05d0de","added_by":"auto","created_at":"2024-11-20 10:36:18","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":8377365,"visible":true,"origin":"","legend":"","description":"","filename":"SimmonsetalPBVSOBIASupplementary20241030.docx","url":"https://assets-eu.researchsquare.com/files/rs-5351238/v1/fa56cb3c0b1a21d4ff8a8b93.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Comparing Pixel-and Object-Based Approaches for Classifying Benthic Habitats","fulltext":[{"header":"Introduction","content":"\u003cp\u003eThe global ocean accounts for 71% of the Earth\u0026rsquo;s surface, plays a central role in the structure and functioning of our biosphere and is critical for achieving the sustainable development of human society (Field et al., \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e1998\u003c/span\u003e; Mayer et al., \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; von Schuckmann et al., \u003cspan citationid=\"CR77\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Marine ecosystems provide several ecosystem services that include but are not limited to climate regulation (Reid et al., \u003cspan citationid=\"CR64\" class=\"CitationRef\"\u003e2009\u003c/span\u003e), coastal protection (Spalding et al., \u003cspan citationid=\"CR73\" class=\"CitationRef\"\u003e2014\u003c/span\u003e), biodiversity (Gamfeldt et al., \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2015\u003c/span\u003e), fisheries and aquaculture (Beardmore et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e1997\u003c/span\u003e; Fontoura et al., \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) and help ensure food security and economic benefits for millions (The and Sumaila 2013). Additionally, oceans offer recreational and cultural value, enriching human lives and sustaining coastal communities (Martin et al., \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). Without healthy marine ecosystems, these services become compromised. Vital marine ecosystems like coral reefs, kelp forests, and seagrass meadows are experiencing losses worldwide due to increased sea surface temperatures (Ram\u0026iacute;rez et al., \u003cspan citationid=\"CR62\" class=\"CitationRef\"\u003e2017\u003c/span\u003e), ocean acidification (Barry et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2011\u003c/span\u003e), habitat destruction (Nikolaou and Katsanevakis \u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e2023\u003c/span\u003e), and pollution (Buonocore et al., \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) (Wernberg et al., \u003cspan citationid=\"CR83\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; McKenzie 2020; Eddy et al., \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). The majority of these pressures can be linked to anthropogenic sources.\u003c/p\u003e \u003cp\u003eDespite these challenges, there is hope. On an international scale, Goal 14 of the United Nations Sustainable Development Goals focuses on conserving and sustainably using the marine resources (Secretary-General \u003cspan citationid=\"CR69\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Monoly et al., 2022). Whilst nationally, Australia\u0026rsquo;s Marine Science Plan has been established (National Marine Science Committee, \u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e2015\u003c/span\u003e) with the aim to support sustainable management and conservation of marine resources by improving access and allocation of fisheries, enhancing productivity, maintaining habitat health, and addressing climate change impacts. The plan also calls for integrated efforts from government, industry, and the scientific community to support sustainable development and conservation of marine biodiversity​.\u003c/p\u003e \u003cp\u003eRecognition of the need to protect marine ecosystems is not a recent development; it has been a priority for decades (Ward et al., \u003cspan citationid=\"CR81\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Within Australia, the state of Victoria has a strong history of marine conservation. In 2002, Victoria implemented one of the world\u0026rsquo;s first representative systems of fully protected marine areas (Victoria \u003cspan citationid=\"CR76\" class=\"CitationRef\"\u003e2003\u003c/span\u003e), increasing the states \u0026ldquo;no-take\u0026rdquo; Marine Protected Areas (MPAs) 100-fold to cover 5% of its coastal waters (Wescott \u003cspan citationid=\"CR84\" class=\"CitationRef\"\u003e2006\u003c/span\u003e). In 2013, Australian national conservation efforts were expanded with the creation of the Southeast Commonwealth Marine Reserves Network, comprising 14 marine parks that allowed for regulated and sustainable commercial and recreational use (Zone and Zone \u003cspan citationid=\"CR87\" class=\"CitationRef\"\u003e2013\u003c/span\u003e). These reserves are a part of Australia's commitment to the conservation and sustainable management of its marine environment, aiming to balance conservation goals with responsible resource use. Each reserve includes specific zoning and management plans tailored to its unique ecological and socio-economic context.\u003c/p\u003e \u003cp\u003eHowever, effective marine park management requires detailed assessments of the distribution and abundance of communities within and beyond their boundaries. This can be complicated by several issues. For example, collecting in-situ data can be arduous, expensive, and time-consuming, especially in waters exposed to dangerous sea conditions. Consequently, remote sensing and modelling platforms have become essential tools for researchers tasked with mapping and monitoring marine environments and resources (Melesse et al., \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e2007\u003c/span\u003e; Sanders and Masri, \u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Strong and Elliot, 2017; Fingas, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). By integrating remote sensing technologies with geospatial analyses and recent improvements in computing power, researchers can generate high-resolution maps of marine habitats. These maps can depict the presence, condition, and spatial arrangement of habitats present and can be repeated over time to identify trends in community composition and habitat condition. This, in turn, provides critical baseline data to inform effective marine reserve management and conservation efforts (Cogan et al., \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2009\u003c/span\u003e; Brown et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2011\u003c/span\u003e; Buhl-Mortensen et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2015\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eIn recent years, evolution of habitat mapping techniques has shifted from traditional pixel-based (PB) approaches to more advanced Object-Based Image Analysis (OBIA) methods (Benz et al., \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2004\u003c/span\u003e; Lu and Weng \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2007\u003c/span\u003e; Belgiu and Drăguţ \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). PB approaches rely on individual pixel values to classify and map habitats based on spectral signatures. However, PB methods often struggle with mixed pixels and spectral confusion in complex environments, leading to a \u0026lsquo;salt and pepper\u0026rsquo; effect when classifying (Turner et al., \u003cspan citationid=\"CR75\" class=\"CitationRef\"\u003e2003\u003c/span\u003e). The development of OBIA in the late 20th century represented a paradigm shift by considering spatial relationships and context between pixels (Blaschke \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2010\u003c/span\u003e). Specifically, OBIA employs segmentation algorithms to group pixels into meaningful objects or regions based on spectral, spatial, and contextual information (Hay et al., \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2001\u003c/span\u003e), allowing for more precise delineation of habitat boundaries and characteristics. OBIA has since become prominent in ecological studies and habitat mapping, enhancing the accuracy and applicability of remote sensing data in environmental management and conservation efforts (Grebby et al. \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Esetlili et al. \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2018\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eDespite advancements in the production and application of OBIA, several knowledge gaps remain. There has been little comparison between the application of OBIA vs PB approaches in benthic habitat classifications. In one paper, Ierodiaconou et al. (\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2018\u003c/span\u003e) compared the two methods in a relatively shallow, largely homogeneous, and spatially small cove (0.39km\u003csup\u003e2\u003c/sup\u003e) in Victoria, Australia, where only five different habitat types were identified. Apart from this paper, comprehensive comparisons between OBIA and PB approaches have been sparse and typically limited to satellite data focused on shallow coastal waters, wetlands, or terrestrial areas (Whiteside et al., \u003cspan citationid=\"CR85\" class=\"CitationRef\"\u003e2011\u003c/span\u003e; Wahidin et al., \u003cspan citationid=\"CR79\" class=\"CitationRef\"\u003e2015\u003c/span\u003e; Berhane et al., \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Anggoro et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). As such, more examples comparing OBIA and PB approaches that classify acoustic data (e.g. MBES or SSS) are required to assess trade-offs between approaches. Furthermore, the application of hierarchical machine learning approaches remains unexplored.\u003c/p\u003e \u003cp\u003eCrucially, the decision between using PB or OBIA should be informed by the specific management goals of the study. The interpretability of each method varies depending on the level of detail required for different management or conservation objectives. Previous studies have noted differences in accuracy, scalability, and practical applications between the two approaches (Whiteside et al. \u003cspan citationid=\"CR85\" class=\"CitationRef\"\u003e2011\u003c/span\u003e; Nursamsi et al. \u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e2024\u003c/span\u003e), but little research has specifically examined how these goals influence the choice of methodology, particularly in benthic habitat classification. Exploring these aspects could provide clearer guidance on which technique might best serve conservation or resource management needs, helping to ensure that data collection is fit for purpose in informing decision-making.\u003c/p\u003e \u003cp\u003eIn this study, we aim to address these knowledge gaps by conducting a comparative analysis of OBIA and PB approaches using high-resolution bathymetric data within Apollo Bay Marine Park, VIC, Australia. Specifically, we seek to compare the accuracy, interpretability, and applicability of PB and OBIA approaches in benthic habitat mapping, using this locale as a test case. The research not only aims to contribute to the field of marine habitat mapping but also aims to provide valuable insights for effective marine spatial management and conservation strategies in marine parks.