Distribution models of the branched hexactinellid Sarostegia oculata in Rio Grande Rise (SW Atlantic)

Distribution models of the branched hexactinellid Sarostegia oculata in Rio Grande Rise (SW Atlantic) | 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 Article Distribution models of the branched hexactinellid Sarostegia oculata in Rio Grande Rise (SW Atlantic) Paulo Corrêa, Paulo Sumida This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-1663413/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted You are reading this latest preprint version Abstract The mining of ferromanganese (Fe-Mn) crusts in the deep sea have gained more attention in the last decade due to increased demand for rare earth elements that are critical for low-carbon technologies, which makes exploitation of this resource feasible and profitable. The Rio Grande Rise (RGR) is a distinct feature located in the South Atlantic and has become a region of great commercial and scientific interest because of its potential for mining Fe-Mn crusts. Extraction of Fe-Mn deposits may cause irreversible changes by removal of substrate, creation of sediment plumes, among others. Here, we use species distribution models (SDMs) to predict the occurrence of Sarostegia oculata , a branched hexactinellid that mimics the 3D skeletal framework of actual corals. It is the dominant organism in areas rich in Fe-Mn crusts and has relevant ecological importance in RGR. The models had excellent or good performance statistics and a high discrimination power between presence and absence sites. Our results support the relationship between S. oculata , the Fe-Mn crusts, and the rift at RGR, and may help to create a management plan and preserve unique marine habitats and biodiversity from the deep sea. Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Introduction Deep-sea sponges have importance as organisms that can provide multiple ecosystem goods and services, and thus have increasingly been mentioned in the literature in recent years 1 . Habitats dominated by sponges are globally distributed, forming structurally complex and often highly diverse communities, along with corals 2 , 3 . They have key roles in ecosystem functions, including habitat provision for associated fauna 4 , 5 , increasing biodiversity 6 – 9 , promoting silicon and organic carbon cycling 10 – 12 , and potentially acting as nursery habitat for many species 13 . Vulnerable Marine Ecosystems (VMEs) are characterized as areas with low resilience and slow recovery from anthropogenic disturbances such as bottom trawling and mining (extraction of fossil fuels, gas, or minerals) 14 . Some taxonomic groups, including sponges, are considered indicators of VMEs as a consequence of having slow growth rates, longevity, late maturity, fragility, as they form three-dimensional structures associated with diverse communities 15 . They have been used to assist agencies responsible for the protection of particular ocean regions, such as slopes, seamounts, and canyons 16 . Increasing knowledge of the distribution and biology of species is necessary to create a consistent network of marine protected areas (MPAs), which can diminish human impacts on the ocean and prevent loss of biodiversity. The Rio Grande Rise (RGR) is a major positive feature in the South Atlantic, between the parallels 28°–35° South and meridians 28°–39° West, located about 1200 km east off the Brazilian coast, and forming a volcanic plateau that rises from depths of about 5000 m to above 1000 m in some areas 17 . The western region of RGR (WRGR) forms an elliptical bulge of 500 km × 300 km that includes seamounts and guyots towering over or juxtaposing with this massive platform 18 ; it is intersected by a NW-SE trending submarine rift between 20–40 km wide (also referred to as ‘graben’ or Cruzeiro do Sul Lineament). RGR has gained special attention from researchers and the Brazilian government in recent years 19 , 20 due to prospecting surveys that revealed its potential for the mining of Fe-Mn crusts 18 , 21 , 22 . Impacts created by mining activities may range from direct destruction of habitats and formation of toxic and particle-rich sediment plumes 23 – 25 to noise, vibration, and light, which may lead to significant (potentially irreversible) damage to deep-sea communities and ecosystems and loss of biodiversity on mining sites and surrounding areas 19,26−29 . Sarostegia oculata is a branched hexactinellid that forms an erect, complex surface, similar to the shape of a stony coral. It has a close association with a zoanthid, likely Thoractis topsenti 30 , growing on and within the skeleton in a similar way to coral polyps, thus mimicking the 3D skeletal framework of actual corals 31 . It was first described at Cape Verde by Topsent 32 , and further sampled throughout the Atlantic 33 , 34 and Indian 35 oceans. More recently, Hajdu et al. 31 recorded extensive sponge gardens in RGR, being the dominant organism in areas rich in Fe-Mn crusts that play a fundamental role in supporting communities in the area. Species distribution models (SDMs; also known as habitat suitability models) are a method to predict the likelihood of a species to occur in a given location, based on their observed relationship with environmental conditions. SDM is widely used in ecology and conservation, especially in the deep sea where data is scarce, and such models are useful for the development of conservation measures and marine ecosystem management by predicting species distributions beyond surveyed areas 36 – 38 . In the present study, we built such models to predict the distribution of S. oculata in two areas with available high-resolution multibeam data at RGR, surrounding the rift within the bulge of WRGR. The growing use of mineral resources in our global economy will surely lead to future commercial extraction of mineral resources in the deep sea, including, but not limited to, Fe-Mn crusts 39 – 41 . Such models can be used to inform spatial management planning for protecting VMEs and environmental impact assessments that minimize the effects of mining on deep-sea ecosystems 42 , 43 . Our study is the first to create a high-resolution species distribution model of a VME indicator species in Rio Grande Rise. Sarostegia oculata is a dominant organism on substrates rich in Fe-Mn crusts, which play a fundamental role to support communities and provide key ecosystem functions in RGR. The five models built here predicted a high likelihood of S. oculata along with the rift borders with a low degree of uncertainty, in an area aimed at exploration of Fe-Mn crust deposits. This reinforces the relationship between S. oculata , the Fe-Mn crusts, and the rift at RGR, implicating that direct destruction of habitats caused by the mining of Fe-Mn crusts could lead to a severe impact in the communities sustained by S. oculata , which should be addressed in the management and MPA network design for RGR. There are still large gaps in the knowledge of the benthic fauna distribution and ecosystem functions on RGR, but our study provides expert advice on the occurrence of deep-sea sponges in the area and represents initial steps to guide future surveys and conservation plans in the region. Methods Data collection Two surveyed areas near the center of WRGR were used in this study. The first area (here named Area 1) was delimited between latitudes 30º35’S–31°03’S and longitudes 35º36’W–36°16’W (Fig. 1 a). This area is ~ 2,000 km², ranging from 600 m to 1,800 m depth, and encompasses a small segment of the rift and a plateau on both sides, here named NE and SW plateaus (Fig. 1 b). It also contains a terrace in the NE plateau, which runs almost parallel to the rift border, a slide south of the SW plateau, and a canyon that splits a section of the SW plateau in two, one next to the rift border and another named the inner SW plateau. The south and west of the SW plateau are neighbored by a lower plane, which extends beyond the MBES data. The second area (Area 2) is smaller, located about 30 km northwest of Area 1, delimited between latitudes 30°19’S–30°25’S and longitudes 35°57’W–36°06’W, is ~ 77 km², ranging from 600 to 1400 m depth. It encompasses a small area of the northeastern plateau and the rift (Fig. 1 c). Area 1 (Figs. 1 b and 2 ) was surveyed with the N/Oc. Alpha Crucis , from January 30 to February 21, 2018, and RSS Discovery , from October 26 to November 8, 2018, as part of the FAPESP/RCUK funded project Marine E-Tech. N/Oc. Alpha Crucis used the Reason 7160 multibeam echosounder (MBES) operating at 41 kHz. RSS Discovery was equipped with a ship-mounted Kongsberg EM122, and multibeam was processed using Caris HIPS and SIPS v9.1.8. The data in both products were gridded at 25 m using NaviModel Producer 4.3.2, then fused with ArcGIS Pro 2.9, using the Mosaic tool with the “Blend” mosaic operator. The survey in the RSS Discovery included 13 dives (codes HY31–43) of the Robotic Underwater Vehicle, RUV HyBIS (Hydraulic Benthic Interactive Sampler). HyBIS was equipped with a Sony Full HD camera that recorded over 36 hours and 26 km of video footage of the seafloor and it was connected to the ship's USBL system to record its position (latitude, longitude, and depth). Area 2 (Figs. 1 c and 3 ) was surveyed with the submersible Shinkai 6500 and Yokosuka support vessel, from April 13 to May 5, 2013, as part of the ‘Iata-Piuna’ expedition. Yokosuka was equipped with multibeam Sea Beam 2112.004, operating at 12 kHz. The data was also gridded at 25 m using NaviModel Producer. Shinkai 6500 was equipped with two HD video cameras. One of them was fixed at the bow, which recorded 4.5 hours and 2,282 m of footage from the seabed during the dive 6K1338. The other video camera was mobile and was ignored in this study so the analysis would be more consistent with the RUV HyBIS . The HyBIS and Shinkai videos were annotated and counted for benthic fauna while classifying them to the lowest possible taxonomic group. The branched hexactinellid Sarostegia oculata was the most commonly observed habitat-forming VME indicator species in both Area 1 and a 2. The timestamp of the video was used to obtain the coordinates of the observations from the HyBIS and Shinkai track data. Grid cells where at least one individual of S. oculata were classified as “presence”, and grid cells with overlapping track and no observations of S. oculata were classified as “absence”. Each of these grid cells are hereafter called records or sites, used to train and evaluate the models. A set of additional seafloor variables (Figs. 2 and 3 ) was derived from the bathymetric data using the Benthic Terrain Model 3.0 tool 44 in ArcGIS. These variables were slope, aspect (measured in terms of northness and eastness), roughness (11 neighborhood size), curvature, fine-scale (3/30 radius) Bathymetric Position Index (BPI), and broad-scale (20/200 radius) BPI. The rift is a major feature in both areas that may play an important role in the distribution of the species in RGR. As such, we included the distance from the rift border as another environment variable in our analysis. This was achieved by creating a contour at value 0 from the broad-scale BPI raster. This will result in two contours outlining the NE and SW rift borders, which were used as source for the “Euclidean Distance” tool in ArcGIS. Before data analysis, collinearity between variables was checked using Pearson’s correlation coefficient and the Variance Inflation Factors (VIFs) 45 using the usdm package 46 . Modeling methods The distribution of S. oculata was modeled using five algorithms widely used in species distribution models: Random Forest (RF), Boosted Regression Trees (BRT), MaxEnt, Generalized Additive Models (GAMs), and Artificial Neural Networks (ANN). In general, these models differ in the way they determine the fitted function, and how they handle interactions, model complexity, and overfitting 47 . Models were developed using R 48 with the framework caret 49 . The caret package is a set of functions that streamline the process of creating predictive models and provides a uniform interface for data splitting, variable selection, and model tuning. It uses other packages internally to create the models. We used default parameters for each package, except when stated in the text below or Table 1 . The function rfe was used to perform recursive feature elimination in the models, but no advantage was gained in model performance from removing variables from the models, so we kept all variables in the final models. RF builds multiple classification trees by taking many bootstrap samples from the training data and making an average prediction over all fitted trees 50 . RF uses only a randomly smaller set of the predictor variables (parameter mtry ) on each split while growing each tree, which reduces the variance of the final model, with subsequent gains for predictive performance. We set the number of trees to grow (parameter ntree ) as 1000. BRT builds several regression trees (parameter n.trees ) in a forward stagewise fashion. At each step of model fitting, the algorithm fits each new tree to the residuals of the previously fitted trees and uses only a random subset of data 51 . This gradually increases emphasis on observations modeled poorly by the existing collection of trees. The contribution of each tree is controlled by the shrinkage parameter, as the model-building process usually performs best if it moves slowly down the gradient. Interaction.depth controls the maximum depth of the individual trees and n.minobsinnode controls the minimum number of observations in terminal nodes. Maxent estimates species distribution by minimizing the relative entropy between two probability densities 52 , 53 . Although it was initially developed as a machine-learning algorithm, MaxEnt has known links to Poisson point process models 54 and it can be categorized as a “regression-based” model. MaxEnt has the ability to fit complex models, controlled by the use of transformed features of the predictor covariates (parameter feature types ). Regularization multiplier can also be used to smooth and avoid fitting overly complex models. In this work, we employed real absences derived from the dives instead of more commonly used pseudo-absences and used the Java implementation to train the MaxEnt models, version 3.4.4. GAMs are an extension of General Linear Models (GLMs) but use nonparametric smoothing functions which allow modeling of non-linear relationships between the response and explanatory functions. Multiple smoothing parameter estimation methods are available in the package (parameter method ) 55 . It can also add an extra penalty to each term (parameter select ), meaning that the smoothing parameter estimation can completely remove terms from the model. We used a binomial distribution with the logit function and thin plate regression splines for the smoothing basis of all terms. ANNs are a deep learning approach that consist of a large number of nodes and connections, typically organized in layers 56 . The input layer contains the environmental data, with each input node representing one environmental variable. The information from each input node is fed into the hidden layers, which may contain multiple layers with a variable number of nodes. Data from the last hidden layer is then fed to the output layer which represents the result of the model. The connections between the nodes from one layer to the next, also known as weights, are optimized using a back-propagation process. Our model has a single hidden layer, with 1 to 21 