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e\u003cem\u003eStudy Area\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe Apollo Marine Park (AMP) was established in 2007 and incorporated into the Southeast Commonwealth Marine Reserves Network in 2013 (Figure 1). Located three nautical miles south of Cape Otway in Victoria, Australia, the park covers 1184 km\u0026sup2; of the Bass Strait continental shelf, with depths ranging from less than 50 m near the Cape to deeper waters in the Otway Depression, an ancient river valley from the Last Glacial Maximum (~21 ka) (Wass et al., 1970; Amini et al., 2004). The AMP is exposed to strong south-westerly swells and tidal flows, resulting in high geoform complexity at the southern tip of the Cape. Biological observations indicate that mesophotic reefs and non-reef forming benthic assemblages dominated by sponges, as well as complex rock habitats with sea plumes and hydroid fans, extend from the north into the park (VEAC, 2019). The AMP supports diverse marine life, including seabirds, dolphins, seals, and white sharks, and serves as a critical migration route for blue, fin, sei, and humpback whales (Gill, 2002). Despite its ecological significance, there is limited knowledge of the species and habitats beyond the water surface.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eAcoustic Data Acquisition and Processing\u0026nbsp;\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eGeophysical data were collected between 2006 and 2021 by the Marine Mapping Group from Deakin University, Victoria, Australia, across several multibeam surveys, using varied equipment. Historic multibeam data were acquired in November and December of 2006 and 2007 respectively as part of the ongoing Victorian Marine Habitat Mapping Project. A Reson Seabat 8101 multibeam echosounder (MBES) (operating 240 kHz, 150\u0026deg; angular sector coverage) was used to collect data before real-time quality control, display, navigation, and post-processing were carried out using the Starfix Suite 7.1 (Fugro proprietary software).\u003c/p\u003e\n\u003cp\u003eContemporary multibeam surveys were also conducted within the northern sector of Apollo Marine Park between 2020 and 2021 using a Kongsberg Maritime EM2040C MBES (operating frequency of 300 kHz, 60\u0026deg; angular sector coverage). Positioning and real time positioning was completed using an Applanix POS MV Wavemaster, and a dual-frequency Trimble Global Navigation Satellite System (GNSS) receiver. For all sets of surveys, a sound velocity profiler was used regularly to correct the sonar data for local variations in sound speed. Twenty-nine linear km of data were acquired on and outside the parks northern boundary to map subtidal components of the Otway ridge, fill data gaps, and combine it with previous data collection initiatives.\u003c/p\u003e\n\u003cp\u003eMBES data from both surveys were used to produce a bathymetric and backscatter grid at 3 m resolution. Given the considerable time difference between surveys, variable multibeam technologies used, and differing technical specifications a normalisation procedure was needed to harmonize the two datasets. A mosaic was created for the bathymetry dataset, as measured depth did not vary significantly between the two systems. Backscatter data was combined using the \u0026lsquo;bulk shift\u0026rsquo; method Misiuk et al., (2020) which performs relative calibration of sonar backscatter datasets by modelling the error between overlapping surveys and adjusting the values to minimise discrepancies. This ensures consistent and accurate backscatter data across different surveys, facilitating seamless seabed mapping. The two datasets were then clipped to the study area (Fig. 1).\u003cstrong\u003e\u003cem\u003e\u0026nbsp;\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eMultibeam Derived Variables (Derivatives) \u0026nbsp;\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThirty-two derivatives from the bathymetry data were calculated to provide information on seafloor terrain attributes (Table 1) using the R package MultiscaleDTM (version 0.8.3, Ilich et al., 2023). MultiscaleDTM was chosen because it allows for the calculation of several terrain attributes within each of the five thematic groups (slope, aspect, curvature, relative position, and roughness) using a variable window size. Various neighbourhood scales were used for each terrain attribute using the 3m resolution harmonised bathymetry data (3, 5, 9, 21, 51, and 101 cell neighbourhoods). A full outline of measures of seafloor structure calculated are included in Table 1.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable.1\u003c/strong\u003e. These derivatives are calculated using different functions within the MutiscaleDTM (Ilich et al. 2023) package including multiscale slope and aspect (SlpAsp), quadratic surface to a rectangular focal window (Qfit), relative position (RelPos), difference from mean value (DMV), topographic position index (TPI), bathymetric position index (BPI), vector ruggedness measure (VRM), surface area to planar area ratio (SAPA), standard deviation of depth values after removing the influence of slope (AdjSD), and roughness index-elevation (RIE). For a full overview of these functions see Ilich et al. (2023).\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"604\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.8344%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eVariable\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.1656%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDescription\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.8344%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSlope\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.1656%;\"\u003e\n \u003cp\u003eThe maximum rate of change in elevation values.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.8344%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAspect\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.1656%;\"\u003e\n \u003cp\u003eThe orientation of the slope in the downslope direction.\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.8344%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eEastness\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.1656%;\"\u003e\n \u003cp\u003eThe east/west component of the orientation of the slope are calculated as the sine of aspect. Thus, it ranges between -1 (due west) and 1 (due east).\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.8344%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eNorthness\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.1656%;\"\u003e\n \u003cp\u003eThe north/south components of the orientation of the slope are calculated as the cosine of aspect. Thus, it ranges between -1 (due South), and 1 (due North)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.8344%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eProfile curvature\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.1656%;\"\u003e\n \u003cp\u003eThe level of convexity or concavity along the direction of the maximum slope.\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.8344%;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePlanform curvature\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.1656%;\"\u003e\n \u003cp\u003eThe level of convexity or concavity perpendicular to the direction of the maximum slope.\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.8344%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTwisting curvature\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.1656%;\"\u003e\n \u003cp\u003eThe amount of local twisting is measured as the change in slope angle per unit distance along the direction perpendicular to the slope.\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.8344%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMinimum curvature\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.1656%;\"\u003e\n \u003cp\u003ethe minimum curvature in any plane.\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.8344%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMaximum curvature\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.1656%;\"\u003e\n \u003cp\u003eThe maximum curvature in any plane.\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.8344%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMean curvature\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.1656%;\"\u003e\n \u003cp\u003eThe average of the minimum and maximum curvatures.\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.8344%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRelative position\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.1656%;\"\u003e\n \u003cp\u003eAn indication of whether an area is a local topographic high (positive values) or low (negative values). It can be measured in either units of the DTM or standardised based on local topography to be a unitless measure.\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.8344%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eVector ruggedness measure\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.1656%;\"\u003e\n \u003cp\u003eThe dispersion of unit vectors normal to the terrain surface.\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.8344%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSurface area to planar area ratio\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.1656%;\"\u003e\n \u003cp\u003eThe surface area divided by the planar area.\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.8344%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAdjusted standard deviation\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.1656%;\"\u003e\n \u003cp\u003eThe local standard deviation of elevation values in a focal window after removing the influence of slope by fitting a plane to the focal window using ordinary least squares and extracting the residuals.\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.8344%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRoughness index-elevation\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.1656%;\"\u003e\n \u003cp\u003eThe local standard deviation of the residual topography surface, where the residual topography surface is calculated as the DTM minus the focal mean of the DTM.\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cem\u003eSupercells Object-Based Image Analysis\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe R package \u0026lsquo;supercells\u0026rsquo; (Nowosad 2022) was used to create \u0026lsquo;superpixels\u0026rsquo; (groups of spectrally similar pixels) using the Simple Linear Iterative Clustering (SLIC) algorithm (Achanta et al., 2012). SLIC uses an adapted k-means clustering approach modified to work with arbitrary dissimilarity values to efficiently generate superpixels. Superpixels were generated using the backscatter layer as an indicator of seafloor variation using a minimum superpixel size of 3 cells. This size was chosen to prevent the amalgamation of fine-scale habitats into larger, more generalised ones, thereby retaining the heterogeneity of the input data. All other settings were left at default values.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eGround Truth Data\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eTowed Video 2006-2009\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eAcoustically located towed video was used to collect observational data from various habitat mapping studies conducted between 2006 and 2009. Sampling was designed to capture a range of depths and topographic and textural diversity across the region. A total of seven transects perpendicular to the coast (~19.5 linear kilometres) were collected. The system was piloted via a live-stream video at an average speed of 1-1.5 knots and at approximately 1-2m above the seabed with a 45\u0026deg; angle to the seabed. The resulting field of view was approximately 2 x 2 m. An ultra-short baseline transponder attached to the video unit allowed 3-dimensional positioning relative to the vessel\u0026rsquo;s dGPS antenna, resulting in a total propagated error of \u0026plusmn; 3.6 m for video seafloor positioning (Rattray et al. 2014). Video data were classified for biotope using the Combined Classification Scheme (CBiCS; \u003cem\u003esee section below and table.2\u003c/em\u003e).\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eBaited Remote Underwater Video Stations (BRUVS) 2023\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eBRUVS were deployed to assess the diversity, abundance, distribution, and habitat use of marine taxa. Between the 10th and 14th of May 2023, a total of 104 BRUV deployments were conducted within Apollo Marine Park, ranging from 52 to 94 meters depth. Each BRUV array consisted of a weighted stainless-steel frame with two independent stereo-video systems positioned at 90\u0026deg; from the seabed facing forward. Each system included a GoPro Hero7 Black action camera, with pairs of cameras fitted 0.7m apart and angled in at 8 degrees to allow for stereo imaging. BRUV frames were calibrated prior to deployment. From each deployment a still image was extracted presenting a clear view of the surrounding habitat area. Benthic habitats were classified for each cell within a 20-cell grid (4 rows x 5 columns) overlaid on the still image, resulting in a single habitat classification for each site using the CBiCS classification. These habitats were annotated using TransectMeasure (Version 4.21, \u0026reg; SeaGIS).\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eAutonomous Underwater Vehicle (AUV) 2023\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe AUV Nimbus was deployed in September and October 2023 to collect fine-scale imagery of the seabed over representative but spatially discrete areas. AUV deployment sites were selected using a feature space representation of bathymetric data combined with robotic information gathering to plan surveys that comprehensively explored the feature space (Shields et al., 2021). This process created a survey plan comprising 20 missions throughout the AMP. Each mission included three transects of 500 m, split by a 200 m parallel connecting transect. Managed by the Integrated Marine Observing System (IMOS) Autonomous Underwater Vehicle facility at the University of Sydney, the AUV navigated a predetermined course at 2 metres above the seabed and took continuous stereo photographs at 1 photo per second. Each stereo image covered an area of 1.8 x 1.5 m at a resolution of 1360x1024 pixels. More than 150,000 images were taken across the 20 missions. Habitat types were classified for every 50th AUV image post-processed using SQUIDLE+ (Proctor et al., 2018), using 25 randomised point counts for habitat and morphotype categories (Perkins et al. 2016) and later translated into CBiCS to match the other habitat data.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eHabitat Classifications Scheme\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe Combined Biotope Classification Scheme (CBiCS; Edmunds and Flynn 2018) is a hierarchical system used to classify and categorise marine habitats based on environmental and biological characteristics. It integrates components from the Joint Nature Conservation Committee-European Nature Information System (JNCC-EUNIS) and the United States Coastal and Marine Ecological Classification Standard (CMECS), emphasising factors such as biota, substrate, geoform, exposure, and biogeographic area. Beyond habitat types, CBiCS offers crucial ecological context by describing the communities within these habitats, thereby facilitating standardised habitat classification for ecological studies and conservation efforts. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eData Preparation\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eTo prepare the ground truth data for modelling, standard point data extraction methods were utilised for both PB and OBIA approaches. Prior to modelling raster layer values were extracted at ground truthing locales. Within PB approaches this was completed using the pixel value whilst for the OBIA approach the superpixel value was used. If a supercell contained more than one biotope label, the most frequent label was selected, or, in case of a tie, the least represented label in the full dataset was chosen. This resulted in a significant reduction in the number of data points for the OBIA models compared to the PB approach. Sample sizes for either approach is shown in Table 2. Once prepared, data were split into training and testing sets with an 70-30 ratio.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eModel Approach\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eHierarchical random forests models were used to predict class membership. Random Forest (RF) is an ensemble machine learning method that combines tree-based classifiers with bootstrap aggregation (Breiman 2001) and mitigates the tendency of individual decision trees to overfit by aggregating results from multiple trees trained on random subsets of the data (Cutler et al., 2007). RF minimises tree correlation through random variable selection at each split, reducing biases associated with the training data (Rigatti 2017). RF models have proven effective in mapping marine substrates (Misiuk et al., 2019; Gregr et al., 2021; Dalby et al. 2024) and identifying dominant biological communities (Rubidge et al., 2016; Martinez-Santos et al., 2021). Hierarchical random forests (HRF) use predictions from previous models as predictor variables in subsequent models. RFs were created using the \u0026apos;ranger\u0026apos; package in R (version 0.16.0).\u003c/p\u003e\n\u003cp\u003eUsing HRFs allowed the nested nature of habitat classifications to be modelled explicitly. This approach can capture dependencies between habitat types and their subcategories, providing a more accurate and nuanced prediction. For example, \u003cem\u003eHigh energy circalittoral rock with seabed covering sponges\u003c/em\u003e can only occur on \u003cem\u003eCircalittoral rock and other hard substrata\u003c/em\u003e, and so knowing the distribution of circalittoral rock aids predictions of the distribution of sponge reefs.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo optimise the RF model, we tuned the algorithm by varying the following hyperparameters (1) the number of trees (50, 100, 500, 600, 1000), (2) the method used to select the number of features at each split (Square root and Log functions), and (3) the impurity functions (Gini coefficient and Entropy). The best combination of hyperparameters (those that maximised the F1 score) were then retained as the chosen model for that level in the HRF.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eVariable Importance\u0026nbsp;\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eVariable importance in Random Forests is quantified as the mean decrease in overall model accuracy when the values of a focal variable are permuted, and all other variables are held constant (Archer and Kimes 2008). This helps identify predictors with the greatest impact on the model performance. In the context of HRF, variable importance is assessed for both approaches, providing a more nuanced understanding of why certain variables are more critical than others (Fig.S1 \u0026amp; S2.). During analyses, 48 predictors that demonstrated less than 5% variable importance were removed as including too many variables, especially those with\u0026nbsp;minimal importance, can introduce noise into the model, potentially leading to overfitting and reducing interpretability (Galelli et al., 2014; Gregorutti et al., 2017).\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eAccuracy Assessments and Uncertainty\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eAccuracy for both PB and OBIA models was assessed using confusion matrices, overall accuracy, precision, recall, and F1 scores. Additionally, the classifications were visually compared as habitat maps to emphasise interpretability, which is a crucial factor in evaluating the effectiveness of classifiers. The class-based confusion matrix along with uncertainty maps were also used to identify class accuracy imbalances and where these imbalances appear in space. This dual approach ensured a comprehensive assessment of model performance, balancing quantitative metrics with practical visual analysis to gauge the accuracy and reliability of habitat classifications. By analysing the results from both models, we aimed to gain deeper insights into their practical effectiveness, identify potential areas for improvement, investigate the effects of data loss in our methodologies, and explore how each method contributes to the overall goal of accurate and comprehensive benthic habitat mapping.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIn a classification task, Random Forests provide not only class predictions but also prediction probabilities for each class. This capability allows us to calculate and visualise the uncertainty associated with each prediction. By plotting these measures spatially, we can create uncertainty or confidence maps that highlight areas of high and low accuracy. Lower uncertainty typically corresponds to higher accuracy, making these maps useful for identifying regions where predictions are reliable or not (Mitchell et al. 2018). By examining these maps, researchers can better understand the reliability of the classification and identify where additional validation or refinement may be needed (Wager et al. 2014).\u0026nbsp;\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eMapping of biotopes revealed that Apollo Marine Park and the adjacent Otway coastline was dominated by sand substrata, with sponge and macroalgae communities to the north and northwest (Fig.5). Increased biodiversity in north and northwestern areas coincided with the shallower depths and the presence of a hard seafloor.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003ePaired t-tests showed no significant differences in precision, recall, or F1-score between the PB and OBIA models across BC levels and found only negligible differences in accuracy (See Table 2).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2\u003c/strong\u003e The results of the paired t-tests, and a comparison of accuracy, for PB and OBIA approaches at each hierarchical level, where a = 0.05.\u0026nbsp;\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" align=\"\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 25%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 25%;\"\u003e\n \u003cp\u003eBC2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 25%;\"\u003e\n \u003cp\u003eBC3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 25%;\"\u003e\n \u003cp\u003eBC4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 25%;\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 25%;\"\u003e\n \u003cp\u003et2 = -0.99, p = 0.42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 25%;\"\u003e\n \u003cp\u003et2 = -1.12, p = 0.34\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 25%;\"\u003e\n \u003cp\u003et2 = 1.37, p = 0.21\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 25%;\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 25%;\"\u003e\n \u003cp\u003et2 = -0.15, p = 0.89\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 25%;\"\u003e\n \u003cp\u003et2 = 0.66, p = 0.60\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 25%;\"\u003e\n \u003cp\u003et2 = 1.07, p = 0.32\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 25%;\"\u003e\n \u003cp\u003eF1-score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 25%;\"\u003e\n \u003cp\u003et2 = -1.10, p = 0.39\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 25%;\"\u003e\n \u003cp\u003et2 = 0.49, p = 0.66\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 25%;\"\u003e\n \u003cp\u003et2 = 1.7, p = 0.14\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 25%;\"\u003e\n \u003cp\u003eAccuracy Difference\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 25%;\"\u003e\n \u003cp\u003e0.015\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 25%;\"\u003e\n \u003cp\u003e0.014\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 25%;\"\u003e\n \u003cp\u003e0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cem\u003eComparison of PB BC2 and OBIA BC2 \u003cem\u003eModels\u003c/em\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cem\u003e\u003cstrong\u003eTable 3\u003c/strong\u003e Comparison of Precision, Recall, F1-Score, and Accuracy between Object-Based Image Analysis (OBIA) and Pixel-Based (PB) random forest models for biotope classification at the BC2 level of hierarchy.