nodes (parameter size ). Decay is used as a penalty for these connections that both help the optimization process and avoid over-fitting. The feed-forward and back-propagation processes are repeated until the model reaches a pre-defined accuracy, or a maximum set number of runs controlled by the parameter maxit (in our model we set maxit as 500). Table 1 Modeling methods, their implementations during tuning, and the tuned parameters used in the final models. Method R package Parameter Values Tuned RF randomForest mtry 1–9 1 BRT gbm n.trees from 100 to 3,000 by 50 2600 interaction.depth from 1 to 16 by 3 10 shrinkage 10 − 1 , 10 − 2 , 10 − 3 , 10 − 4 10 − 3 n.minobsinnode 5, 15 5 Maxent custom¹ feature types l, lq, lqp, lqph, lqh, lqpt, lqt, lqpth² lqpth regularization multiplier from 0.2 to 2 by 0.2, 2.5, 3, 3.5, 4, 4.5, 5 1 GAM mgcv method REML, ML ML select TRUE, FALSE TRUE ANN nnet size 1, 3, 5, 7, 9, 11, 13, 21 21 decay 0, 10 − 1 , 10 − 2 , 10 − 3 , 10 − 4 , 10 − 5 0.1 ¹ Custom script was used to run the MaxEnt java application and read its output within caret. ² l = linear, q = quadratic, h = hinge, p = product and t = threshold. The relationship between the environmental layers and the predicted probability of presence was analyzed using response curves produced using the methodology described by Elith et al. 57 . All variables were held constant at their mean except the target variable, which varied at 100 points across its range. Then, the prediction of each algorithm was computed for each of the 100 values of the target variable and used to produce the response curves. The advantage of this method is that it can be applied to any model and be compared against different model techniques, but it does not account for interactions between variables. The importance of each variable in the final output was computed using the methodology described by Thuiller et al. 58 . The target variable was shuffled, while the others were left untouched. The Pearson's correlation coefficient between the reference and shuffled predictions was calculated and the score 1 - correlation is returned. Scores near one mean the variable has a high influence on the model, and a score of zero assumes no influence of the variable on the model. The process was repeated 100 times for each variable. An ensemble model is a popular technique for reducing the uncertainty of model predictions 59 . They may reduce prediction errors and decrease mean bias due to the choice of method 60 . This is usually achieved by computing the weighted or unweighted averages of the different model outputs. Ensemble models also permit uncertainty caused by the different predictions to be calculated. In this study, all model predictions were rescaled between 0 and 1, and their unweighted average was used to build the ensemble model. Spatial measures of uncertainty were calculated using a bootstrap technique 15 , 61 . A random sample of the presence-absence data, of equal size, was drawn with replacement, and the five models were constructed with the same settings as the originals. Predictions were then made to the study area, and this process was repeated 200 times. Model uncertainty was represented as the coefficient of variation (CV) of the bootstrap output, i.e, the standard deviation divided by the mean. Model evaluation The data in this study was divided into train, validation (also referred to as development in the literature), and test sets. The train set is used to train the model, while the validation set is used to evaluate the model during tuning. After the best parameters are selected, both train and validation sets are used to train a final model, which is evaluated using the test set. Data obtained from Area 1 with HyBIS was used as train and validation sets. Such data inherits a high spatial dependency, as records are available next to each other throughout each dive, and can lead to overestimation of model performance 62 . To provide a larger level of spatial independence between the training and the validation data, the first or last 20% records in each dive were selected to be used as a validation set. Consequently, the remaining 80% were used to train the model. The region of the dive which was cut from training, whether from the beginning or towards the end, was selected at random for each dive, and this process was repeated 25 times. During model tuning, the AUC ROC metric was used to select the best models. Data obtained from Area 2 with Shinkai is completely independent from Area 1 and was used exclusively as a test set for the final model. The evaluation was performed 25 times using a random selection of 70% of the test data (subsampling) in each interaction. Models were evaluated using five metrics to cover different aspects of the modeling performance: (1) area under the receiver operating characteristic curve (AUC ROC ), (2) area under the precision-recall gain curve (AUC PRG ), (3) Sensitivity, (4) Specificity, and (5) True Skill Statistics (TSS). AUC ROC and AUC PRG are threshold-independent measures, while Sensitivity, Specificity, and TSS need to convert the predicted likelihood of the model (a value between 0 and 1) into a presence/absence classification using a threshold. Here we used the threshold which balances sensitivity and specificity, which minimizes both commission and omission errors, and is recommended for SDMs model 63 . Note that thresholds were selected using the validation data set. AUC ROC assesses the ability of models to discriminate presence from absence sites and is widely used in species distribution models. The ROC curve is created by calculating the true positive rate (the proportion of presences correctly predicted, also known as sensitivity or recall) and the false positive rate (the proportion of absences falsely predicted, also calculated as 1 - specificity) at various threshold settings. AUC ROC ranges from 0 to 1, where values above 0.9 indicate excellent performance, values between 0.7 and 0.9 indicate good performance, and a value of 0.5 or below indicates no better discrimination than a random classification. The area under the precision-recall curve (AUC PR ) measures the ability to capture true presences and does not include false positives or absences in its calculation. AUC PR is recommended when the user is more interested in an accurate prediction of the presences or when the number of absences is much larger than presences 64 . Like the ROC curve, the precision-recall curve calculates precision (the proportion between correctly predicted presences and all predicted presences) and recall across multiple thresholds. The baseline in the AUC PR depends on the prevalence in the testing data 65 , making it difficult to compare between species or studies. As such, the precision-recall curve is plotted in a new coordinate system, called the precision-recall gain curve, as described by Flach and Kull 66 . AUC PRG values from 0 to 1 indicate a performance better than random, where 1 is a perfect discrimination model. Negative values indicate predictions worse than random. Specificity is the proportion of absences correctly predicted. Both sensitivity and specificity are considered high with values above 0.8 63 . TSS normalizes the overall accuracy and accounts for both omission and commission errors 67 . TSS is calculated as sensitivity plus specificity minus one and ranges from − 1 to 1, where + 1 indicates perfect agreement and values of zero or less indicate a performance no better than random. TSS is also considered less sensitive to prevalence compared to Cohen’s Kappa statistic. Because we used presence-absence data to train the models, we also produced calibration plots to evaluate the agreement between predicted probabilities of occurrence and observation of presence and absence. If the plotted tend is lined up close to the diagonal, we conclude that the model is well-calibrated. Calibration plots were drawn using the scripts in Phillips and Elith 68 . Statistical significance of differences between the models for the five statistics were tested using Friedman’s Aligned Rank test ( p < 0.05) 69 , using the R package scmamp v0.2.55 70 . We also applied the post hoc test to conduct pairwise comparisons for each metric, using the Shaffer correction. Results From the 1193 records (grid cells) that were observed by HyBIS in Area 1, 231 had the presence of at least one Sarostegia oculata . In Area 2, 29 out of 255 records were observed with the presence of S. oculata by Shinkai . The hexactinellid showed a higher abundance near the rift walls in both areas and was found only on hard substrates, of which the majority had a ferromanganese crust. The performance of the models was good for each metric calculated, for both validation and test data. AUC ROC (Fig. 4 a) values were > 0.9, except for ANN which performed slightly worse. All models had good AUC PRG scores > 0.8 (Fig. 4 b) and the RF, BRT, MaxEnt, GAM, and Ensemble models had nearly perfect values close to one for the test data, indicating a good discrimination ability to detect presence records. Similarly, all models showed high values for Sensitivity (Fig. 4 c) and Specificity (Fig. 4 d) for the validation data, suggesting a high proportion of observed presences and absences correctly predicted, respectively. Models RF, BRT, MaxEnt, and Ensemble had a higher sensitivity for the test data compared to the validation data. In contrast, GAM and ANN models had worse performance. Models had high specificity for test data as well, although the RF model was slightly worse. For the TSS metric, the ANN model had the overall worst performance, while models BRT, MaxEnt, and Ensemble had the highest values for test data (Fig. 4 e). Friedman’s Aligned Rank test showed a significant difference between models for the five statistics. The post-hoc test showed that the most significant differences were found when comparing GAM and ANN with other models (Appendix D). The threshold values used to discriminate between presence and absence were 0.248, 0.126, 0.304, 0.121, 0.187, and 0.225 for the RF, BRT, MaxEnt, ANN, GAM, and Ensemble models, respectively. The calibration plots show that the true probability of presence compared to the predicted presence of the five models are badly calibrated (Fig. 5 ). The ideal curve (dotted line) is below the lower confidence interval of the fitted calibration curve, indicating that the true probability of presence is much larger than the estimate given by the models. Only models RF (Fig. 5 a) and MaxEnt (Fig. 5 c) had the ideal curves above the higher confidence interval for low probability values. These results indicate models have a high discrimination power, i.e. the ability of a model to correctly distinguish between occupied and unoccupied sites, but the model output should not be interpreted as estimates of conditional probability of presence. Generally, all models predicted a suitable habitat along with the full extent of both NE and SW rift borders in Area 1, and the NE rift border in Area 2. For Area 1 (Fig. 6 ), RF model predicted the distribution to be more extensive, with a relatively higher likelihood on the bottom of the rift compared to the middle of the plateaus. Other models had a relatively low (< 0.1) likelihood on the bottom of the rift, on top of the plateaus, on the south canyon, and the lower plane regions at southwest. All models predicted a larger area with a high likelihood around the north end of the NE rift border. However, only models RF and ANN extended this area throughout the small terrace at the NE plateau. Models predicted high suitability nearby the area between the east side of the canyon and the SW plateau. The predicted likelihood extended to the other side of the canyon for the MaxEnt model, and even further around the inner SW plateau for models RF and BRT. The south slide of the SW plateau showed high predicted suitability as well, except for the GAM model. For Area 2 (Fig. 7 ), models predicted a high likelihood near the slope of the NE rift border, between 700 and 1000 m. All models, except for RF, showed low (< 0.1) suitability at the rift bottom and the top to the plateau. The predicted suitable habitat by the ensemble model reflected the average of all five models accordingly. This predicted distribution reflected environmental variables included in the model, namely depth and slope. The region with bottom depths between 700 and 1000 m that had nearby slopes > 20 degrees contained a continuous band of high prediction of suitable habitat. The spatial patterns of low-modeled uncertainty corresponded to the main areas predicted as highly suitable on the rift borders in both Area 1 and Area 2 (Fig. 8 ), together with the majority of the SW plateau and around the inner SW plateau. Regions of high uncertainty were obtained for the rift bottom, the canyon, the lower plane regions, and in some areas on top of the plateaus. The importance of the environmental variables varied across the modeling algorithms, but depth, fine BPI, and northness usually had a high influence across all models (Table 2 ). For the RF model, the variables showed a low importance index (< 0.1), indicating that this model uses all variables to predict the presence of S. oculata , and changing a single variable has little effect on its output. Only depth and fine BPI had higher importance relative to other variables. For BRT, MaxEnt, and GAM models, the variables depth, fine BPI, northness, and curvature had a high influence in the likelihood of S. oculata . However, depth had a higher influence for BRT and GAM compared to MaxEnt, and northness had a higher influence for MaxEnt and GAM compared to BRT models. For the ANN model, most variables showed a high influence in its output, except for broad BPI and eastness. Table 2 Mean index of the importance of each predictor variable across 100 permutations in the training dataset, for the Random Forest (RF), Boosted Regression Trees (BRT), MaxEnt, Generalized Additive Models (GAM), and Artificial Neural Networks (ANN) models. Variables RF BRT MaxEnt GAM ANN depth 0.09 0.388 0.113 0.572 0.553 slope 0.035 0.009 0.034 0.051 0.183 broad BPI 0.016 0.002 0.000 0.000 0.066 fine BPI 0.072 0.296 0.441 0.588 0.44 rift distance 0.049 0.008 0.004 0.000 0.349 northness 0.044 0.112 0.435 0.762 0.338 eastness 0.031 0.003 0.007 0.000 0.081 rugosity 0.036 0.011 0.008 0.000 0.19 curvature 0.037 0.046 0.106 0.094 0.304 In general, models showed similar response patterns across the gradient of each environmental variable. Depth (Fig. 9 a): MaxEnt, GAM, and ANN models showed a low response at higher depths below ~ 1000 m and shallower waters above ~ 700 m, but showed a peak in the predicted likelihood of S. oculata within ~ 700–1000 m. For RF and BRT models, the response remained high from deeper sites until ~ 800 m where it reached a peak, and then the response was low at shallow depths. Slope (Fig. 9 b): models had a low response at flat sites, which increased at steeper slopes. The biggest variation in the response for slope was found in the ANN model. Broad BPI (Fig. 9 c): it had a high response around 70 for RF, BRT, and MaxEnt models, but GAM and ANN predicted likelihood was higher at lower broad BPI values ( 100. Fine BPI (Fig. 9 d): RF, BRT, and ANN showed a lower response at fine BPI 0, with a larger variation produced by the ANN model. MaxEnt response was higher with negative fine BPI, while GAM was unresponsive for this variable, characterized by a flat horizontal line in the plot. Rift distance (Fig. 9 e): All models except ANN showed a similar