\u0026nbsp;\u003c/em\u003e\u003c/em\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" align=\"\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003eOBIA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003ePB\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e0.76\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e0.89\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003eF1-score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e0.79\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e0.86\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e84%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e85%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eWhen compared at the BC2 level, model accuracy was nearly identical (Table 3). Both models predict three main habitat classes: \u003cem\u003eCircalittoral rock, Infralittoral rock, and Sublittoral sediment\u003c/em\u003e. The spatial extent of these classes in the PB map appears more fragmented, with smaller patches, indicative of the PB model\u0026rsquo;s tendency to produce more granular \u0026ldquo;salt-and-pepper\u0026rdquo; patterns, especially in transitional zones between classes. In contrast, the OBIA model predicted classes in broader, more continuous patches, reflecting a more generalized and smoother representation of biotopes present.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eSupplementary Figure 3 visualised differences in areal extents predicted for each biotope. Although both the PB and OBIA models predicted similar spatial extents of key biotopes, differences were present when assessed at the fine scale. The OBIA model predicted slightly larger extents for Circalittoral rock and Infralittoral rock, while the PB model predicted a slightly greater coverage of Sublittoral sediment. Further, the PB model had a prediction ranging from 0.33 to 1, with a mean of 0.80. This was marginally more uncertain than the equivalent OBIA model (rand of 0.35 to 1, mean of 0.86).\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eComparison of PB BC3 and OBIA BC3 ModelS\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eTable 4\u003c/strong\u003e Comparison of Precision, Recall, F1-Score, and Accuracy between Object-Based Image Analysis (OBIA) and Pixel-Based (PB) random forest models for biotope classification at the BC3 level of hierarchy.\u0026nbsp;\u003c/em\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" align=\"\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 204px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 199px;\"\u003e\n \u003cp\u003eOBIA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 198px;\"\u003e\n \u003cp\u003e\u0026nbsp;PB\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 204px;\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 199px;\"\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 198px;\"\u003e\n \u003cp\u003e0.79\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 204px;\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 199px;\"\u003e\n \u003cp\u003e0.87\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 198px;\"\u003e\n \u003cp\u003e0.92\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 204px;\"\u003e\n \u003cp\u003eF1-score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 199px;\"\u003e\n \u003cp\u003e0.88\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 198px;\"\u003e\n \u003cp\u003e0.82\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 204px;\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 199px;\"\u003e\n \u003cp\u003e93%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 198px;\"\u003e\n \u003cp\u003e92%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eAs with BC2 predictions, the accuracy of PB and OBIA BC3 model was largely comparable (Table 4). Although, marginal difference was found in precision between the model types. Both models predict five habitat classes: \u003cem\u003eSublittoral course sediment, High energy open coast circalittoral rock, High energy infralittoral rock, Sublittoral sand and muddy sand, Sublittoral seaweed on sediment.\u003c/em\u003e Mirroring the results from the BC2 level of hierarchy; the PB model predicted more fragmented biotope classes with stronger \u0026ldquo;salt-and-pepper\u0026rdquo; effects that were particularly noticeable in transitional zones between biotopes. In contrast, the OBIA model presents even more pronounced generalisation, with broader and more continuous patches. This generalised representation is particularly evident in the shallow, diverse habitats, where the OBIA model predicts a much less complex arrangement of biotopes (Fig.7.).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAt the BC3 hierarchical level, the pixel-based (PB) model shows an uncertainty range of 0.20 to 1, with a mean of 0.91, indicating a high level of confidence in its predictions. The OBIA model, at this level, has an uncertainty range of 0.26 to 1 and a mean of 0.88, reflecting slightly lower but still high confidence. Supplementary Figure 4 illustrates that both models maintain similar spatial patterns of confidence. Compared to the previous level, this higher hierarchical level shows increased overall confidence while preserving the same general patterns across both approaches.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eComparison of Pixel-based BC4 and OBIA BC4 Models\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eTable 5\u003c/strong\u003e Comparison of Precision, Recall, F1-Score, and Accuracy between Object-Based Image Analysis (OBIA) and Pixel-Based (PB) random forest models for biotope classification at the BC3 level of hierarchy.\u0026nbsp;\u003c/em\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" align=\"\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003eOBIA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003ePB\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e0.69479\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e0.74242\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e0.67942\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e0.72448\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003eF1-score\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e0.68169\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e0.72957\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003eAccuracy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e82.6%\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 33.3333%;\"\u003e\n \u003cp\u003e84.4%\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eAs observed at previous hierarchical levels, the accuracy of the PB and OBIA BC4 models is largely comparable (Table 5). Both models predicted the following nine habitat classes: \u003cem\u003eCircalittoral fine sand, High-energy circalittoral rock with seabed covering sponges, High-energy Durvillaea communities, Low-complexity circalittoral rock with non-erect sponges, High-energy lower infralittoral zone, Moderate-complexity circalittoral rock with prominent sea plumes, sea tulips, and hydroid fans, thallose red algae on sand with ground surge, sandy low-profile reef wave surge communities, and infralittoral shell and sand mixes.\u0026nbsp;\u003c/em\u003eNotably, the OBIA model fails to identify \u003cem\u003eHigh-energy Durvillaea communities\u003c/em\u003e, which are captured by the PB model.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe PB model predicted more fragmented biotope classes with a stronger \u0026ldquo;salt-and-pepper\u0026rdquo; effect, particularly noticeable in transitional zones between biotopes. In contrast, the OBIA model showed more pronounced generalization, resulting in broader and more continuous patches. These differences were particularly evident in shallow, diverse habitats, where the OBIA model predicted a much less complex arrangement of biotopes (Fig. 9).\u003c/p\u003e\n\u003cp\u003eAt the final hierarchical level, the PB model displayed an uncertainty range of 0.17 to 1, with a mean of 0.89, indicating a high level of confidence in its predictions. Similarly, the OBIA model exhibited an uncertainty range of 0.20 to 1, with the same mean of 0.89, reflecting comparable overall confidence. Supplementary Figure 5 shows that both models continue to display similar spatial patterns of confidence. This level maintains high confidence, consistent with previous hierarchical levels, while preserving general patterns across both approaches.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eVisual comparisons of biotope classification between the PB and OBIA models reveal differences at both broad and fine scales. The highlighted areas illustrate how the models vary in their predictions, particularly in regions with higher heterogeneity. The OBIA model tends to capture broader biotope categories, while the PB model offers more detailed and nuanced classifications, leading to discrepancies in area coverage for certain biotopes. These differences are evident in the zoomed-in sections, which demonstrate how each model performs in predicting biotope distributions. This comparison provides insights into the varying levels of detail and coverage between the models.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we developed benthic biotope maps for a northern section of Apollo Marine Park and the adjacent Cape Otway coastline, VIC, Australia. Maps were created using habitat/environment relationships derived from hierarchical random forest models with either PB or OBIA sets of MBES-derived input features. Our findings revealed that the study site is dominated by shell and sand mixes, with increased biodiversity in sponge and macroalgae communities towards the north and northwest, corresponding with geoform complexities and hard-bottom seafloors. We compared various metrics such as the accuracy and interpretability of the hierarchical random forest classifications. While the PB approach achieved a higher overall accuracy than the OBIA approaches, the differences were not statistically significant. The major differences between the models' classifications are evident when examining specific areas of interest (see Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e). Notably, the \u0026lsquo;salt and pepper\u0026rsquo; effect in the PB map contrasts with the more homogenous outputs of the OBIA map, which provides more clearly defined habitat boundaries. Analysing these results and their underlying causes highlight the weaknesses and strengths of both the PB and OBIA approaches.\u003c/p\u003e \u003cp\u003eAcross the three classification levels presented in this study, generally, the PB approach demonstrates higher average precision and recall, leading to better overall performance metrics, particularly in achieving a more refined and detailed classification of biotopes. This can be attributed to their fine-grain approach, evaluating each pixel independently and capturing subtle variations in habitat characteristics, making them effective for detecting and correctly classifying a broader range of biotopes on a smaller scale. Although the majority of studies comparing OBIA and pixel-based (PB) classification tend to favour OBIA due to its ability to group pixels into meaningful objects, there are some studies that have reported PB methods outperforming OBIA under specific conditions. For example, research from Universiti Teknologi MARA (Othman et al., \u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) found that PB methods outperformed OBIA for coral reef mapping off Mabul Island, Malaysia, using Sentinel-2B and SPOT-7 imagery. The PB approach yielded higher accuracy (97.5% for Sentinel-2 and 90% for SPOT-7) compared to OBIA (82.81% and 87.05%, respectively). Othman explained that this is likely due to PB\u0026rsquo;s ability to capture fine-scale habitat details in highly complex and heterogeneous environments. Additionally, Blaschke (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2010\u003c/span\u003e) emphasized that while object-based image analysis (OBIA) has its advantages in some situations, pixel-based methods are essential for accurately capturing detailed spatial patterns, especially in high-resolution environments.\". This study underscored the advantages of PB methods in achieving refined habitat boundaries and detecting subtle habitat features. These findings highlight the robustness of PB methods in delivering high-resolution habitat maps, which is critical for effective marine spatial planning and management.\u003c/p\u003e \u003cp\u003eIn contrast, OBIA models offer more streamlined classifications with a tendency towards broader habitat categories. This understanding of OBIA models is mirrored in Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e within the \u003cspan refid=\"Sec16\" class=\"InternalRef\"\u003eresults\u003c/span\u003e section. Although there are no significant differences identified in the PB and OBIA approach\u0026rsquo;s performance metrics, there are clear differences in their classification approaches, whereby the OBIA model classifies with a far more general, or broader, brush than the PB approach. OBIA\u0026rsquo;s strengths lie in object-based segmentation, which groups adjacent pixels into meaningful objects and leverages spatial relationships for classification. Previous literature has shown that OBIA methods outperform PB approaches in various ecological and remote sensing applications (Mishra and Crews \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2014\u003c/span\u003e; Wallace et al., \u003cspan citationid=\"CR80\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Nababan et al., \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). For example, Cui et al., (\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) found that OBIA significantly improved classification accuracy for benthic habitats due to its ability to incorporate spatial context and shape information. Similarly, Zheng et al., (\u003cspan citationid=\"CR86\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) demonstrated that OBIA was more effective in classifying coastal wetland vegetation, highlighting the method\u0026rsquo;s advantage in utilising both spectral and spatial data for better delineation of habitat boundaries and diverse vegetation types. However, this trend was not observed in our study. The OBIA model's tendency to generalise segments may have reduced the predictive utility of the data, especially given the heterogeneous nature of the habitat area in our research compared to the more homogenous environments in the cited studies.\u003c/p\u003e \u003cp\u003eWhile the OBIA approach generally results in higher accuracy for broader biotope categories and improved specificity, it may overlook finer details and lead to misclassification. This is especially relevant for rarer or less represented biotopes. For instance, Othman et al., (2013) demonstrated that while object-based image analysis (OBIA) methods are generally effective, they can overlook finer details compared to pixel-based classifications, particularly in complex benthic habitats. This limitation was evident in the OBIA model\u0026rsquo;s performance from this study when classifying \u0026ldquo;Thallose red algae on sand with ground surge\u0026rdquo;, \u0026ldquo;Ecklonia-Phyllospora communities\u0026rdquo;, and the \u0026ldquo;High energy infralittoral zone\u0026rdquo;. Markedly, the OBIA model failed to learn to classify \u0026ldquo;Durvillaea communities\u0026rdquo;, as the pathway of the OBIA approach significantly reduced the number of sample points, which adversely influenced the model's performance and consequently lowered its overall performance metrics. This limitation meant that the OBIA approach suffered in comparison to the data-rich PB approach, which was better equipped to capture finer details and provide more accurate classifications for a broader range of biotopes. Trade-offs between specificity and map complexity have considerable implications for ecological management (Lecours et al., \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2015\u003c/span\u003e) and it is likely that a middle ground is needed. However, it is probable that the preferred level of complexity is situation and application specific based on the needs of managers and available data. As such, authors suggest that trade-offs between PB and OBIA approaches be assessed more commonly in remote sensing projects and the differences clearly communicated to managers to ensure that informing conservation efforts remains centre stage.\u003c/p\u003e \u003cp\u003eAnalysing both approaches' variable importance (Fig. \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e.) and their density plots (Fig. S2.) within the supplementary materials sheds light on the underlying mechanisms driving the two models\u0026rsquo; workings. The density plot shows that the OBIA model distributes importance across a broader range of variables, reflecting its tendency to utilise spatial relationships and context in its classifications (Hossain and Chen \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). Gao and Mas (\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2008\u003c/span\u003e) found that OBIA methods excel in leveraging multiple scales of information, as evidenced here by the OBIA model\u0026rsquo;s top ten most important variables, which include various scales of elevation and landscape features. This aligns with Wahidin et al. (\u003cspan citationid=\"CR79\" class=\"CitationRef\"\u003e2015\u003c/span\u003e), who demonstrated that OBIA techniques can effectively classify diverse coral reef habitats by incorporating contextual information and spatial relationships\u0026mdash;an ability mirrored in the broader variable importance observed in the OBIA model.\u003c/p\u003e \u003cp\u003eBy comparison, the PB model displays a more concentrated distribution of variable importance, with key variables dominating the classification process. Sibarddin et al., (2018) showed that PB methods rely heavily on specific, detailed features to achieve high classification accuracy, which likely contributes to the PB approach\u0026rsquo;s higher precision and recall by capturing fine-scale variations in habitat characteristics. This is supported by Anggoro et al., (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2018\u003c/span\u003e), who found that pixel-based methods performed better in classifying detailed benthic habitats due to their fine-scale feature extraction capabilities. Furthermore, Whiteside et al., (\u003cspan citationid=\"CR85\" class=\"CitationRef\"\u003e2011\u003c/span\u003e) compared OBIA and PB methods for savanna mapping and found that while OBIA provided a more generalised classification, PB methods achieved greater detail and accuracy in distinguishing between different vegetation types. The concentrated variable importance observed in the PB model, as also reported by Othman et al., (\u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e2023\u003c/span\u003e), highlights its effectiveness in leveraging specific features to enhance precision and recall in classification tasks (Baker et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2013\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eFigure S3, S4, \u0026amp; S5 in the supplementary materials illustrates the model certainty across all biotope hierarchy classifications. Notably, regions with the highest heterogeneity, where biotope types are likely to overlap, also show the greatest uncertainty in both the PB and OBIA models. This pattern highlights the difficulties in distinguishing closely related biotope types in complex or transitional environments. These overlapping biotopes are often those that were least represented or under-sampled in the ground truth data (see Table.3.). While the PB approach captures these areas with higher resolution, providing finer-scale detail, both models reveal that these high-heterogeneity regions are prone to increased ambiguity. This suggests that targeted data collection in these under-sampled, high-uncertainty areas could substantially enhance model performance and improve biotope classification accuracy for both approaches.\u003c/p\u003e \u003cp\u003eThe implications of these findings are significant for marine spatial planning and management. The higher precision and recall of the PB models suggest they are more adept at capturing the fine-grained details necessary for detailed biotope classification. This can be particularly valuable in conservation planning, where understanding the specific distribution of various biotopes can inform more targeted and effective management strategies (Rubbens et al. \u003cspan citationid=\"CR66\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). On the other hand, the OBIA approach\u0026rsquo;s ability to provide clearer, more homogeneous classification makes it an excellent tool for broader scale mapping efforts where the focus may be on general habitat distribution rather than detailed biotope distinctions. This duality of PB and OBIA methods underscores the importance of selecting the appropriate classification technique based on the specific objectives of a given marine spatial planning project. The integration of PB and OBIA methods, as suggested by Ierodiaconou et al., (\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2018\u003c/span\u003e) and mirrored here, has the potential to leverage the strengths of both approaches, resulting in improved biotope classification accuracy. This approach could provide detailed and precise classifications while maintaining clear and coherent habitat boundaries, offering a more accurate and interpretable representation of marine biotopes.\u003c/p\u003e \u003cp\u003eFuture improvements in classification techniques could significantly enhance the accuracy and utility of both PB and OBIA methods. Advances in machine learning, such as deep learning techniques, have the potential to refine PB methods by capturing more complex patterns and relationships in the data (Misiuk et al., \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). For OBIA methods, developments in spatial statistics and the incorporation of higher-resolution data from emerging technologies like drones or autonomous vehicles could provide more detailed inputs and improve segmentation and classification performance. Exploring these technological advancements and their applications in marine habitat mapping can drive further innovation and effectiveness.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn this study, we compared the performance of hierarchical random forests built using pixel-based (PB) or object-based image analysis (OBIA) methods in the classification of benthic biotopes for the northern section of Apollo Marine Park and the adjacent Cape Otway coastline, Victoria, Australia. The findings reveal that there was little to separate the two approaches\u0026rsquo; accuracy and performance metrics. However, when visualised, their large-scale predictions exhibited clear differences. These results underscore the importance of selecting the appropriate sampling and classification techniques based on specific project objectives. PB methods are well-suited for detailed biotope classification, which is valuable for targeted conservation efforts, whereas OBIA methods are effective for broader-scale habitat mapping. Future research should explore integrating PB and OBIA methods to combine their strengths, potentially enhanced by advancements in machine learning and emerging technologies such as deep learning and higher resolution bathymetric data. This study highlights the potential for future innovations to improve both the accuracy and utility of habitat classification techniques for marine spatial planning and management.