pattern for the rift distance. A peak in response near the rift (< 2000 m) and from 4500 to 7000 m, along with low response between 2000 and 4500 m and regions more distant than 7000 m. The ANN model was different, with a high response near the rift, lowering constantly as it moves away until 10,000 m. The MaxEnt, GAM, and ANN had very low predicted outputs when far away from the rift (> 20,000 m). Northness (Fig. 9 f): peaks in response were produced at sides facing north and south in models RF, BRT, and MaxEnt. GAM generated a slightly higher response in sites facing north than south. ANN, instead, generated a higher response in sites facing south than north. Eastness (Fig. 9 g): models predicted a slightly higher response on sites facing either east or west. However, the GAM was unresponsive for this variable as well. Rugosity (Fig. 9 h): models had a low response at sites with low rugosity, that increased at areas with higher values. Curvature (Fig. 9 i): only ANN model had a large variation in response for curvature, with a peak in sites with curvature close to zero. RF and BRT, instead, showed a smaller response in these sites. MaxEnt and GAM were unresponsive for this variable. Discussion The Rio Grande Rise contains diverse benthic communities which include several VME indicator species such as sponges, scleractinians, octocorals, and black corals (Corrêa et al., unpublished data). Although RGR is located in oligotrophic waters, this diversity is likely related to strong currents and complex geomorphology with multiple habitats 19 . Our study provides high-resolution distribution maps (DM) for an important VME indicator species recorded in RGR, the branched hexactinellid Sarostegia oculata . This sponge was the most abundant species found in the footage near the rift 31 , closely associated with Fe-Mn crust substrates. A few fragments of S. oculata sampled by dredges 71 revealed 18 species associated with this sponge, the most notorious being a zoanthid and an annelid, Thoracactis topsenti and Hermadion fauveli cf. Gravier 30 respectively, which also emphasize the importance S. oculata has in this environment. Our work is the first to create a species distribution model of S. oculata , and the first to produce high-resolution DMs at RGR, with potential use for deep-sea conservation and management in this area. Deep-sea sponges have a high conservation and management significance because of their low resilience 43 . These species and their associated fauna are particularly vulnerable to anthropogenic impacts such as fishing and deep-sea mining due to their slow growth rates and low or unpredictable recruitment 72 , proving to have a very slow or nonexistent recovery 5 . For being suspension filters, they will likely suffer from sediment loads caused by deep-sea mining 29 . The water canal system of such sponges are at risk of becoming clogged 73 by high loads of suspended particulate matter. These mining plumes can be of low or no nutritional value, further threatening the maintenance of these animals. Hence, S. oculata not only will be primarily affected by crust removal, but also in the area surrounding mining operations High-resolution distribution models of benthic species, using seafloor camera imagery and bathymetric data obtained from multibeam surveys, have proven useful in the last two decades. Similar models for VME indicator taxa have been developed for conservation and management of areas of interest in the deep sea 15,43,74−80 . Such models are able to provide expert advice on the occurrence of VME indicator taxa in efforts to limit anthropogenic threats from future marine resource exploitation Species distribution model The five models trained in this study predicted a high likelihood of S. oculata along with all extent of both rift borders in the multibeam survey. These findings reinforce the idea that S. oculata distribution is highly related to the NE-SW rift border that runs throughout RGR, as suggested by Hajdu et al. 31 found in our study. This area is known to be an optimal place for Fe-Mn crusts formation 19 , highly supported by the first claims by CPRM to ISA, which consists of 150 blocks of 20 km 2 each, the maximum allowed by Regulations on Prospecting and Exploration for Fe-Mn crusts 20 . The majority of the blocks are situated along the plateaus of WRGR, on both sides of the rift 19 , where CPRM likely prospected areas of high Fe-Mn deposits before submitting the claim. In addition, S. oculata is somehow more closely associated with Fe-Mn crusts pavements in the plateaus near the rift and along the rift walls compared to other substrates 31 . Our study reinforces this relationship between S. oculata , Fe-Mn crusts, and the rift at RGR. A second notable suitable area was predicted between the southwestern of the SW plateau and on the top northeastern canyon side. However, we have no footage of this area that can confirm or not the presence of S. oculata there. Sampled Fe-Mn crusts near the rift suggest that they were eroded by the strong currents that impacted the RGR plateau 21 , and bottom currents of more than 0.2 m/s may be capable of eroding Fe-Mn crust surfaces 81 . These currents may provide a high food supply for the development of S. oculata , and can be an important key factor in the development of this organism. Unfortunately, we do not have a hydrodynamic model in the study area that could provide currents velocity and direction, which could be used as predictors in the distribution models. Future studies should focus on obtaining and applying such variables in the models, as they likely play an important role in the distribution of species in RGR. The complex outlines of the rift walls may generate vortices 31 and areas of slope may facilitate the propagation of internal tidal waves 1 , 82 , which can cause resuspensions favoring greater development of sponges. This may explain why the response for rift distance and fine BPI had such importance to explain the distribution of Sarostegia . Steeper slopes also predicted higher response in the models, although this variable did not output great importance in the bootstrapped correlation test. The aspect of the seabed, in terms of northness, predicted the highest likelihood on the northern and southern slopes, which may correspond with the direction of currents in RGR as well. Another variable considered important in our results was depth. The prediction of S. oculata was restricted between ~ 700–1000 m, which corresponds to the plateaus and upper rift walls. Depth is usually an important variable that predicts the occurrence of VME indicator species and can act as a surrogate for other important variables such as temperature, salinity, oxygen, nutrients, water masses, exported surface production, and aragonite saturation 15 , 83 , 84 . S. oculata was collected/observed at 598–1311 m depth in Cape Verde 32 , at 745 m in east of Miami Terrace, south-west of Bimini 33 , at 900–790 m in the Vitória Trindade seamounts chain 34 , at 738‒1040 m in RGR by Hajdu el al. 31 , and in our study, from 681 to 1203 m. The similarity in depth ranges in the Atlantic between these records suggests that this sponge occurs in bathymetrically constrained bands. However, it is important to note that this bathymetric range alone is not sufficient to predict the distribution, as the inner SW plateau is within this range, but the models had a low likelihood in this region. There are a few global distribution models of VME indicator taxa, namely scleractinians, octocorals, and black corals, that predicted their occurrence on RGR 19,85−87 . Similar models were built using data exclusively from the Brazilian continental margin by Barbosa et al. 88 , who addressed the distribution of scleractinians and octocorals in RGR. Overall, they predicted a high suitability of VME indicator taxa on the plateau of RGR, especially in regions near the rift, somewhat similar to the predicted output of our models. However, these models use coarse resolution data ( > = 0.0083°, ~ 1 km) as environmental predictors, and none had access to biological records from RGR to train or evaluate the model. High-resolution bathymetric data can better represent seabed physiographic features and improve regional and local suitability models 89 . Implications for Spatial Management and Conservation There is a growing number of examples in the literature that demonstrate the potential of using species distribution models to predict the occurrence of deep-sea sponges for their conservation and management from impacts caused by anthropogenic activities such as bottom trawling and mining 1 , 15 . They provide fundamental ecosystem functions, and even in low densities, hexactinellid sponges may create a suitable substrate for colonization and development of several invertebrate taxa, serving as island habitats on the deep-sea floor 90 . Thus, increasing knowledge of their distribution and environmental conditions responsible for their formation and persistence are key factors to ensure the protection of the marine environment, especially from harmful effects resulting from human activities. The models performance were excellent or good in most cases, correctly identifying 85.4% of sites for validation and 88.8% for testing. Our study uses a completely different dataset to test the models (Area 2), independent from the data used to train and validate the model (Area 1). This method was intended to simulate how models would perform in case they were used to predict the distribution of S. oculata in a neighboring, unexplored region compared to the original area where models were built. The high performance models had in the test dataset suggests they could be used to predict the distribution of S. oculata in unsurveyed areas, at least to some extent. In addition, our study suggests an overlap in the potential distribution of Fe-Mn crusts and S. oculata , which should be addressed in the management of mining to minimize impacts in this community and diminish the loss of biodiversity in RGR. There are a few limitations that should be considered during the modeling approach. Species occurrence in different areas can be difficult to model, as SDMs can fail to account for biotic processes, such as competition and predator-prey interactions, and due to shortfalls in the available data, such as sampling bias and lack of key drivers of habitat suitability 91 . The absolute uncertainty of the model prediction is unknown, but the bootstrap procedures provided a measure of internal consistency across the models 61 . Uncertainty maps are a key resource when applying predictions of distribution models to management measures. The uncertainty of the ensemble model was the lowest near the rift border on both sides, and higher below 1,000 m depth or in some areas on top of the plateaus. Predictions in the rift border had more confidence, which could be explained by a higher sampling effort in this region. Thus, it is advised to use our models carefully to predict areas away from the rift. Increase in model performance and reduction in uncertainty may be achieved with (a) a regularly spaced sampling regime that covers the entirety of the environmental conditions observed in the region of interest 92 , (b) inclusion of key environmental drivers for sponges, such as sediment type, current regimes, and nutrients 76 , 77 , (c) and broader high-resolution surveys, that are still scarce for the area 71 . Rio Grange Rise had been treated as an area beyond national jurisdictions (ABNJ) and under regulations of the International Seabed Authority (ISA). In December 2018, Brazil presented a partial revised submission to the Commission on Limits of The Continental Shelf (CLCS), which includes RGR as part of its continental margin 20 . If approved, RGR will become a region under the jurisdiction and sovereign rights of Brazil, along with its mineral resources. This creates unforeseen implications for the management of RGR, as the regulations that will govern this area are still uncertain. Nevertheless, this issue should not undermine research that produces data and results which could be used to inform management and conservation planning in the area. Declarations Data availability The data that support the findings of this study are available from the corresponding author upon reasonable request. Code availability Code is available at the GitHub repositories https://github.com/ correapvf/SDM-RGR-2022 and https://github.com/correapvf/caretSDM. Acknowledgements This study was funded by the Fundação de Amparo a Pesquisa do Estado de São Paulo (FAPESP, BR) grant number 2014/50820-7 and the Natural Environment Research Council (NERC, UK), as part of the Brazil-UK joint project “Marine ferromanganese deposits: a major resource of E-tech elements”. PVFC was funded by FAPESP grant number 2017/11884-8 and PYGS was funded by CNPq grant number 301554/2019-6. We are grateful to the Captain and Crew Members of the N/Oc. AlphaCrucis and RSS Discovery for their professionalism and dedication to data acquisition. Author contributions The authors contributed equally for this work. Competing interests The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. References Beazley, L. et al. Predicted distribution of the glass sponge Vazella pourtalesi on the Scotian Shelf and its persistence in the face of climatic variability. PLOS ONE 13 , e0205505 (2018). Hogg, M. M. et al. Deep-sea Sponge Grounds: Reservoirs of Biodiversity . (UNEP-WCMC, 2010). Maldonado, M. et al. Sponge Grounds as Key Marine Habitats: A Synthetic Review of Types, Structure, Functional Roles, and Conservation Concerns. in Marine Animal Forests: The Ecology of Benthic Biodiversity Hotspots (eds. Rossi, S., Bramanti, L., Gori, A. & Orejas Saco del Valle, C.) 1–39 (Springer International Publishing, 2016). doi: 10.1007/978-3-319-17001-5_24-1 . Cook, S. E., Conway, K. W. & Burd, B. Status of the glass sponge reefs in the Georgia Basin. Marine Environmental Research 66 , S80–S86 (2008). Klitgaard, A. B. The fauna associated with outer shelf and upper slope sponges (Porifera, Demospongiae) at the Faroe Islands, northeastern Atlantic. Sarsia 80 , 1–22 (1995). Beazley, L., Kenchington, E., Yashayaev, I. & Murillo, F. J. Drivers of epibenthic megafaunal composition in the sponge grounds of the Sackville Spur, northwest Atlantic. Deep Sea Research Part I: Oceanographic Research Papers 98 , 102–114 (2015). Bett, B. J. & Rice, A. L. The influenceof hexactinellid sponge (Pheronema carpenteri) spicules on the patchy distribution of macrobenthos in the porcupine seabight (bathyal ne atlantic). Ophelia 36 , 217–226 (1992). Bo, M. et al. Role of deep sponge grounds in the Mediterranean Sea: a case study in southern Italy. in Ancient Animals, New Challenges: Developments in Sponge Research (eds. Maldonado, M., Turon, X., Becerro, M. & Jesús Uriz, M.) 163–177 (Springer Netherlands, 2012). doi: 10.1007/978-94-007-4688-6_14 . Marliave, J. B., Conway, K. W., Gibbs, D. M., Lamb, A. & Gibbs, C. Biodiversity and rockfish recruitment in sponge gardens and bioherms of southern British Columbia, Canada. Mar Biol 156 , 2247–2254 (2009). Cathalot, C. et al. Cold-water coral reefs and adjacent sponge grounds: hotspots of benthic respiration and organic carbon cycling in the deep sea. Frontiers in Marine Science 2 , (2015). Maldonado, M., Navarro, L., Grasa, A., Gonzalez, A. & Vaquerizo, I. Silicon uptake by sponges: a twist to understanding nutrient cycling on continental margins. Sci Rep 1 , 30 (2011). Pile, A. J. & Young, C. M. The natural diet of a hexactinellid sponge: Benthic–pelagic coupling in a deep-sea microbial food web. Deep Sea Research Part I: Oceanographic Research Papers 53 , 1148–1156 (2006). Rossi, S. The destruction of the ‘animal forests’ in the oceans: Towards an over-simplification of the benthic ecosystems. Ocean & Coastal Management 84 , 77–85 (2013). Ramirez-Llodra, E. et al. Man and the Last Great Wilderness: Human Impact on the Deep Sea. PLOS ONE 6 , e22588 (2011). Rowden, A. A. et al. High-Resolution Habitat Suitability Models for the Conservation and Management of Vulnerable Marine Ecosystems on the Louisville Seamount Chain, South Pacific Ocean. Front. Mar. Sci. 4 , 335 (2017). International Guidelines for the Management of Deep-Sea Fisheries in the High-Seas . (Food and Agriculture Organization of the United Nations, 2009). Mohriak, W. Genesis and evolution of the South Atlantic volcanic islands offshore Brazil. Geo-Mar Lett 40 , 1–33 (2020). Cavalcanti, J. A. D., Santos, R. V., Lacasse, C. M., Rojas, J. N. L. & Nobrega, M. Potential Mineral Resources of Phosphates and Trace Elements on the Rio Grande Rise, South Atlantic Ocean. 44th Underwater Mining Conference (2015). Montserrat, F. et al. Deep-sea mining on the Rio Grande Rise (Southwestern Atlantic): A review on environmental baseline, ecosystem services and potential impacts. Deep Sea Research Part I: Oceanographic Research Papers 145 , 31–58 (2019). da Silva, A. P. Brazil advances over the Area: The inclusion of the Rio Grande Rise within the Brazilian outer continental shelf and its consequences for other states and for the common heritage of mankind. Marine Policy 125 , 104399 (2021). Benites, M., Hein, J. R., Mizell, K., Blackburn, T. & Jovane, L. Genesis and Evolution of Ferromanganese Crusts from the Summit of Rio Grande Rise, Southwest Atlantic Ocean. Minerals 10 , 349 (2020). Sousa, I. M. C. et al. Mineralogy and chemical composition of ferromanganese crusts from the Cruzeiro do Sul Lineament - Rio Grande Rise, South Atlantic. Journal of South American Earth Sciences 108 , 103207 (2021). Christiansen, B., Denda, A. & Christiansen, S. Potential effects of deep seabed mining on pelagic and benthopelagic biota. Marine Policy 114 , 103442 (2020). Levin, L. A. et al. Defining “serious harm” to the marine environment in the context of deep-seabed mining. Marine Policy 74 , 245–259 (2016). Schmidt, C. W. Going Deep: Cautious Steps toward Seabed Mining. Environmental Health Perspectives 123 , A234–A241 (2015). Dunn, D. C. et al. A strategy for the conservation of biodiversity on mid-ocean ridges from deep-sea mining. Science Advances 4 , eaar4313 (2018). Jones, D. O. B., Ardron, J. A., Colaço, A. & Durden, J. M. Environmental considerations for impact and preservation reference zones for deep-sea polymetallic nodule mining. Marine Policy 118 , (2020). Lodge, M. W. & Verlaan, P. A. Deep-Sea Mining: International Regulatory Challenges and Responses. Elements 14 , 331–336 (2018). Miller, K. A., Thompson, K. F., Johnston, P. & Santillo, D. An Overview of Seabed Mining Including the Current State of Development, Environmental Impacts, and Knowledge Gaps. Front. Mar. Sci. 4 , 418 (2018). Gravier, Ch. Note sur une Actinie ( Thoracactis n. g., Topsenti n. sp.) et un Annélide Polychète ( Hermadion Faiweli n. sp.), commensaux d’une Éponge siliceuse ( Sarostegia oculata Topsent). Bulletin du Musée Océanographique de Monaco 344 , 1–20 (1918). Hajdu, E., Castello-Branco, C., Lopes, D. A., Sumida, P. Y. G. & Perez, J. A. A. Deep-sea dives reveal an unexpected hexactinellid sponge garden on the Rio Grande Rise (SW Atlantic). A mimicking habitat? Deep Sea Research Part II: Topical Studies in Oceanography 146 , 93–100 (2017). Topsent, E. Sarostegia oculata , Hexactinellide nouvelle des îles du Cap-Vert. Bulletin du Musée Océanographique de Monaco 10 , 1–8 (1904). Dohrmann, M. et al. Systematics and spicule evolution in dictyonal sponges (Hexactinellida: Sceptrulophora) with description of two new species. Zoological Journal of the Linnean Society 163 , 1003–1025 (2011). Tabachnick, K. R., Menshenina, L. L., Lopes, D. A. & Hajdu, E. Two new Hyalonema species (Hyalonematidae: Amphidiscosida) from eastern and south-eastern Brazil, and further Hexactinellida (Porifera) collected from seamounts off south-eastern Brazil by the RV ‘Marion Dufresne’ MD55 expedition. J. Mar. Biol. Ass. 89 , 1243–1250 (2009). Reiswig, H. M. Family Farreidae Gray, 1872. in Systema Porifera: a guide to the classification of sponges (eds. Hooper, J. N. A., Van Soest, R. W. M. & Willenz, P.) 1332–1340 (Springer, 2002). doi: 10.1007/978-1-4615-0747-5_136 . Howell, K.-L., Piechaud, N., Downie, A.-L. & Kenny, A. The distribution of deep-sea sponge aggregations in the North Atlantic and implications for their effective spatial management. Deep Sea Research Part I: Oceanographic Research Papers 115 , 309–320 (2016). Marshall, C. E., Glegg, G. A. & Howell, K. L. Species distribution modelling to support marine conservation planning: The next steps. Marine Policy 45 , 330–332 (2014). Reiss, H. et al. Benthos distribution modelling and its relevance for marine ecosystem management. ICES Journal of Marine Science 72 , 297–315 (2015). Guilhon, M., Montserrat, F. & Turra, A. Recognition of ecosystem-based management principles in key documents of the seabed mining regime: implications and further recommendations. ICES Journal of Marine Science 78 , 884–899 (2021). Wedding, L. M. et al. Managing mining of the deep seabed. Science 349 , 144–145 (2015). Wedding, L. M. et al. From principles to practice: a spatial approach to systematic conservation planning in the deep sea. Proceedings of the Royal Society B: Biological Sciences 280 , 20131684–20131684 (2013). Kenchington, E. L. et al. Use of Species Distribution Modeling in the Deep Sea. Canadian Technical Report of Fisheries and Aquatic Sciences 3296 , 1–76 (2019). Ramiro-Sánchez, B. et al. Characterization and Mapping of a Deep-Sea Sponge Ground on the Tropic Seamount (Northeast Tropical Atlantic): Implications for Spatial Management in the High Seas. Frontiers in Marine Science 6 , (2019). Walbridge, S., Slocum, N., Marjean Pobuda & Wright, D. Unified Geomorphological Analysis Workflows with Benthic Terrain Modeler. Geosciences 8 , 94 (2018). Guisan, A., Thuiller, W. & Zimmermann, N. E. Habitat Suitability and Distribution Models: With Applications in R . (Cambridge University Press, 2017). doi: 10.1017/9781139028271 . Naimi, B., Hamm, N. A. S., Groen, T. A., Skidmore, A. K. & Toxopeus, A. G. Where is positional uncertainty a problem for species distribution modelling? Ecography 37 , 191–203 (2014). Merow, C. et al. What do we gain from simplicity versus complexity in species distribution models? Ecography 37 , 1267–1281 (2014). R Core Team. R: A Language and Environment for Statistical Computing . (R Foundation for Statistical Computing, 2020). Kuhn, M. Building Predictive Models in R Using the caret Package. J. Stat. Soft. 28 , (2008). Strobl, C., Malley, J. & Tutz, G. An introduction to recursive partitioning: Rationale, application, and characteristics of classification and regression trees, bagging, and random forests. Psychological Methods 14 , 323–348 (2009). Elith, J., Leathwick, J. R. & Hastie, T. A working guide to boosted regression trees. Journal of Animal Ecology 77 , 802–813 (2008). Elith, J. et al. A statistical explanation of MaxEnt for ecologists: Statistical explanation of MaxEnt. Diversity and Distributions 17 , 43–57 (2011). Phillips, S. J., Anderson, R. P. & Schapire, R. E. Maximum entropy modeling of species geographic distributions. Ecological Modelling 190 , 231–259 (2006). Renner, I. W. & Warton, D. I. Equivalence of MAXENT and Poisson Point Process Models for Species Distribution Modeling in Ecology. Biometrics 69 , 274–281 (2013). Wood, S. N. Fast stable restricted maximum likelihood and marginal likelihood estimation of semiparametric generalized linear models. Journal of the Royal Statistical Society: Series B (Statistical Methodology) 73 , 3–36 (2011). Basheer, I. A. & Hajmeer, M. Artificial neural networks: fundamentals, computing, design, and application. Journal of Microbiological Methods 43 , 3–31 (2000). Elith, J., Ferrier, S., Huettmann, F. & Leathwick, J. The evaluation strip: A new and robust method for plotting predicted responses from species distribution models. Ecological Modelling 186 , 280–289 (2005). Thuiller, W., Lafourcade, B., Engler, R. & Araújo, M. B. BIOMOD – a platform for ensemble forecasting of species distributions. Ecography 32 , 369–373 (2009). Dormann, C. F. et al. Model averaging in ecology: a review of Bayesian, information-theoretic, and tactical approaches for predictive inference. Ecological Monographs 88 , 485–504 (2018). Araújo, M. B. & New, M. Ensemble forecasting of species distributions. Trends in Ecology & Evolution 22 , 42–47 (2007). Anderson, O. F. et al. Habitat suitability models for predicting the occurrence of vulnerable marine ecosystems in the seas around New Zealand. Deep Sea Research Part I: Oceanographic Research Papers 115 , 265–292 (2016). Fourcade, Y., Besnard, A. G. & Secondi, J. Paintings predict the distribution of species, or the challenge of selecting environmental predictors and evaluation statistics. Global Ecology and Biogeography 27 , 245–256 (2018). Liu, C., Berry, P. M., Dawson, T. P. & Pearson, R. G. Selecting thresholds of occurrence in the prediction of species distributions. Ecography 28 , 385–393 (2005). Valavi, R., Guillera-Arroita, G., Lahoz-Monfort, J. J. & Elith, J. Predictive performance of presence-only species distribution models: a benchmark study with reproducible code. Ecological Monographs 92 , e01486 (2022). Saito, T. & Rehmsmeier, M. The Precision-Recall Plot Is More Informative than the ROC Plot When Evaluating Binary Classifiers on Imbalanced Datasets. PLOS ONE 10 , e0118432 (2015). Flach, P. & Kull, M. Precision-Recall-Gain Curves: PR Analysis Done Right. in Advances in Neural Information Processing Systems (eds. Cortes, C., Lawrence, N., Lee, D., Sugiyama, M. & Garnett, R.) vol. 28 838–846 (Massachusetts Institute of Technology (MIT) Press, 2015). Allouche, O., Tsoar, A. & Kadmon, R. Assessing the accuracy of species distribution models: prevalence, kappa and the true skill statistic (TSS): Assessing the accuracy of distribution models. Journal of Applied Ecology 43 , 1223–1232 (2006). Phillips, S. J. & Elith, J. POC plots: calibrating species distribution models with presence-only data. Ecology 91 , 2476–2484 (2010). García, S., Fernández, A., Luengo, J. & Herrera, F. Advanced nonparametric tests for multiple comparisons in the design of experiments in computational intelligence and data mining: Experimental analysis of power. Information Sciences 180 , 2044–2064 (2010). Calvo, B. & Santafé, G. scmamp: Statistical Comparison of Multiple Algorithms in Multiple Problems. The R Journal 8 , 248–256 (2016). Jovane, L. et al. Multidisciplinary Scientific Cruise to the Rio Grande Rise. Front. Mar. Sci. 6 , 252 (2019). Gollner, S. et al. Resilience of benthic deep-sea fauna to mining activities. Marine Environmental Research 129 , 76–101 (2017). Abdul Wahab, M. A., Fromont, J., Gomez, O., Fisher, R. & Jones, R. Comparisons of benthic filter feeder communities before and after a large-scale capital dredging program. Marine Pollution Bulletin 122 , 176–193 (2017). Dolan, M. F. J., Grehan, A. J., Guinan, J. C. & Brown, C. Modelling the local distribution of cold-water corals in relation to bathymetric variables: Adding spatial context to deep-sea video data. Deep Sea Research Part I: Oceanographic Research Papers 55 , 1564–1579 (2008). Howell, K. L., Holt, R., Endrino, I. P. & Stewart, H. When the species is also a habitat: Comparing the predictively modelled distributions of Lophelia pertusa and the reef habitat it forms. Biological Conservation 144 , 2656–2665 (2011). Rengstorf, A. M., Mohn, C., Brown, C., Wisz, M. S. & Grehan, A. J. Predicting the distribution of deep-sea vulnerable marine ecosystems using high-resolution data: Considerations and novel approaches. Deep Sea Research Part I: Oceanographic Research Papers 93 , 72–82 (2014). Rengstorf, A. M., Yesson, C., Brown, C. & Grehan, A. J. High-resolution habitat suitability modelling can improve conservation of vulnerable marine ecosystems in the deep sea. Journal of Biogeography 40 , 1702–1714 (2013). Robert, K., Jones, D. O. B., Roberts, J. M. & Huvenne, V. A. I. Improving predictive mapping of deep-water habitats: Considering multiple model outputs and ensemble techniques. Deep Sea Research Part I: Oceanographic Research Papers 113 , 80–89 (2016). Rodríguez-Basalo, A. et al. High Resolution Spatial Distribution for the Hexactinellid Sponges Asconema setubalense and Pheronema carpenteri in the Central Cantabrian Sea. Front. Mar. Sci. 8 , 612761 (2021). Tong, R., Purser, A., Guinan, J. & Unnithan, V. Modeling the habitat suitability for deep-water gorgonian corals based on terrain variables. Ecological Informatics 13 , 123–132 (2013). Yeo, I. A. et al. Distribution of and hydrographic controls on ferromanganese crusts: Tropic Seamount, Atlantic. Ore Geology Reviews 114 , 103131 (2019). Klitgaard, A. B. & Tendal, O. S. Distribution and species composition of mass occurrences of large-sized sponges in the northeast Atlantic. Progress in Oceanography 61 , 57–98 (2004). Leathwick, J. R., Elith, J., Francis, M. P., Hastie, T. & Taylor, P. Variation in demersal fish species richness in the oceans surrounding New Zealand: an analysis using boosted regression trees. Marine Ecology Progress Series 321 , 267–281 (2006). Thresher, R. et al. Strong Depth-Related Zonation of Megabenthos on a Rocky Continental Margin (∼700–4000 m) off Southern Tasmania, Australia. PLOS ONE 9 , e85872 (2014). Davies, A. J. & Guinotte, J. M. Global Habitat Suitability for Framework-Forming Cold-Water Corals. PLoS ONE 6 , e18483 (2011). Yesson, C., Bedford, F., Rogers, A. D. & Taylor, M. L. The global distribution of deep-water Antipatharia habitat. Deep Sea Research Part II: Topical Studies in Oceanography (2015) doi: 10.1016/j.dsr2.2015.12.004 . Yesson, C. et al. Global habitat suitability of cold-water octocorals: Global distribution of deep-sea octocorals. Journal of Biogeography 39 , 1278–1292 (2012). Barbosa, R. V., Davies, A. J. & Sumida, P. Y. G. Habitat suitability and environmental niche comparison of cold-water coral species along the Brazilian continental margin. Deep Sea Research Part I: Oceanographic Research Papers 155 , 103147 (2020). Ross, L. K., Ross, R. E., Stewart, H. A. & Howell, K. L. The Influence of Data Resolution on Predicted Distribution and Estimates of Extent of Current Protection of Three ‘Listed’ Deep-Sea Habitats. PLoS ONE 10 , e0140061 (2015). Beaulieu, S. E. Life on glass houses: sponge stalk communities in the deep sea. Marine Biology 138 , 803–817 (2001). Vierod, A. D. T., Guinotte, J. M. & Davies, A. J. Predicting the distribution of vulnerable marine ecosystems in the deep sea using presence-background models. Deep Sea Research Part II: Topical Studies in Oceanography 99 , 6–18 (2014). Hirzel, A. & Guisan, A. Which is the optimal sampling strategy for habitat suitability modelling. Ecological Modelling 157 , 331–341 (2002). Appendix D Appendix D is not available with this version. Additional Declarations There is NO Competing Interest. Cite Share Download PDF Status: Under Review 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-1663413","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":111596770,"identity":"50734e37-62cc-4a51-8f3a-ef7ce78f3c66","order_by":0,"name":"Paulo Corrêa","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAz0lEQVRIiWNgGAWjYDAD9vYGIGlgQYIWnjMHQFokSNFyIwFEEaGFv7332YcPNXcSeySfX93wo0ACKNKdgFeLxJnjxjNnHHuW2COdU3azB+gwiTNnN+DVYiCRxszMw3bY2F46J+0GD1CLgUQuEVr+/DtszCN5Ju3mH6K1MLYdluORYD92myhbJM4cY2bs7Xsmx8OTw3ZbxkCCh6Bf+NvbmBl+fLvDw8N+/NnNN39s5IBhiF8LFBwAYh4DEIuHGOUwLewPiFU9CkbBKBgFIwwAAMCfRAY0EWlhAAAAAElFTkSuQmCC","orcid":"https://orcid.org/0000-0001-6474-7355","institution":"Instituto Oceanográfico da Universidade de São Paulo","correspondingAuthor":true,"submittingAuthor":false,"prefix":"","firstName":"Paulo","middleName":"","lastName":"Corrêa","suffix":""},{"id":111596771,"identity":"a0d057c2-ee42-49e7-ab67-b5e83e36f9cf","order_by":1,"name":"Paulo Sumida","email":"","orcid":"","institution":"","correspondingAuthor":false,"submittingAuthor":false,"prefix":"","firstName":"Paulo","middleName":"","lastName":"Sumida","suffix":""}],"badges":[],"createdAt":"2022-05-17 02:15:54","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-1663413/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-1663413/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":22632058,"identity":"909e3d76-dfbc-4225-ae67-e1a4f3c71926","added_by":"auto","created_at":"2022-06-14 13:30:47","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":853642,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003e(a)\u003c/strong\u003e Rio Grande Rise, showing the locations of panels (b) and (c). \u003cstrong\u003e(b)\u003c/strong\u003e MBES of Area 1, used for model training and validation. \u003cstrong\u003e(c)\u003c/strong\u003e MBES of Area 2, used for model testing. Red lines represent dives from (a) \u003cem\u003eHyBIS\u003c/em\u003e and (c) \u003cem\u003eShinkai\u003c/em\u003e. Maps (b) and (c) are in the same scale.