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cem\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eFor this project, Deakin University secured the Marine Parks grant (4-HAZCNTR) to provide funding for the collection of the data used in the Apollo Marine Park and along the Cape Otway coastline.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eCompeting Interests \u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe authors have no relevant financial or non-financial interests to disclose.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eAuthor Contributions\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eAll authors contributed to the study conception and design. Material preparation and data collection were performed by Daniel Ierodiaconou and Henry Simmons. Analysis was performed by Henry Simmons. The first draft of the manuscript was written by Henry Simmons and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eData Availability \u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003eMultibeam Echosounder (MBES) Data: Available through the \u003cem\u003eAusSeabed\u003c/em\u003e initiative, a national repository for seabed mapping data in Australian waters (https://www.ausseabed.gov.au/).\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eBaited Remote Underwater Video Systems (BRUVS) Data: Accessible via \u003cem\u003eGlobalArchive\u003c/em\u003e, an online platform for managing, storing, and sharing underwater imagery and associated metadata collected from BRUVS surveys worldwide (https://globalarchive.org/).\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eTowed Video Data: Hosted on \u003cem\u003eCoastKit\u003c/em\u003e, a centralized database that provides access to towed video imagery collected along Australia\u0026rsquo;s coastline (https://mapshare.vic.gov.au/coastkit/).\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eAutonomous Underwater Vehicle (AUV) Data: Available through \u003cem\u003eSQUIDLE+\u003c/em\u003e, a platform designed for the annotation and analysis of underwater imagery, including data collected by AUVs. SQUIDLE+ facilitates the classification of marine species and habitats using advanced tools and AI-driven workflows (https://squidle.org/). \u0026nbsp;\u0026nbsp;\u003c/li\u003e\n\u003c/ul\u003e\u003ch2\u003eAcknowledgement\u003c/h2\u003e\u003cp\u003eThe authors respectfully acknowledge the Eastern Maar Aboriginal Corporation, the Traditional Owners of the land, sea, and sky country where this work was conducted. We extend our gratitude to Parks Australia for their funding through the Our Marine Parks Grant Program, which supported multibeam sonar data collection, as well as the deployment of stereo-baited remote underwater video and autonomous underwater video (AUV) in the Apollo Marine Park.We also thank the Department of Energy, Environment, and Climate Action for funding the multibeam data and towed video collection in State waters through the Victorian Coastal Monitoring Program. Further appreciation is extended to the IMOS AUV Facility for supporting the AUV deployments, and to the IMOS Understanding of Marine Imagery (UMI) Sub-Facility for image scoring and archiving via the Squidle+ platform.Data were sourced from Australia\u0026rsquo;s Integrated Marine Observing System (IMOS) \u0026ndash; IMOS is enabled by the National Collaborative Research Infrastructure Strategy (NCRIS). We are grateful to Dr Ben Misiuk for his expertise in backscatter harmonization processes, and to Dr Sasha Whitmarsh, Dr Jacquomo Monk, and Dr Ariell Friedman for their training and assistance in habitat classification.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eAchanta, R., Shaji, A., Smith, K., Lucchi, A., Fua, P. and S\u0026uuml;sstrunk, S., 2012. SLIC superpixels compared to state-of-the-art superpixel methods. IEEE Trans. Pattern Anal. Mach., \u003cem\u003e34\u003c/em\u003e(11), pp.2274-2282. https://doi.org/10.1109/TPAMI.2012.120\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eAmini, Z.Z., Adabi, M.H., Burrett, C.F. and Quilty, P.G., 2004. Bryozoan distribution and growth form associations as a tool in environmental interpretation, Tasmania, Australia. \u003cem\u003eSediment. Geol.\u003c/em\u003e, \u003cem\u003e167\u003c/em\u003e(1-2), pp.1-15. https://doi.org/10.1016/j.sedgeo.2004.01.010\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eAnggoro, A., Sumartono, E., Siregar, V.P., Agus, S.B., Purnama, D., Puspitosari, D.A., Listyorini, T. and Sulistyo, B., 2018, November. Comparing object-based and pixel-based classifications for benthic habitats mapping in Pari Islands. In \u003cem\u003eJ. Phys. Conf. Ser.\u0026nbsp;\u003c/em\u003e(Vol. 1114, No. 1, p. 012049). https://doi.org/10.1088/1742-6596/1114/1/012049\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eArcher, K.J. and Kimes, R.V., 2008. Empirical characterization of random forest variable importance measures. \u003cem\u003eComput. Stat. Data Anal.\u003c/em\u003e, \u003cem\u003e52\u003c/em\u003e(4), pp.2249-2260. https://doi.org/10.1016/j.csda.2007.08.015\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eBaker, B.A., Warner, T.A., Conley, J.F. and McNeil, B.E., 2013. Does spatial resolution matter? A multi-scale comparison of object-based and pixel-based methods for detecting change associated with gas well drilling operations. \u003cem\u003eInt. J. Remote Sens.\u003c/em\u003e, \u003cem\u003e34\u003c/em\u003e(5), pp.1633-1651. https://doi.org/10.1080/01431161.2012.724540\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eBarry, J.P., Widdicombe, S. and Hall-Spencer, J.M., 2011. Effects of ocean acidification on marine biodiversity and ecosystem function. \u003cem\u003eOcean acidification\u003c/em\u003e, pp.192-209. https://doi.org/10.1093/oso/9780199591091.003.0015\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eBeardmore, J.A., Mair, G.C. and Lewis, R.I., 1997. Biodiversity in aquatic systems in relation to aquaculture. \u003cem\u003eAquac. Res.\u003c/em\u003e, \u003cem\u003e28\u003c/em\u003e(10), pp.829-839. https://doi.org/10.1046/j.1365-2109.1997.00947.x\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eBelgiu, M. and Drăguţ, L., 2016. Random forest in remote sensing: A review of applications and future directions.\u003cem\u003e\u0026nbsp;ISPRS J. Photogramm. Remote Sens\u003c/em\u003e, \u003cem\u003e114\u003c/em\u003e, pp.24-31. https://doi.org/10.1016/j.isprsjprs.2016.01.011\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eBenz, U.C., Hofmann, P., Willhauck, G., Lingenfelder, I. and Heynen, M., 2004. Multi-resolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready information. \u003cem\u003eISPRS J. Photogramm. Remote Sens\u003c/em\u003e, \u003cem\u003e58\u003c/em\u003e(3-4), pp.239-258. https://doi.org/10.1016/j.isprsjprs.2003.10.002\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eBerhane, T.M., Lane, C.R., Wu, Q., Anenkhonov, O.A., Chepinoga, V.V., Autrey, B.C. and Liu, H., 2017. Comparing pixel-and object-based approaches in effectively classifying wetland-dominated landscapes. \u003cem\u003eJ. Remote Sens.\u003c/em\u003e, \u003cem\u003e10\u003c/em\u003e(1), p.46. https://doi.org/10.3390/rs10010046\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eBlaschke, T., 2010. Object based image analysis for remote sensing. \u003cem\u003eISPRS J. Photogramm. Remote Sens\u003c/em\u003e, \u003cem\u003e65\u003c/em\u003e(1), pp.2-16. https://doi.org/10.1016/j.isprsjprs.2009.06.004\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eBreiman, L., 2001. Random forests. \u003cem\u003eMach. Learn.\u003c/em\u003e, \u003cem\u003e45\u003c/em\u003e, pp.5-32. http://dx.doi.org/10.1023/A:1010933404324\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eBrown, 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. \u003cem\u003eEstuar. Coast. Shelf Sci., 92\u003c/em\u003e(3), pp.502-520. https://doi.org/10.1016/j.ecss.2011.02.007\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eBuhl-Mortensen, L., Buhl-Mortensen, P., Dolan, M.J.F. and Gonzalez-Mirelis, G., 2015. Habitat mapping as a tool for conservation and sustainable use of marine resources: Some perspectives from the MAREANO Programme, Norway. \u003cem\u003eJ. Sea Res.\u003c/em\u003e, \u003cem\u003e100\u003c/em\u003e, pp.46-61. https://doi.org/10.1016/j.seares.2014.10.014\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eBuonocore, E., Grande, U., Franzese, P.P. and Russo, G.F., 2021. Trends and evolution in the concept of marine ecosystem services: an overview. \u003cem\u003eWater\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e(15), p.2060. https://doi.org/10.3390/w13152060\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eCogan, C.B., Todd, B.J., Lawton, P. and Noji, T.T., 2009. The role of marine habitat mapping in ecosystem-based management. \u003cem\u003eICES\u0026nbsp;Mar. Sci.\u003c/em\u003e, \u003cem\u003e66\u003c/em\u003e(9), pp.2033-2042. https://doi.org/10.1093/icesjms/fsp214\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eCui, Q., Pan, H., Zhang, K., Li, X. and Sun, H., 2023. Multiscale and multisubgraph-based segmentation method for ocean remote sensing images. \u003cem\u003eIEEE Trans. Geosci. Remote Sens\u003c/em\u003e, \u003cem\u003e61\u003c/em\u003e, pp.1-20. https://doi.org/10.1109/TGRS.2023.3247697\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eCutler, D.R., Edwards Jr, T.C., Beard, K.H., Cutler, A., Hess, K.T., Gibson, J. and Lawler, J.J., 2007. Random forests for classification in ecology. \u003cem\u003eEcol.\u003c/em\u003e, \u003cem\u003e88\u003c/em\u003e(11), pp.2783-2792. https://doi.org/10.1890/07-0539.1\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eDalby, O., Kennedy, D.M., McCarroll, R.J., Young, M. and Ierodiaconou, D., 2024. Mapping surface sediment characteristics in enclosed shallow‐marine environments using spatially balanced designs and the random forest algorithm. \u003cem\u003eEarth Surf. Process. Landf.\u003c/em\u003e https://doi.org/10.1002/esp.5864\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eEddy, T.D., Lam, V.W., Reygondeau, G., Cisneros-Montemayor, A.M., Greer, K., Palomares, M.L.D., Bruno, J.F., Ota, Y. and Cheung, W.W., 2021. Global decline in capacity of coral reefs to provide ecosystem services. \u003cem\u003eOne Earth\u003c/em\u003e, \u003cem\u003e4\u003c/em\u003e(9), pp.1278-1285. https://doi.org/10.1016/j. oneear.2021.08.016\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eEdmunds, M. and Flynn, A., 2018. CBiCS classification of Victorian biotopes. \u0026nbsp;https://doi.org/10.13140/RG.2.2.16919.29603\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eEsetlili, M.T., Balcik, F.B., Sanli, F.B., Kalkan, K., Ustuner, M., Goksel, C., Gazioğlu, C. and Kurucu, Y., 2018. Comparison of object and pixel-based classifications for mapping crops using Rapideye imagery: a case study of Menemen Plain, Turkey. \u003cem\u003eInt. j. environ. geoinformatics\u003c/em\u003e, 5(2), pp.231-243. https://doi.org/10.30897/ijegeo.442002\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eField, C.B., Behrenfeld, M.J., Randerson, J.T. and Falkowski, P., 1998. Primary production of the biosphere: integrating terrestrial and oceanic components. \u003cem\u003eJ. Sci\u003c/em\u003e, \u003cem\u003e281\u003c/em\u003e(5374), pp.237-240. https://doi.org/10.1126/science.281.5374.237 \u0026nbsp;\u003c/li\u003e\n \u003cli\u003eFingas, M., 2019. Remote sensing for marine management. In \u003cem\u003eWorld seas: An environmental evaluation\u003c/em\u003e (pp. 103-119). Academic Press. https://doi.org/10.1016/B978-0-12-805052-1.00005-X\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eFontoura, L., D\u0026rsquo;agata, S., Gamoyo, M., Barneche, D.R., Luiz, O.J., Madin, E.M., Eggertsen, L. and Maina, J.M., 2022. Protecting connectivity promotes successful biodiversity and fisheries conservation. \u003cem\u003eJ. Sci.