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-1663413/v1/f9ace7892f5efcfccc193660.png"},{"id":22632056,"identity":"15bbff6f-619f-4cc3-bd28-f20c10fed22b","added_by":"auto","created_at":"2022-06-14 13:30:46","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":516017,"visible":true,"origin":"","legend":"\u003cp\u003eThe bathymetry-derived variables in Area 1, which were used as explanatory variables in the models for training and validation. \u003cstrong\u003e(a)\u003c/strong\u003e Depth. \u003cstrong\u003e(b)\u003c/strong\u003e Slope. \u003cstrong\u003e(c)\u003c/strong\u003e Broad-scale Bathymetric Position Index (BPI). \u003cstrong\u003e(d)\u003c/strong\u003e Fine-scale BPI. \u003cstrong\u003e(e)\u003c/strong\u003e Euclidian distance from the border of the rift. \u003cstrong\u003e(f)\u003c/strong\u003e Northness. \u003cstrong\u003e(g)\u003c/strong\u003e Eastness. \u003cstrong\u003e(h)\u003c/strong\u003e Rugosity. \u003cstrong\u003e(i) \u003c/strong\u003eCurvature\u003cstrong\u003e.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-1663413/v1/f4f6fb5828b46c23ab8f1e1c.png"},{"id":22632324,"identity":"d4150b5c-3abc-48e6-a265-b5df8245d26d","added_by":"auto","created_at":"2022-06-14 13:35:46","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":540205,"visible":true,"origin":"","legend":"\u003cp\u003eThe bathymetry-derived variables in Area 2, which were used as explanatory variables in the models for testing. \u003cstrong\u003e(a)\u003c/strong\u003e Depth. \u003cstrong\u003e(b)\u003c/strong\u003e Slope. \u003cstrong\u003e(c)\u003c/strong\u003e Broad-scale Bathymetric Position Index (BPI). \u003cstrong\u003e(d)\u003c/strong\u003e Fine-scale BPI. \u003cstrong\u003e(e)\u003c/strong\u003e Euclidian distance from the border of the rift. \u003cstrong\u003e(f)\u003c/strong\u003e Northness. \u003cstrong\u003e(g)\u003c/strong\u003e Eastness. \u003cstrong\u003e(h)\u003c/strong\u003e Rugosity. \u003cstrong\u003e(i) \u003c/strong\u003eCurvature\u003cstrong\u003e.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-1663413/v1/fc7c9b106a2c0eca75f563e6.png"},{"id":22631752,"identity":"9ecc6134-15e6-4693-b63b-0a21a3a1415f","added_by":"auto","created_at":"2022-06-14 13:25:46","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":81294,"visible":true,"origin":"","legend":"\u003cp\u003eMean and standard error of model performance statistics for both validation (orange) and test (blue) data across replicates. \u003cstrong\u003e(a)\u003c/strong\u003e AUC\u003csub\u003eROC\u003c/sub\u003e, \u003cstrong\u003e(b)\u003c/strong\u003e AUC\u003csub\u003ePRG\u003c/sub\u003e, \u003cstrong\u003e(c)\u003c/strong\u003e Sensitivity, \u003cstrong\u003e(d)\u003c/strong\u003e Specificity and \u003cstrong\u003e(e)\u003c/strong\u003e TSS.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-1663413/v1/3c35590bb9c60cbdff9731e9.png"},{"id":22631749,"identity":"433c90d5-a989-4d86-879b-4e31c204d0f1","added_by":"auto","created_at":"2022-06-14 13:25:46","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":96901,"visible":true,"origin":"","legend":"\u003cp\u003ePresence-absence calibration plots for each model fitted with natural splines. \u003cstrong\u003e(a)\u003c/strong\u003e Random Forest, \u003cstrong\u003e(b)\u003c/strong\u003e Boosted Regression Trees, \u003cstrong\u003e(c)\u003c/strong\u003e MaxEnt, \u003cstrong\u003e(d)\u003c/strong\u003e Generalized Additive Models, and \u003cstrong\u003e(e)\u003c/strong\u003e Artificial Neural Networks. The calibration curve is in blue and a confidence interval of ± 2 SD is in orange. The rug plots show model predictions at presences (orange) and absences (blue). The dotted line indicates a 1:1 relationship representing ‘perfect’ calibration.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-1663413/v1/770cc7211f9884912cc7c3ae.png"},{"id":22631756,"identity":"2426330c-f06b-452d-87c9-888578804b3a","added_by":"auto","created_at":"2022-06-14 13:25:47","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":569290,"visible":true,"origin":"","legend":"\u003cp\u003ePrediction maps of \u003cem\u003eSarostegia oculata\u003c/em\u003e in Area 1, using \u003cstrong\u003e(a)\u003c/strong\u003e Random Forest, \u003cstrong\u003e(b)\u003c/strong\u003e Boosted Regression Trees, \u003cstrong\u003e(c)\u003c/strong\u003e MaxEnt, \u003cstrong\u003e(d)\u003c/strong\u003e Generalized Additive Models, \u003cstrong\u003e(e)\u003c/strong\u003e Artificial Neural Networks, and \u003cstrong\u003e(f)\u003c/strong\u003e Ensemble models.\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-1663413/v1/5def6324f47c5b3e20569dbb.png"},{"id":22631758,"identity":"1c10af9a-3c02-4372-a7a1-f2765daa2936","added_by":"auto","created_at":"2022-06-14 13:25:47","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":574797,"visible":true,"origin":"","legend":"\u003cp\u003ePrediction maps of \u003cem\u003eSarostegia oculata\u003c/em\u003e in Area 2, using \u003cstrong\u003e(a)\u003c/strong\u003e Random Forest, \u003cstrong\u003e(b)\u003c/strong\u003e Boosted Regression Trees, \u003cstrong\u003e(c)\u003c/strong\u003e MaxEnt, \u003cstrong\u003e(d)\u003c/strong\u003e Generalized Additive Models, \u003cstrong\u003e(e)\u003c/strong\u003e Artificial Neural Networks, and \u003cstrong\u003e(f)\u003c/strong\u003e Ensemble models.\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-1663413/v1/138fd34e2dfb76aa543325fa.png"},{"id":22631759,"identity":"6b112397-eebd-47b7-92f9-c0e848a95681","added_by":"auto","created_at":"2022-06-14 13:25:47","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":329175,"visible":true,"origin":"","legend":"\u003cp\u003eUncertainty (CV) for the ensemble distribution model of \u003cstrong\u003e(a)\u003c/strong\u003e Area 1 and \u003cstrong\u003e(b)\u003c/strong\u003e Area 2.\u003c/p\u003e","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-1663413/v1/a99a865e374d87581d54a02d.png"},{"id":22631755,"identity":"4c21bbc5-6357-4aff-a957-d660a33c22d0","added_by":"auto","created_at":"2022-06-14 13:25:46","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":104320,"visible":true,"origin":"","legend":"\u003cp\u003eResponse curves of the likelihood of presence of \u003cem\u003eSarostegia oculata\u003c/em\u003e for each predictor variable for the Random Forest (RF), Boosted Regression Trees (BRT), MaxEnt, Generalized Additive Models (GAM), and Artificial Neural Networks (ANN) models. Rug plot inside bottom of panels show distribution of sites across that variable, in percentiles.\u003c/p\u003e","description":"","filename":"floatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-1663413/v1/04aa1a0c6491c16a71351158.png"},{"id":22632325,"identity":"97a63eb1-ec0f-41c7-b711-18a5ff9958ea","added_by":"auto","created_at":"2022-06-14 13:35:50","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":485863,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-1663413/v1/bf25b210-d048-40d7-95c2-1e40ce48c652.pdf"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"Distribution models of the branched hexactinellid Sarostegia oculata in Rio Grande Rise (SW Atlantic)","fulltext":[{"header":"Introduction","content":"\u003cp\u003eDeep-sea sponges have importance as organisms that can provide multiple ecosystem goods and services, and thus have increasingly been mentioned in the literature in recent years\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. Habitats dominated by sponges are globally distributed, forming structurally complex and often highly diverse communities, along with corals\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e,\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. They have key roles in ecosystem functions, including habitat provision for associated fauna\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e,\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e, increasing biodiversity\u003csup\u003e\u003cspan additionalcitationids=\"CR7 CR8\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e, promoting silicon and organic carbon cycling\u003csup\u003e\u003cspan additionalcitationids=\"CR11\" citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e, and potentially acting as nursery habitat for many species\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eVulnerable Marine Ecosystems (VMEs) are characterized as areas with low resilience and slow recovery from anthropogenic disturbances such as bottom trawling and mining (extraction of fossil fuels, gas, or minerals)\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. Some taxonomic groups, including sponges, are considered indicators of VMEs as a consequence of having slow growth rates, longevity, late maturity, fragility, as they form three-dimensional structures associated with diverse communities\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. They have been used to assist agencies responsible for the protection of particular ocean regions, such as slopes, seamounts, and canyons\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e. Increasing knowledge of the distribution and biology of species is necessary to create a consistent network of marine protected areas (MPAs), which can diminish human impacts on the ocean and prevent loss of biodiversity.\u003c/p\u003e \u003cp\u003eThe Rio Grande Rise (RGR) is a major positive feature in the South Atlantic, between the parallels 28\u0026deg;\u0026ndash;35\u0026deg; South and meridians 28\u0026deg;\u0026ndash;39\u0026deg; West, located about 1200 km east off the Brazilian coast, and forming a volcanic plateau that rises from depths of about 5000 m to above 1000 m in some areas\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e. The western region of RGR (WRGR) forms an elliptical bulge of 500 km \u0026times; 300 km that includes seamounts and guyots towering over or juxtaposing with this massive platform\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e; it is intersected by a NW-SE trending submarine rift between 20\u0026ndash;40 km wide (also referred to as \u0026lsquo;graben\u0026rsquo; or Cruzeiro do Sul Lineament). RGR has gained special attention from researchers and the Brazilian government in recent years\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e,\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e due to prospecting surveys that revealed its potential for the mining of Fe-Mn crusts\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e,\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e,\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e. Impacts created by mining activities may range from direct destruction of habitats and formation of toxic and particle-rich sediment plumes\u003csup\u003e\u003cspan additionalcitationids=\"CR24\" citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e to noise, vibration, and light, which may lead to significant (potentially irreversible) damage to deep-sea communities and ecosystems and loss of biodiversity on mining sites and surrounding areas\u003csup\u003e19,26\u0026minus;29\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003e \u003cem\u003eSarostegia oculata\u003c/em\u003e is a branched hexactinellid that forms an erect, complex surface, similar to the shape of a stony coral. It has a close association with a zoanthid, likely \u003cem\u003eThoractis topsenti\u003c/em\u003e\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e, growing on and within the skeleton in a similar way to coral polyps, thus mimicking the 3D skeletal framework of actual corals\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e. It was first described at Cape Verde by Topsent\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e, and further sampled throughout the Atlantic\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e,\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e and Indian\u003csup\u003e\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e oceans. More recently, Hajdu et al.\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e recorded extensive sponge gardens in RGR, being the dominant organism in areas rich in Fe-Mn crusts that play a fundamental role in supporting communities in the area.\u003c/p\u003e \u003cp\u003eSpecies distribution models (SDMs; also known as habitat suitability models) are a method to predict the likelihood of a species to occur in a given location, based on their observed relationship with environmental conditions. SDM is widely used in ecology and conservation, especially in the deep sea where data is scarce, and such models are useful for the development of conservation measures and marine ecosystem management by predicting species distributions beyond surveyed areas\u003csup\u003e\u003cspan additionalcitationids=\"CR37\" citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eIn the present study, we built such models to predict the distribution of \u003cem\u003eS. oculata\u003c/em\u003e in two areas with available high-resolution multibeam data at RGR, surrounding the rift within the bulge of WRGR. The growing use of mineral resources in our global economy will surely lead to future commercial extraction of mineral resources in the deep sea, including, but not limited to, Fe-Mn crusts\u003csup\u003e\u003cspan additionalcitationids=\"CR40\" citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e\u003c/sup\u003e. Such models can be used to inform spatial management planning for protecting VMEs and environmental impact assessments that minimize the effects of mining on deep-sea ecosystems\u003csup\u003e\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e,\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e\u003c/sup\u003e. Our study is the first to create a high-resolution species distribution model of a VME indicator species in Rio Grande Rise. \u003cem\u003eSarostegia oculata\u003c/em\u003e is a dominant organism on substrates rich in Fe-Mn crusts, which play a fundamental role to support communities and provide key ecosystem functions in RGR. The five models built here predicted a high likelihood of \u003cem\u003eS. oculata\u003c/em\u003e along with the rift borders with a low degree of uncertainty, in an area aimed at exploration of Fe-Mn crust deposits. This reinforces the relationship between \u003cem\u003eS. oculata\u003c/em\u003e, the Fe-Mn crusts, and the rift at RGR, implicating that direct destruction of habitats caused by the mining of Fe-Mn crusts could lead to a severe impact in the communities sustained by \u003cem\u003eS. oculata\u003c/em\u003e, which should be addressed in the management and MPA network design for RGR. There are still large gaps in the knowledge of the benthic fauna distribution and ecosystem functions on RGR, but our study provides expert advice on the occurrence of deep-sea sponges in the area and represents initial steps to guide future surveys and conservation plans in the region.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eData collection\u003c/h2\u003e \u003cp\u003eTwo surveyed areas near the center of WRGR were used in this study. The first area (here named Area 1) was delimited between latitudes 30\u0026ordm;35\u0026rsquo;S\u0026ndash;31\u0026deg;03\u0026rsquo;S and longitudes 35\u0026ordm;36\u0026rsquo;W\u0026ndash;36\u0026deg;16\u0026rsquo;W (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ea). This area is ~\u0026thinsp;2,000 km\u0026sup2;, ranging from 600 m to 1,800 m depth, and encompasses a small segment of the rift and a plateau on both sides, here named NE and SW plateaus (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eb). It also contains a terrace in the NE plateau, which runs almost parallel to the rift border, a slide south of the SW plateau, and a canyon that splits a section of the SW plateau in two, one next to the rift border and another named the inner SW plateau. The south and west of the SW plateau are neighbored by a lower plane, which extends beyond the MBES data. The second area (Area 2) is smaller, located about 30 km northwest of Area 1, delimited between latitudes 30\u0026deg;19\u0026rsquo;S\u0026ndash;30\u0026deg;25\u0026rsquo;S and longitudes 35\u0026deg;57\u0026rsquo;W\u0026ndash;36\u0026deg;06\u0026rsquo;W, is ~\u0026thinsp;77 km\u0026sup2;, ranging from 600 to 1400 m depth. It encompasses a small area of the northeastern plateau and the rift (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ec).\u003c/p\u003e\u003cp\u003eArea 1 (Figs.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eb and \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) was surveyed with the \u003cem\u003eN/Oc. Alpha Crucis\u003c/em\u003e, from January 30 to February 21, 2018, and \u003cem\u003eRSS Discovery\u003c/em\u003e, from October 26 to November 8, 2018, as part of the FAPESP/RCUK funded project Marine E-Tech. \u003cem\u003eN/Oc. Alpha Crucis\u003c/em\u003e used the Reason 7160 multibeam echosounder (MBES) operating at 41 kHz. \u003cem\u003eRSS Discovery\u003c/em\u003e was equipped with a ship-mounted Kongsberg EM122, and multibeam was processed using Caris HIPS and SIPS v9.1.8. The data in both products were gridded at 25 m using NaviModel Producer 4.3.2, then fused with ArcGIS Pro 2.9, using the Mosaic tool with the \u0026ldquo;Blend\u0026rdquo; mosaic operator. The survey in the \u003cem\u003eRSS Discovery\u003c/em\u003e included 13 dives (codes HY31\u0026ndash;43) of the Robotic Underwater Vehicle, \u003cem\u003eRUV HyBIS\u003c/em\u003e (Hydraulic Benthic Interactive Sampler). \u003cem\u003eHyBIS\u003c/em\u003e was equipped with a Sony Full HD camera that recorded over 36 hours and 26 km of video footage of the seafloor and it was connected to the ship's USBL system to record its position (latitude, longitude, and depth).\u003c/p\u003e\u003cp\u003eArea 2 (Figs.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ec and \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e) was surveyed with the submersible \u003cem\u003eShinkai 6500\u003c/em\u003e and \u003cem\u003eYokosuka\u003c/em\u003e support vessel, from April 13 to May 5, 2013, as part of the \u0026lsquo;Iata-Piuna\u0026rsquo; expedition. Yokosuka was equipped with multibeam Sea Beam 2112.004, operating at 12 kHz. The data was also gridded at 25 m using NaviModel Producer. \u003cem\u003eShinkai\u003c/em\u003e 6500 was equipped with two HD video cameras. One of them was fixed at the bow, which recorded 4.5 hours and 2,282 m of footage from the seabed during the dive 6K1338. The other video camera was mobile and was ignored in this study so the analysis would be more consistent with the \u003cem\u003eRUV HyBIS\u003c/em\u003e.\u003c/p\u003e \u003cp\u003eThe \u003cem\u003eHyBIS\u003c/em\u003e and \u003cem\u003eShinkai\u003c/em\u003e videos were annotated and counted for benthic fauna while classifying them to the lowest possible taxonomic group. The branched hexactinellid \u003cem\u003eSarostegia oculata\u003c/em\u003e was the most commonly observed habitat-forming VME indicator species in both Area 1 and a 2. The timestamp of the video was used to obtain the coordinates of the observations from the \u003cem\u003eHyBIS\u003c/em\u003e and \u003cem\u003eShinkai\u003c/em\u003e track data. Grid cells where at least one individual of \u003cem\u003eS. oculata\u003c/em\u003e were classified as \u0026ldquo;presence\u0026rdquo;, and grid cells with overlapping track and no observations of \u003cem\u003eS. oculata\u003c/em\u003e were classified as \u0026ldquo;absence\u0026rdquo;. Each of these grid cells are hereafter called records or sites, used to train and evaluate the models.\u003c/p\u003e \u003cp\u003eA set of additional seafloor variables (Figs.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e and \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e) was derived from the bathymetric data using the Benthic Terrain Model 3.0 tool\u003csup\u003e\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e\u003c/sup\u003e in ArcGIS. These variables were slope, aspect (measured in terms of northness and eastness), roughness (11 neighborhood size), curvature, fine-scale (3/30 radius) Bathymetric Position Index (BPI), and broad-scale (20/200 radius) BPI. The rift is a major feature in both areas that may play an important role in the distribution of the species in RGR. As such, we included the distance from the rift border as another environment variable in our analysis. This was achieved by creating a contour at value 0 from the broad-scale BPI raster. This will result in two contours outlining the NE and SW rift borders, which were used as source for the \u0026ldquo;Euclidean Distance\u0026rdquo; tool in ArcGIS. Before data analysis, collinearity between variables was checked using Pearson\u0026rsquo;s correlation coefficient and the Variance Inflation Factors (VIFs)\u003csup\u003e\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e\u003c/sup\u003e using the \u003cem\u003eusdm\u003c/em\u003e package\u003csup\u003e\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003eModeling methods\u003c/h2\u003e \u003cp\u003eThe distribution of \u003cem\u003eS. oculata\u003c/em\u003e was modeled using five algorithms widely used in species distribution models: Random Forest (RF), Boosted Regression Trees (BRT), MaxEnt, Generalized Additive Models (GAMs), and Artificial Neural Networks (ANN). In general, these models differ in the way they determine the fitted function, and how they handle interactions, model complexity, and overfitting\u003csup\u003e\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e\u003c/sup\u003e. Models were developed using R\u003csup\u003e\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e\u003c/sup\u003e with the framework caret\u003csup\u003e\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e\u003c/sup\u003e. The caret package is a set of functions that streamline the process of creating predictive models and provides a uniform interface for data splitting, variable selection, and model tuning. It uses other packages internally to create the models. We used default parameters for each package, except when stated in the text below or Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. The function \u003cem\u003erfe\u003c/em\u003e was used to perform recursive feature elimination in the models, but no advantage was gained in model performance from removing variables from the models, so we kept all variables in the final models.\u003c/p\u003e \u003cp\u003eRF builds multiple classification trees by taking many bootstrap samples from the training data and making an average prediction over all fitted trees\u003csup\u003e\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e\u003c/sup\u003e. RF uses only a randomly smaller set of the predictor variables (parameter \u003cb\u003emtry\u003c/b\u003e) on each split while growing each tree, which reduces the variance of the final model, with subsequent gains for predictive performance. We set the number of trees to grow (parameter \u003cb\u003entree\u003c/b\u003e) as 1000.\u003c/p\u003e \u003cp\u003eBRT builds several regression trees (parameter \u003cb\u003en.trees\u003c/b\u003e) in a forward stagewise fashion. At each step of model fitting, the algorithm fits each new tree to the residuals of the previously fitted trees and uses only a random subset of data \u003csup\u003e\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e\u003c/sup\u003e. This gradually increases emphasis on observations modeled poorly by the existing collection of trees. The contribution of each tree is controlled by the \u003cb\u003eshrinkage\u003c/b\u003e parameter, as the model-building process usually performs best if it moves slowly down the gradient. \u003cb\u003eInteraction.depth\u003c/b\u003e controls the maximum depth of the individual trees and \u003cb\u003en.minobsinnode\u003c/b\u003e controls the minimum number of observations in terminal nodes.\u003c/p\u003e \u003cp\u003eMaxent estimates species distribution by minimizing the relative entropy between two probability densities\u003csup\u003e\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e,\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e\u003c/sup\u003e. Although it was initially developed as a machine-learning algorithm, MaxEnt has known links to Poisson point process models\u003csup\u003e\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e\u003c/sup\u003e and it can be categorized as a \u0026ldquo;regression-based\u0026rdquo; model. MaxEnt has the ability to fit complex models, controlled by the use of transformed features of the predictor covariates (parameter \u003cb\u003efeature types\u003c/b\u003e). \u003cb\u003eRegularization multiplier\u003c/b\u003e can also be used to smooth and avoid fitting overly complex models. In this work, we employed real absences derived from the dives instead of more commonly used pseudo-absences and used the Java implementation to train the MaxEnt models, version 3.4.4.\u003c/p\u003e \u003cp\u003eGAMs are an extension of General Linear Models (GLMs) but use nonparametric smoothing functions which allow modeling of non-linear relationships between the response and explanatory functions. Multiple smoothing parameter estimation methods are available in the package (parameter \u003cb\u003emethod\u003c/b\u003e)\u003csup\u003e\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e\u003c/sup\u003e. It can also add an extra penalty to each term (parameter \u003cb\u003eselect\u003c/b\u003e), meaning that the smoothing parameter estimation can completely remove terms from the model. We used a binomial distribution with the logit function and thin plate regression splines for the smoothing basis of all terms.\u003c/p\u003e \u003cp\u003eANNs are a deep learning approach that consist of a large number of nodes and connections, typically organized in layers\u003csup\u003e\u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e56\u003c/span\u003e\u003c/sup\u003e. The input layer contains the environmental data, with each input node representing one environmental variable. The information from each input node is fed into the hidden layers, which may contain multiple layers with a variable number of nodes. Data from the last hidden layer is then fed to the output layer which represents the result of the model. The connections between the nodes from one layer to the next, also known as weights, are optimized using a back-propagation process. Our model has a single hidden layer, with 1 to 21 nodes (parameter \u003cb\u003esize\u003c/b\u003e). \u003cb\u003eDecay\u003c/b\u003e is used as a penalty for these connections that both help the optimization process and avoid over-fitting. The feed-forward and back-propagation processes are repeated until the model reaches a pre-defined accuracy, or a maximum set number of runs controlled by the parameter maxit (in our model we set \u003cb\u003emaxit\u003c/b\u003e as 500).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eModeling methods, their implementations during tuning, and the tuned parameters used in the final models.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMethod\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eR package\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eParameter\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eValues\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTuned\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRF\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003erandomForest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003emtry\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1\u0026ndash;9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003eBRT\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003egbm\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003en.trees\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003efrom 100 to 3,000 by 50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2600\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003einteraction.depth\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003efrom 1 to 16 by 3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eshrinkage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e10\u003csup\u003e\u0026minus;\u0026thinsp;1\u003c/sup\u003e, 10\u003csup\u003e\u0026minus;\u0026thinsp;2\u003c/sup\u003e, 10\u003csup\u003e\u0026minus;\u0026thinsp;3\u003c/sup\u003e, 10\u003csup\u003e\u0026minus;\u0026thinsp;4\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e10\u003csup\u003e\u0026minus;\u0026thinsp;3\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003en.minobsinnode\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5, 15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eMaxent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003ecustom\u0026sup1;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003efeature types\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003el, lq, lqp, lqph, lqh, lqpt, lqt, lqpth\u0026sup2;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003elqpth\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eregularization multiplier\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003efrom 0.2 to 2 by 0.2, 2.5, 3, 3.5, 4, 4.5, 5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eGAM\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003emgcv\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003emethod\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eREML, ML\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eML\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eselect\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTRUE, FALSE\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTRUE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003ennet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003esize\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1, 3, 5, 7, 9, 11, 13, 21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e21\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003edecay\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0, 10\u003csup\u003e\u0026minus;\u0026thinsp;1\u003c/sup\u003e, 10\u003csup\u003e\u0026minus;\u0026thinsp;2\u003c/sup\u003e, 10\u003csup\u003e\u0026minus;\u0026thinsp;3\u003c/sup\u003e, 10\u003csup\u003e\u0026minus;\u0026thinsp;4\u003c/sup\u003e, 10\u003csup\u003e\u0026minus;\u0026thinsp;5\u003c/sup\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.1\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003e\u0026sup1; Custom script was used to run the MaxEnt java application and read its output within caret.