\u003c/em\u003e, \u003cem\u003e375\u003c/em\u003e(6578), pp.336-340. https://doi.org/10.1126/science.abg4351\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eGalelli, S., Humphrey, G.B., Maier, H.R., Castelletti, A., Dandy, G.C. and Gibbs, M.S., 2014. An evaluation framework for input variable selection algorithms for environmental data-driven models. \u003cem\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;Environ. Model. Softw.\u003c/em\u003e, \u003cem\u003e62\u003c/em\u003e, pp.33-51. https://doi.org/10.1016/j.envsoft.2014.08.015\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eGamfeldt, L., Lefcheck, J.S., Byrnes, J.E., Cardinale, B.J., Duffy, J.E. and Griffin, J.N., 2015. Marine biodiversity and ecosystem functioning: what\u0026apos;s known and what\u0026apos;s next?. \u003cem\u003eOikos\u003c/em\u003e, \u003cem\u003e124\u003c/em\u003e(3), pp.252-265. https://doi.org/10.1111/oik.01549\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eGao, Y. and Mas, J.F., 2008. A comparison of the performance of pixel-based and object-based classifications over images with various spatial resolutions. \u003cem\u003eJ. Earth Sci.\u003c/em\u003e, \u003cem\u003e2\u003c/em\u003e(1), pp.27-35. https://doi.org/10.30897/ijegeo.442002\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eGill, P.C. (2002). A blue whale (Balaenoptera musculus) feeding ground in a southern Australian coastal upwelling zone. \u003cem\u003eJ. Cetac. Res. Manage.,\u003c/em\u003e 4(2): 179-184. https://doi.org/10.47536/jcrm.v4i2.854\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eGrebby, S., Field, E. and Tansey, K., 2016. Evaluating the use of an object-based approach to lithological mapping in vegetated terrain. \u003cem\u003eRemote Sens.,\u003c/em\u003e 8(10), p.843. https://doi.org/10.3390/rs8100843\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eGregorutti, B., Michel, B. and Saint-Pierre, P., 2017. Correlation and variable importance in random forests. \u003cem\u003eStat. Comput., 27\u003c/em\u003e, pp.659-678. https://doi.org/10.1007/s11222-016-9646-1\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eGregr, E.J., Haggarty, D.R., Davies, S.C., Fields, C. and Lessard, J., 2021. Comprehensive marine substrate classification applied to Canada\u0026rsquo;s Pacific shelf. \u003cem\u003ePloS one\u003c/em\u003e, \u003cem\u003e16\u003c/em\u003e(10), p.e0259156. https://doi.org/10.1371/journal.pone.0259156\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eHasan, R. C., Ierodiaconou, D., Laurenson, L. and Schimel, A., 2014. Integrating multibeam backscatter angular response, mosaic and bathymetry data for benthic habitat mapping. \u003cem\u003ePlos one\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e(5), p.e97339. https://doi.org/10.1371/journal.pone.0097339\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eHasan, R.C., Ierodiaconou, D. and Laurenson, L., 2012. Combining angular response classification and backscatter imagery segmentation for benthic biological habitat mapping. \u003cem\u003eEstuar. Coast. Shelf Sci.\u003c/em\u003e, \u003cem\u003e97\u003c/em\u003e, pp.1-9. https://doi.org/10.1016/j.ecss.2011.10.004\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eHay, G.J., Marceau, D.J., Dube, P. and Bouchard, A., 2001. A multiscale framework for landscape analysis: object-specific analysis and upscaling. \u003cem\u003eLandsc. Ecol.\u003c/em\u003e, \u003cem\u003e16\u003c/em\u003e, pp.471-490. https://doi.org/10.1023/A:1013101931793\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eHossain, M.D. and Chen, D., 2019. Segmentation for Object-Based Image Analysis (OBIA): A review of algorithms and challenges from remote sensing perspective. \u003cem\u003eISPRS J. Photogramm. Remote Sens\u003c/em\u003e, \u003cem\u003e150\u003c/em\u003e, pp.115-134. https://doi.org/10.1016/j.isprsjprs.2019.02.009\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eIerodiaconou, D., Schimel, A.C., Kennedy, D., Monk, J., Gaylard, G., Young, M., Diesing, M. and Rattray, A., 2018. Combining pixel and object based image analysis of ultra-high resolution multibeam bathymetry and backscatter for habitat mapping in shallow marine waters. \u003cem\u003eMar. Geophys. Res.\u003c/em\u003e, \u003cem\u003e39\u003c/em\u003e, pp.271-288. https://doi.org/10.1007/s11001-017-9338-z\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eIlich, A.R., Misiuk, B., Lecours, V. and Murawski, S.A., 2023. MultiscaleDTM: An open‐source R package for multiscale geomorphometric analysis. \u003cem\u003eTrans. GIS.\u003c/em\u003e, \u003cem\u003e27\u003c/em\u003e(4), pp.1164-1204. https://doi.org/10.1111/tgis.13067\u003c/li\u003e\n \u003cli\u003eIntergovernmental Oceanographic Commission, 2018. United Nations Decade of Ocean Science for Sustainable Development (2021\u0026ndash;2030): call for nomination to the Executive Planning Group (EPG).\u003c/li\u003e\n \u003cli\u003eLecours, V., Devillers, R., Schneider, D.C., Lucieer, V.L., Brown, C.J. and Edinger, E.N., 2015. Spatial scale and geographic context in benthic habitat mapping: review and future directions. \u003cem\u003eMar. Ecol. Prog. Ser.\u003c/em\u003e, \u003cem\u003e535\u003c/em\u003e, pp.259-284. https://doi.org/10.3354/meps11378\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eLu, D. and Weng, Q., 2007. A survey of image classification methods and techniques for improving classification performance. \u003cem\u003eInt. J. Remote Sens.\u003c/em\u003e, \u003cem\u003e28\u003c/em\u003e(5), pp.823-870. https://doi.org/10.1080/01431160600746456\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eMartin, C.L., Momtaz, S., Gaston, T. and Moltschaniwskyj, N.A., 2016. A systematic quantitative review of coastal and marine cultural ecosystem services: current status and future research. \u003cem\u003eMar. Policy\u003c/em\u003e, \u003cem\u003e74\u003c/em\u003e, pp.25-32. https://doi.org/10.1016/j.marpol.2016.09.004\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eMart\u0026iacute;nez-Santos, P., Aristiz\u0026aacute;bal, H.F., D\u0026iacute;az-Alcaide, S. and G\u0026oacute;mez-Escalonilla, V., 2021. Predictive mapping of aquatic ecosystems by means of support vector machines and random forests. \u003cem\u003eJ. Hydrol.\u003c/em\u003e, \u003cem\u003e595\u003c/em\u003e, p.126026. https://doi.org/10.1016/j.jhydrol.2021.126026\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eMayer, L., Jakobsson, M., Allen, G., Dorschel, B., Falconer, R., Ferrini, V., Lamarche, G., Snaith, H. and Weatherall, P., 2018. The Nippon Foundation\u0026mdash;GEBCO seabed 2030 project: The quest to see the world\u0026rsquo;s oceans completely mapped by 2030. \u003cem\u003eGeosciences\u003c/em\u003e, \u003cem\u003e8\u003c/em\u003e(2), p.63. https://doi.org/10.3390/geosciences8020063\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eMcKenzie, L.J., Nordlund, L.M., Jones, B.L., Cullen-Unsworth, L.C., Roelfsema, C. and Unsworth, R.K., 2020. The global distribution of seagrass meadows. \u003cem\u003eEnviron. Res. Lett.\u003c/em\u003e, \u003cem\u003e15\u003c/em\u003e(7), p.074041. https://doi.org/10.1088/1748-9326/ab7d06\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eMelesse, A.M., Weng, Q., Thenkabail, P.S. and Senay, G.B., 2007. Remote sensing sensors and applications in environmental resources mapping and modelling. \u003cem\u003eSens. J.\u003c/em\u003e, \u003cem\u003e7\u003c/em\u003e(12), pp.3209-3241. https://doi.org/10.3390/s7123209\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eMishra, N.B. and Crews, K.A., 2014. Mapping vegetation morphology types in a dry savanna ecosystem: Integrating hierarchical object-based image analysis with Random Forest. \u003cem\u003eInt. J. Remote Sens.,\u003c/em\u003e\u003cem\u003e35\u003c/em\u003e(3), pp.1175-1198. https://doi.org/10.3390/land4010197\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eMitchell, P.J., Downie, A.-L. \u0026amp; Diesing, M. (2018) How good is my map? A tool for semi-automated thematic mapping and spatially explicit confidence assessment. \u003cem\u003eEnviron. Model. Softw.,\u003c/em\u003e 108, 111\u0026ndash;122. https://doi.org/10.1016/j.envsoft.2018.07.014\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eMisiuk, B., Brown, C.J., Robert, K. and Lacharit\u0026eacute;, M., 2020. Harmonizing multi-source sonar backscatter datasets for seabed mapping using bulk shift approaches. \u003cem\u003eRemote Sens.\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e(4), p.601. https://doi.org/10.3390/rs12040601\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eMisiuk, B., Diesing, M., Aitken, A., Brown, C.J., Edinger, E.N. and Bell, T., 2019. A spatially explicit comparison of quantitative and categorical modelling approaches for mapping seabed sediments using random forest. \u003cem\u003eGeosci.\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e(6), p.254. https://doi.org/10.3390/geosciences9060254\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eMisiuk, B., Tan, Y.L., Li, M.Z., Trappenberg, T., Alleosfour, A., Church, I.W., Ferrini, V. and Brown, C.J., 2024. Multivariate mapping of seabed grain size parameters in the Bay of Fundy using convolutional neural networks. \u003cem\u003eMar. Geol.\u003c/em\u003e, \u003cem\u003e472\u003c/em\u003e, p.107299. https://doi.org/10.1016/j.margeo.2024.107299\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eNababan, B., Mastu, L.O.K., Idris, N.H. and Panjaitan, J.P., 2021. Shallow-water benthic habitat mapping using drone with object based image analyses. \u003cem\u003eRemote sens.\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e(21), p.4452. https://doi.org/10.3390/rs13214452\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eNational Marine Science Committee, 2015. \u003cem\u003eNational marine science plan 2015-2025: driving the development of Australia\u0026apos;s blue economy\u003c/em\u003e. National Marine Science Committee.\u003c/li\u003e\n \u003cli\u003eNational Research Council, Board on Environmental Change and Committee on the Human Dimensions of Global Change, 1998. \u003cem\u003ePeople and pixels: Linking remote sensing and social science\u003c/em\u003e. National Academies Press.\u003c/li\u003e\n \u003cli\u003eNikolaou, A. and Katsanevakis, S., 2023. Marine extinctions and their drivers. \u003cem\u003eReg. Environ. Change\u003c/em\u003e, \u003cem\u003e23\u003c/em\u003e(3), p.88. https://doi.org/10.1007/s10113-023-02081-8\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eNowosad J, Stepinski T (2022). \u0026ldquo;Extended SLIC superpixels algorithm for applications to non-imagery geospatial rasters.\u0026rdquo; \u003cem\u003eInt. J. Appl. Earth Obs. Geoinf.\u003c/em\u003e, pp.112. https://doi.org/10.1016/j.jag.2022.102935\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eNursamsi, I., Phinn, S.R., Levin, N., Luskin, M.S. and Sonter, L.J., 2024. Remote sensing of artisanal and small-scale mining: A review of scalable mapping approaches. \u003cem\u003eSci. Total Environ\u003c/em\u003e., p.175761. https://doi.org/10.1016/j.scitotenv.2024.175761\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eOthman, A.N., Hashim, N., Salim, P.M. and Zaidi, P.N.S.M., 2023. Comparative Study of Pixel-Based and Object-Based Classifications in Benthic Mapping. \u003cem\u003eJ. Adv. Geosp. Sci. Technol. 3\u003c/em\u003e(2), pp.51-62. https://doi.org/10.11113/jagst.v3n2.69\u0026nbsp;\u003c/li\u003e\n \u003cli\u003ePerkins, N.R., Foster, S.D., Hill, N.A. and Barrett, N.S., 2016. Image subsampling and point scoring approaches for large-scale marine benthic monitoring programs. \u003cem\u003eEstuar. Coast. Shelf Sci.\u003c/em\u003e, \u003cem\u003e176\u003c/em\u003e, pp.36-46. https://doi.org/10.1016/j.ecss.2023.108438\u0026nbsp;\u003c/li\u003e\n \u003cli\u003ePorskamp, P., Rattray, A., Young, M. and Ierodiaconou, D., 2018. Multiscale and hierarchical classification for benthic habitat mapping. \u003cem\u003eGeosci.\u003c/em\u003e, \u003cem\u003e8\u003c/em\u003e(4), p.119. https://doi.org/10.3390/geosciences8040119\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eProctor, R., Langlois, T., Friedman, A., Mancini, S., Hoenner, X. and Davey, B., 2018. Cloud-based national on-line services to annotate and analyse underwater imagery.