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e\u0026sup2; l\u0026thinsp;=\u0026thinsp;linear, q\u0026thinsp;=\u0026thinsp;quadratic, h\u0026thinsp;=\u0026thinsp;hinge, p\u0026thinsp;=\u0026thinsp;product and t\u0026thinsp;=\u0026thinsp;threshold.\u003c/p\u003e \u003cp\u003eThe relationship between the environmental layers and the predicted probability of presence was analyzed using response curves produced using the methodology described by Elith et al.\u003csup\u003e\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e\u003c/sup\u003e. All variables were held constant at their mean except the target variable, which varied at 100 points across its range. Then, the prediction of each algorithm was computed for each of the 100 values of the target variable and used to produce the response curves. The advantage of this method is that it can be applied to any model and be compared against different model techniques, but it does not account for interactions between variables. The importance of each variable in the final output was computed using the methodology described by Thuiller et al.\u003csup\u003e\u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e58\u003c/span\u003e\u003c/sup\u003e. The target variable was shuffled, while the others were left untouched. The Pearson's correlation coefficient between the reference and shuffled predictions was calculated and the score \u003cem\u003e1 - correlation\u003c/em\u003e is returned. Scores near one mean the variable has a high influence on the model, and a score of zero assumes no influence of the variable on the model. The process was repeated 100 times for each variable.\u003c/p\u003e \u003cp\u003eAn ensemble model is a popular technique for reducing the uncertainty of model predictions\u003csup\u003e\u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e59\u003c/span\u003e\u003c/sup\u003e. They may reduce prediction errors and decrease mean bias due to the choice of method\u003csup\u003e\u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e60\u003c/span\u003e\u003c/sup\u003e. This is usually achieved by computing the weighted or unweighted averages of the different model outputs. Ensemble models also permit uncertainty caused by the different predictions to be calculated. In this study, all model predictions were rescaled between 0 and 1, and their unweighted average was used to build the ensemble model. Spatial measures of uncertainty were calculated using a bootstrap technique\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e,\u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e\u003c/sup\u003e. A random sample of the presence-absence data, of equal size, was drawn with replacement, and the five models were constructed with the same settings as the originals. Predictions were then made to the study area, and this process was repeated 200 times. Model uncertainty was represented as the coefficient of variation (CV) of the bootstrap output, i.e, the standard deviation divided by the mean.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003eModel evaluation\u003c/h2\u003e \u003cp\u003eThe data in this study was divided into train, validation (also referred to as development in the literature), and test sets. The train set is used to train the model, while the validation set is used to evaluate the model during tuning. After the best parameters are selected, both train and validation sets are used to train a final model, which is evaluated using the test set. Data obtained from Area 1 with \u003cem\u003eHyBIS\u003c/em\u003e was used as train and validation sets. Such data inherits a high spatial dependency, as records are available next to each other throughout each dive, and can lead to overestimation of model performance\u003csup\u003e\u003cspan citationid=\"CR62\" class=\"CitationRef\"\u003e62\u003c/span\u003e\u003c/sup\u003e. To provide a larger level of spatial independence between the training and the validation data, the first or last 20% records in each dive were selected to be used as a validation set. Consequently, the remaining 80% were used to train the model. The region of the dive which was cut from training, whether from the beginning or towards the end, was selected at random for each dive, and this process was repeated 25 times. During model tuning, the AUC\u003csub\u003eROC\u003c/sub\u003e metric was used to select the best models. Data obtained from Area 2 with \u003cem\u003eShinkai\u003c/em\u003e is completely independent from Area 1 and was used exclusively as a test set for the final model. The evaluation was performed 25 times using a random selection of 70% of the test data (subsampling) in each interaction.\u003c/p\u003e \u003cp\u003eModels were evaluated using five metrics to cover different aspects of the modeling performance: (1) area under the receiver operating characteristic curve (AUC\u003csub\u003eROC\u003c/sub\u003e), (2) area under the precision-recall gain curve (AUC\u003csub\u003ePRG\u003c/sub\u003e), (3) Sensitivity, (4) Specificity, and (5) True Skill Statistics (TSS). AUC\u003csub\u003eROC\u003c/sub\u003e and AUC\u003csub\u003ePRG\u003c/sub\u003e are threshold-independent measures, while Sensitivity, Specificity, and TSS need to convert the predicted likelihood of the model (a value between 0 and 1) into a presence/absence classification using a threshold. Here we used the threshold which balances sensitivity and specificity, which minimizes both commission and omission errors, and is recommended for SDMs model\u003csup\u003e\u003cspan citationid=\"CR63\" class=\"CitationRef\"\u003e63\u003c/span\u003e\u003c/sup\u003e. Note that thresholds were selected using the validation data set.\u003c/p\u003e \u003cp\u003eAUC\u003csub\u003eROC\u003c/sub\u003e assesses the ability of models to discriminate presence from absence sites and is widely used in species distribution models. The ROC curve is created by calculating the true positive rate (the proportion of presences correctly predicted, also known as sensitivity or recall) and the false positive rate (the proportion of absences falsely predicted, also calculated as 1 - specificity) at various threshold settings. AUC\u003csub\u003eROC\u003c/sub\u003e ranges from 0 to 1, where values above 0.9 indicate excellent performance, values between 0.7 and 0.9 indicate good performance, and a value of 0.5 or below indicates no better discrimination than a random classification.\u003c/p\u003e \u003cp\u003eThe area under the precision-recall curve (AUC\u003csub\u003ePR\u003c/sub\u003e) measures the ability to capture true presences and does not include false positives or absences in its calculation. AUC\u003csub\u003ePR\u003c/sub\u003e is recommended when the user is more interested in an accurate prediction of the presences or when the number of absences is much larger than presences\u003csup\u003e\u003cspan citationid=\"CR64\" class=\"CitationRef\"\u003e64\u003c/span\u003e\u003c/sup\u003e. Like the ROC curve, the precision-recall curve calculates precision (the proportion between correctly predicted presences and all predicted presences) and recall across multiple thresholds. The baseline in the AUC\u003csub\u003ePR\u003c/sub\u003e depends on the prevalence in the testing data \u003csup\u003e\u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e65\u003c/span\u003e\u003c/sup\u003e, making it difficult to compare between species or studies. As such, the precision-recall curve is plotted in a new coordinate system, called the precision-recall gain curve, as described by Flach and Kull\u003csup\u003e\u003cspan citationid=\"CR66\" class=\"CitationRef\"\u003e66\u003c/span\u003e\u003c/sup\u003e. AUC\u003csub\u003ePRG\u003c/sub\u003e values from 0 to 1 indicate a performance better than random, where 1 is a perfect discrimination model. Negative values indicate predictions worse than random.\u003c/p\u003e \u003cp\u003eSpecificity is the proportion of absences correctly predicted. Both sensitivity and specificity are considered high with values above 0.8\u003csup\u003e63\u003c/sup\u003e. TSS normalizes the overall accuracy and accounts for both omission and commission errors\u003csup\u003e\u003cspan citationid=\"CR67\" class=\"CitationRef\"\u003e67\u003c/span\u003e\u003c/sup\u003e. TSS is calculated as sensitivity plus specificity minus one and ranges from \u0026minus;\u0026thinsp;1 to 1, where +\u0026thinsp;1 indicates perfect agreement and values of zero or less indicate a performance no better than random. TSS is also considered less sensitive to prevalence compared to Cohen\u0026rsquo;s Kappa statistic.\u003c/p\u003e \u003cp\u003eBecause we used presence-absence data to train the models, we also produced calibration plots to evaluate the agreement between predicted probabilities of occurrence and observation of presence and absence. If the plotted tend is lined up close to the diagonal, we conclude that the model is well-calibrated. Calibration plots were drawn using the scripts in Phillips and Elith\u003csup\u003e\u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e68\u003c/span\u003e\u003c/sup\u003e. Statistical significance of differences between the models for the five statistics were tested using Friedman\u0026rsquo;s Aligned Rank test (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05)\u003csup\u003e69\u003c/sup\u003e, using the R package \u003cem\u003escmamp\u003c/em\u003e v0.2.55\u003csup\u003e70\u003c/sup\u003e. We also applied the post hoc test to conduct pairwise comparisons for each metric, using the Shaffer correction.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003eFrom the 1193 records (grid cells) that were observed by \u003cem\u003eHyBIS\u003c/em\u003e in Area 1, 231 had the presence of at least one \u003cem\u003eSarostegia oculata\u003c/em\u003e. In Area 2, 29 out of 255 records were observed with the presence of \u003cem\u003eS. oculata\u003c/em\u003e by \u003cem\u003eShinkai\u003c/em\u003e. The hexactinellid showed a higher abundance near the rift walls in both areas and was found only on hard substrates, of which the majority had a ferromanganese crust.\u003c/p\u003e \u003cp\u003eThe performance of the models was good for each metric calculated, for both validation and test data. AUC\u003csub\u003eROC\u003c/sub\u003e (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea) values were \u0026gt;\u0026thinsp;0.9, except for ANN which performed slightly worse. All models had good AUC\u003csub\u003ePRG\u003c/sub\u003e scores\u0026thinsp;\u0026gt;\u0026thinsp;0.8 (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eb) and the RF, BRT, MaxEnt, GAM, and Ensemble models had nearly perfect values close to one for the test data, indicating a good discrimination ability to detect presence records. Similarly, all models showed high values for Sensitivity (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ec) and Specificity (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ed) for the validation data, suggesting a high proportion of observed presences and absences correctly predicted, respectively. Models RF, BRT, MaxEnt, and Ensemble had a higher sensitivity for the test data compared to the validation data. In contrast, GAM and ANN models had worse performance. Models had high specificity for test data as well, although the RF model was slightly worse. For the TSS metric, the ANN model had the overall worst performance, while models BRT, MaxEnt, and Ensemble had the highest values for test data (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ee). Friedman\u0026rsquo;s Aligned Rank test showed a significant difference between models for the five statistics. The post-hoc test showed that the most significant differences were found when comparing GAM and ANN with other models (Appendix D). The threshold values used to discriminate between presence and absence were 0.248, 0.126, 0.304, 0.121, 0.187, and 0.225 for the RF, BRT, MaxEnt, ANN, GAM, and Ensemble models, respectively.\u003c/p\u003e\u003cp\u003eThe calibration plots show that the true probability of presence compared to the predicted presence of the five models are badly calibrated (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e). The ideal curve (dotted line) is below the lower confidence interval of the fitted calibration curve, indicating that the true probability of presence is much larger than the estimate given by the models. Only models RF (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ea) and MaxEnt (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ec) had the ideal curves above the higher confidence interval for low probability values. These results indicate models have a high discrimination power, i.e. the ability of a model to correctly distinguish between occupied and unoccupied sites, but the model output should not be interpreted as estimates of \u003cem\u003econditional\u003c/em\u003e probability of presence.\u003c/p\u003e \u003cp\u003eGenerally, all models predicted a suitable habitat along with the full extent of both NE and SW rift borders in Area 1, and the NE rift border in Area 2. For Area 1 (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e), RF model predicted the distribution to be more extensive, with a relatively higher likelihood on the bottom of the rift compared to the middle of the plateaus. Other models had a relatively low (\u0026lt;\u0026thinsp;0.1) likelihood on the bottom of the rift, on top of the plateaus, on the south canyon, and the lower plane regions at southwest. All models predicted a larger area with a high likelihood around the north end of the NE rift border. However, only models RF and ANN extended this area throughout the small terrace at the NE plateau. Models predicted high suitability nearby the area between the east side of the canyon and the SW plateau. The predicted likelihood extended to the other side of the canyon for the MaxEnt model, and even further around the inner SW plateau for models RF and BRT. The south slide of the SW plateau showed high predicted suitability as well, except for the GAM model. For Area 2 (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e), models predicted a high likelihood near the slope of the NE rift border, between 700 and 1000 m. All models, except for RF, showed low (\u0026lt;\u0026thinsp;0.1) suitability at the rift bottom and the top to the plateau.