\u003c/li\u003e\n \u003cli\u003eRam\u0026iacute;rez, F., Af\u0026aacute;n, I., Davis, L.S. and Chiaradia, A., 2017. Climate impacts on global hot spots of marine biodiversity. \u003cem\u003eSci. Adv.\u003c/em\u003e, \u003cem\u003e3\u003c/em\u003e(2), p.e1601198. https://doi.org/10.1126/sciadv.1601198\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eRattray, A., Ierodiaconou, D., Monk, J., Laurenson, L.J.B. and Kennedy, P., 2014. Quantification of spatial and thematic uncertainty in the application of underwater video for benthic habitat mapping. \u003cem\u003eMar. Geod.\u003c/em\u003e, 37(3), pp.315-336.\u003c/li\u003e\n \u003cli\u003eReid, P.C., Fischer, A.C., Lewis-Brown, E., Meredith, M.P., Sparrow, M., Andersson, A.J., Antia, A., Bates, N.R., Bathmann, U., Beaugrand, G. and Brix, H., 2009. Impacts of the oceans on climate change. \u003cem\u003eAdv. Mar. Biol\u003c/em\u003e, \u003cem\u003e56\u003c/em\u003e, pp.1-150. https://doi.org/10.1016/S0065-2881(09)56001-4\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eRigatti, S.J., 2017. Random forest. \u003cem\u003eJ Insur. Med\u003c/em\u003e, \u003cem\u003e47\u003c/em\u003e(1), pp.31-39. https://doi.org/10.17849/insm-47-01-31-39.1\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eRubbens, P., Brodie, S., Cordier, T., Destro Barcellos, D., Devos, P., Fernandes-Salvador, J.A., Fincham, J.I., Gomes, A., Handegard, N.O., Howell, K. and Jamet, C., 2023. Machine learning in marine ecology: an overview of techniques and applications. \u003cem\u003eICES J. Mar. Sci.,\u003c/em\u003e 80(7), pp.1829-1853. https://doi.org/10.1093/icesjms/fsad100\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eRubidge, E.M., Gale, K.S. and Curtis, J.M., 2016. Community ecological modelling as an alternative to physiographic classifications for marine conservation planning. \u003cem\u003eBiodivers. Conserv.\u003c/em\u003e, \u003cem\u003e25\u003c/em\u003e, pp.1899-1920. https://doi.org/10.1016/j.ecolind.2021.107849\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eSanders, K.T. and Masri, S.F., 2016. The energy-water agriculture nexus: the past, present and future of holistic resource management via remote sensing technologies. \u003cem\u003eJ. Clean. Prod.\u003c/em\u003e, 117, pp.73-88. https://doi.org/10.1016/j.heliyon.2023.e17016\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eSecretary-General, U.N., 2017. Progress towards the Sustainable Development Goals: report of the Secretary-General.\u003c/li\u003e\n \u003cli\u003eShapiro, S. S., and Wilk, M. B. (1965). An Analysis of Variance Test for Normality (Complete Samples). \u003cem\u003eBiometrika\u003c/em\u003e, 52(3/4), 591\u0026ndash;611. https://doi.org/10.2307/2333709\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eShields, J., Pizarro, O. and Williams, S.B., 2021. Feature space exploration for planning initial benthic AUV surveys. \u003cem\u003earXiv preprint at\u0026nbsp;\u003c/em\u003ehttps://doi.org/10.48550/arXiv.2105.11598\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eSibaruddin, H.I., Shafri, H.Z.M., Pradhan, B. and Haron, N.A., 2018, June. Comparison of pixel-based and object-based image classification techniques in extracting information from UAV imagery data. \u003cem\u003eIOP Conf. Ser.: Earth Environ. Sci.\u0026nbsp;\u003c/em\u003e(Vol. 169, No. 1, p. 012098). IOP Publishing. https://doi.org/10.1088/1755-1315/169/1/012098\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eSpalding, M.D., Ruffo, S., Lacambra, C., Meliane, I., Hale, L.Z., Shepard, C.C. and Beck, M.W., 2014. The role of ecosystems in coastal protection: Adapting to climate change and coastal hazards. \u003cem\u003eOcean Coast. Manag.,\u003c/em\u003e\u003cem\u003e90\u003c/em\u003e, pp.50-57. https://doi.org/10.1016/j.ocecoaman.2013.09.007 \u0026nbsp;\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eStrong, J.A. and Elliott, M., 2017. The value of remote sensing techniques in supporting effective extrapolation across multiple marine spatial scales. \u003cem\u003eMar. Pollut. Bull.\u003c/em\u003e, \u003cem\u003e116\u003c/em\u003e(1-2), pp.405-419. https://doi.org/10.1016/j.marpolbul.2017.01.028\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eTurner, W., Spector, S., Gardiner, N., Fladeland, M., Sterling, E. and Steininger, M., 2003. Remote sensing for biodiversity science and conservation. \u003cem\u003erends Ecol. Evol., 18\u003c/em\u003e(6), pp.306-314. https://doi.org/10.1016/S0169-5347(03)00070-3\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eVictoria, P., 2003. Victoria\u0026rsquo;s System of Marine National Parks and Marine Sanctuaries. \u003cem\u003eManagement Strategy 2003\u0026ndash;2010\u003c/em\u003e.\u003c/li\u003e\n \u003cli\u003eVon Schuckmann, K., Holland, E., Haugan, P. and Thomson, P., 2020. Ocean science, data, and services for the UN 2030 Sustainable Development Goals. \u003cem\u003eMar. Policy\u003c/em\u003e, \u003cem\u003e121\u003c/em\u003e, p.104154. https://doi.org/10.1016/j.marpol.2020.104154\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eWager, S., Hastie, T. \u0026amp; Efron, B. (2014) Confidence intervals for random forests: the jackknife and the infinitesimal jackknife. J. Mach. Learn. Res., 15(1), 1625\u0026ndash;1651.\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eWahidin, N., Siregar, V.P., Nababan, B., Jaya, I. and Wouthuyzen, S., 2015. Object-based image analysis for coral reef benthic habitat mapping with several classification algorithms. \u003cem\u003eProcedia Environ. Sci.\u003c/em\u003e, \u003cem\u003e24\u003c/em\u003e, pp.222-227. https://doi.org/10.1016/j.proenv.2015.03.029\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eWallace, L., Saldias, D.S., Reinke, K., Hillman, S., Hally, B. and Jones, S., 2019. Using orthoimages generated from oblique terrestrial photography to estimate and monitor vegetation cover. \u003cem\u003eEcol. Indic.\u003c/em\u003e, \u003cem\u003e101\u003c/em\u003e, pp.91-101. https://doi.org/10.1080/15481603.2020.1859264\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eWard, D., Melbourne-Thomas, J., Pecl, G.T., Evans, K., Green, M., McCormack, P.C., Novaglio, C., Trebilco, R., Bax, N., Brasier, M.J. and Cavan, E.L., 2022. Safeguarding marine life: conservation of biodiversity and ecosystems. \u003cem\u003eRev. Fish Biol. Fish..\u003c/em\u003e, \u003cem\u003e32\u003c/em\u003e(1), pp.65-100. https://doi.org/10.1007/s11160-022-09700-3\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eWass, R.E., Conolly, J.R. and MacIntyre, R.J., 1970. Bryozoan carbonate sand continuous along southern Australia. \u003cem\u003eMar. Geol.\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e(1), pp.63-73. https://doi.org/10.1016/0025-3227(70)90080-0\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eWernberg, T., Krumhansl, K., Filbee-Dexter, K. and Pedersen, M.F., 2019. Status and trends for the world\u0026rsquo;s kelp forests. In \u003cem\u003eWorld seas: An environmental evaluation\u003c/em\u003e (pp. 57-78). Academic Press. https://doi.org/10.1016/B978-0-12-805052-1.00003-6\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eWescott, G., 2006. The long and winding road: the development of a comprehensive, adequate and representative system of highly protected marine protected areas in Victoria, Australia. \u003cem\u003eOcean Coast. Manag., 49\u003c/em\u003e(12), pp.905-922. https://doi.org/10.1016/j.ocecoaman.2006.08.001\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eWhiteside, T.G., Boggs, G.S. and Maier, S.W., 2011. Comparing object-based and pixel-based classifications for mapping savannas. \u003cem\u003eJ. Appl. Earth Obs. Geoinformation.\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e(6), pp.884-893. https://doi.org/10.1016/j.jag.2011.06.008\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eZheng, J.Y., Hao, Y.Y., Wang, Y.C., Zhou, S.Q., Wu, W.B., Yuan, Q., Gao, Y., Guo, H.Q., Cai, X.X. and Zhao, B., 2022. Coastal wetland vegetation classification using pixel-based, object-based and deep learning methods based on RGB-UAV. \u003cem\u003eLand\u003c/em\u003e, \u003cem\u003e11\u003c/em\u003e(11), p.2039. https://doi.org/10.3390/land11112039\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eZone, S. and Zone, H.P., 2013. South-east Commonwealth Marine Reserves Network.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Benthic Habitat Classification, Remote Sensing, Hierarchical Random Forest, Backscatter Segmentation, GIS, Marine Spatial Planning","lastPublishedDoi":"10.21203/rs.3.rs-5351238/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5351238/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eContext\u003c/h2\u003e \u003cp\u003eBenthic habitat mapping is crucial for effective marine spatial planning. Despite advancements in multibeam echosounder (MBES) technology, selecting appropriate classification methods to accurately map seafloor habitats remains a challenge.\u003c/p\u003e\u003ch2\u003eObjectives\u003c/h2\u003e \u003cp\u003eThis study aims to provide novel comparisons of large spatial scale habitat classifications using pixel-based (PB) and object-based image analysis (OBIA) methods, applied within a hierarchical random forest framework, to classify benthic biotopes in the northern section of Apollo Marine Park and the adjacent Cape Otway coastline, Victoria, Australia.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eWe utilised high-resolution MBES-derived data, implementing a hierarchical random forests algorithm to classify benthic habitats. The PB method treated each pixel independently, allowing for high spatial detail, while the OBIA method grouped pixels into meaningful segments for classification. Prior to segmentation, backscatter data from two different MBES systems were harmonised using a bulk shift method (Misiuk et al., \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) to ensure consistency across datasets. We then applied the Supercells segmentation technique (Nowosad \u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) to the harmonised backscatter data, forming the foundation for the OBIA-based classification. Both methods were evaluated using accuracy, F1 scores, and uncertainty maps were generated to assess classification reliability.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eBoth classification methods demonstrated strong performance, with no statistically significant differences in overall accuracy. However, the complexity of the habitat maps varied: the PB approach excelled in capturing fine-scale habitat details, beneficial for management and conservation efforts requiring high detail. Conversely, the OBIA method produced more interpretable and less complex maps, suitable for general spatial analyses, though it resulted in the omission of some minority classes.\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eThis study emphasises the importance of defining the desired level of complexity in habitat maps before analysis, ensuring that chosen methods yield maps suitable for specific applications\u0026mdash;particularly in datasets with strong class imbalances. Future advancements in machine learning and emerging technologies have the potential to further refine habitat mapping techniques and enhance classification accuracy.\u003c/p\u003e","manuscriptTitle":"Comparing Pixel-and Object-Based Approaches for Classifying Benthic Habitats","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-11-20 10:36:13","doi":"10.21203/rs.3.rs-5351238/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"7a680542-ac52-468a-91d9-c888111ad9a3","owner":[],"postedDate":"November 20th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-04-11T06:38:32+00:00","versionOfRecord":[],"versionCreatedAt":"2024-11-20 10:36:13","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-5351238","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5351238","identity":"rs-5351238","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}