\u003c/p\u003e \u003cp\u003eThe predicted suitable habitat by the ensemble model reflected the average of all five models accordingly. This predicted distribution reflected environmental variables included in the model, namely depth and slope. The region with bottom depths between 700 and 1000 m that had nearby slopes\u0026thinsp;\u0026gt;\u0026thinsp;20 degrees contained a continuous band of high prediction of suitable habitat. The spatial patterns of low-modeled uncertainty corresponded to the main areas predicted as highly suitable on the rift borders in both Area 1 and Area 2 (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e), together with the majority of the SW plateau and around the inner SW plateau. Regions of high uncertainty were obtained for the rift bottom, the canyon, the lower plane regions, and in some areas on top of the plateaus.\u003c/p\u003e \u003cp\u003eThe importance of the environmental variables varied across the modeling algorithms, but depth, fine BPI, and northness usually had a high influence across all models (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). For the RF model, the variables showed a low importance index (\u0026lt;\u0026thinsp;0.1), indicating that this model uses all variables to predict the presence of \u003cem\u003eS. oculata\u003c/em\u003e, and changing a single variable has little effect on its output. Only depth and fine BPI had higher importance relative to other variables. For BRT, MaxEnt, and GAM models, the variables depth, fine BPI, northness, and curvature had a high influence in the likelihood of \u003cem\u003eS. oculata\u003c/em\u003e. However, depth had a higher influence for BRT and GAM compared to MaxEnt, and northness had a higher influence for MaxEnt and GAM compared to BRT models. For the ANN model, most variables showed a high influence in its output, except for broad BPI and eastness.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eMean index of the importance of each predictor variable across 100 permutations in the training dataset, for the Random Forest (RF), Boosted Regression Trees (BRT), MaxEnt, Generalized Additive Models (GAM), and Artificial Neural Networks (ANN) models.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVariables\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRF\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBRT\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMaxEnt\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eGAM\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003edepth\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.09\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.388\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.113\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.572\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.553\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eslope\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.035\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.009\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.034\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.051\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.183\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ebroad BPI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.016\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.002\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.066\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003efine BPI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.072\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.296\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.441\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.588\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.44\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003erift distance\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.049\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.008\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.004\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.349\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003enorthness\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.044\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.112\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.435\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.762\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.338\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eeastness\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.031\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.003\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.007\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.081\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003erugosity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.036\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.011\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.008\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.000\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.19\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ecurvature\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.037\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.046\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.106\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.094\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.304\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eIn general, models showed similar response patterns across the gradient of each environmental variable. Depth (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003ea): MaxEnt, GAM, and ANN models showed a low response at higher depths below ~\u0026thinsp;1000 m and shallower waters above ~\u0026thinsp;700 m, but showed a peak in the predicted likelihood of \u003cem\u003eS. oculata\u003c/em\u003e within ~\u0026thinsp;700\u0026ndash;1000 m. For RF and BRT models, the response remained high from deeper sites until ~\u0026thinsp;800 m where it reached a peak, and then the response was low at shallow depths. Slope (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003eb): models had a low response at flat sites, which increased at steeper slopes. The biggest variation in the response for slope was found in the ANN model. Broad BPI (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003ec): it had a high response around 70 for RF, BRT, and MaxEnt models, but GAM and ANN predicted likelihood was higher at lower broad BPI values (\u0026lt;\u0026thinsp;0), and lower at broad BPI\u0026thinsp;\u0026gt;\u0026thinsp;100. Fine BPI (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003ed): RF, BRT, and ANN showed a lower response at fine BPI\u0026thinsp;\u0026lt;\u0026thinsp;0, and a higher response at values\u0026thinsp;\u0026gt;\u0026thinsp;0, with a larger variation produced by the ANN model. MaxEnt response was higher with negative fine BPI, while GAM was unresponsive for this variable, characterized by a flat horizontal line in the plot. Rift distance (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003ee): All models except ANN showed a similar pattern for the rift distance. A peak in response near the rift (\u0026lt;\u0026thinsp;2000 m) and from 4500 to 7000 m, along with low response between 2000 and 4500 m and regions more distant than 7000 m. The ANN model was different, with a high response near the rift, lowering constantly as it moves away until 10,000 m. The MaxEnt, GAM, and ANN had very low predicted outputs when far away from the rift (\u0026gt;\u0026thinsp;20,000 m). Northness (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003ef): peaks in response were produced at sides facing north and south in models RF, BRT, and MaxEnt. GAM generated a slightly higher response in sites facing north than south. ANN, instead, generated a higher response in sites facing south than north. Eastness (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003eg): models predicted a slightly higher response on sites facing either east or west. However, the GAM was unresponsive for this variable as well. Rugosity (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003eh): models had a low response at sites with low rugosity, that increased at areas with higher values. Curvature (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003ei): only ANN model had a large variation in response for curvature, with a peak in sites with curvature close to zero. RF and BRT, instead, showed a smaller response in these sites. MaxEnt and GAM were unresponsive for this variable.\u003c/p\u003e "},{"header":"Discussion","content":"\u003cp\u003eThe Rio Grande Rise contains diverse benthic communities which include several VME indicator species such as sponges, scleractinians, octocorals, and black corals (Corr\u0026ecirc;a et al., unpublished data). Although RGR is located in oligotrophic waters, this diversity is likely related to strong currents and complex geomorphology with multiple habitats\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e. Our study provides high-resolution distribution maps (DM) for an important VME indicator species recorded in RGR, the branched hexactinellid \u003cem\u003eSarostegia oculata\u003c/em\u003e. This sponge was the most abundant species found in the footage near the rift\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e, closely associated with Fe-Mn crust substrates. A few fragments of \u003cem\u003eS. oculata\u003c/em\u003e sampled by dredges\u003csup\u003e\u003cspan citationid=\"CR71\" class=\"CitationRef\"\u003e71\u003c/span\u003e\u003c/sup\u003e revealed 18 species associated with this sponge, the most notorious being a zoanthid and an annelid, \u003cem\u003eThoracactis topsenti\u003c/em\u003e and \u003cem\u003eHermadion fauveli cf.\u003c/em\u003e Gravier\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e respectively, which also emphasize the importance \u003cem\u003eS. oculata\u003c/em\u003e has in this environment. Our work is the first to create a species distribution model of \u003cem\u003eS. oculata\u003c/em\u003e, and the first to produce high-resolution DMs at RGR, with potential use for deep-sea conservation and management in this area.\u003c/p\u003e \u003cp\u003eDeep-sea sponges have a high conservation and management significance because of their low resilience\u003csup\u003e\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e\u003c/sup\u003e. These species and their associated fauna are particularly vulnerable to anthropogenic impacts such as fishing and deep-sea mining due to their slow growth rates and low or unpredictable recruitment\u003csup\u003e\u003cspan citationid=\"CR72\" class=\"CitationRef\"\u003e72\u003c/span\u003e\u003c/sup\u003e, proving to have a very slow or nonexistent recovery\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. For being suspension filters, they will likely suffer from sediment loads caused by deep-sea mining\u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e. The water canal system of such sponges are at risk of becoming clogged\u003csup\u003e\u003cspan citationid=\"CR73\" class=\"CitationRef\"\u003e73\u003c/span\u003e\u003c/sup\u003e by high loads of suspended particulate matter. These mining plumes can be of low or no nutritional value, further threatening the maintenance of these animals. Hence, \u003cem\u003eS. oculata\u003c/em\u003e not only will be primarily affected by crust removal, but also in the area surrounding mining operations\u003c/p\u003e \u003cp\u003eHigh-resolution distribution models of benthic species, using seafloor camera imagery and bathymetric data obtained from multibeam surveys, have proven useful in the last two decades. Similar models for VME indicator taxa have been developed for conservation and management of areas of interest in the deep sea\u003csup\u003e15,43,74\u0026minus;80\u003c/sup\u003e. Such models are able to provide expert advice on the occurrence of VME indicator taxa in efforts to limit anthropogenic threats from future marine resource exploitation\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eSpecies distribution model\u003c/h2\u003e \u003cp\u003eThe five models trained in this study predicted a high likelihood of \u003cem\u003eS. oculata\u003c/em\u003e along with all extent of both rift borders in the multibeam survey. These findings reinforce the idea that \u003cem\u003eS. oculata\u003c/em\u003e distribution is highly related to the NE-SW rift border that runs throughout RGR, as suggested by Hajdu et al.\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e found in our study. This area is known to be an optimal place for Fe-Mn crusts formation\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e, highly supported by the first claims by CPRM to ISA, which consists of 150 blocks of 20 km\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e each, the maximum allowed by Regulations on Prospecting and Exploration for Fe-Mn crusts\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. The majority of the blocks are situated along the plateaus of WRGR, on both sides of the rift\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e, where CPRM likely prospected areas of high Fe-Mn deposits before submitting the claim. In addition, \u003cem\u003eS. oculata\u003c/em\u003e is somehow more closely associated with Fe-Mn crusts pavements in the plateaus near the rift and along the rift walls compared to other substrates\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e. Our study reinforces this relationship between \u003cem\u003eS. oculata\u003c/em\u003e, Fe-Mn crusts, and the rift at RGR. A second notable suitable area was predicted between the southwestern of the SW plateau and on the top northeastern canyon side. However, we have no footage of this area that can confirm or not the presence of \u003cem\u003eS. oculata\u003c/em\u003e there.\u003c/p\u003e \u003cp\u003eSampled Fe-Mn crusts near the rift suggest that they were eroded by the strong currents that impacted the RGR plateau\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e, and bottom currents of more than 0.2 m/s may be capable of eroding Fe-Mn crust surfaces\u003csup\u003e\u003cspan citationid=\"CR81\" class=\"CitationRef\"\u003e81\u003c/span\u003e\u003c/sup\u003e. These currents may provide a high food supply for the development of \u003cem\u003eS. oculata\u003c/em\u003e, and can be an important key factor in the development of this organism. Unfortunately, we do not have a hydrodynamic model in the study area that could provide currents velocity and direction, which could be used as predictors in the distribution models. Future studies should focus on obtaining and applying such variables in the models, as they likely play an important role in the distribution of species in RGR. The complex outlines of the rift walls may generate vortices\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e and areas of slope may facilitate the propagation of internal tidal waves\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan citationid=\"CR82\" class=\"CitationRef\"\u003e82\u003c/span\u003e\u003c/sup\u003e, which can cause resuspensions favoring greater development of sponges. This may explain why the response for rift distance and fine BPI had such importance to explain the distribution of \u003cem\u003eSarostegia\u003c/em\u003e. Steeper slopes also predicted higher response in the models, although this variable did not output great importance in the bootstrapped correlation test. The aspect of the seabed, in terms of northness, predicted the highest likelihood on the northern and southern slopes, which may correspond with the direction of currents in RGR as well.\u003c/p\u003e \u003cp\u003eAnother variable considered important in our results was depth. The prediction of \u003cem\u003eS. oculata\u003c/em\u003e was restricted between ~\u0026thinsp;700\u0026ndash;1000 m, which corresponds to the plateaus and upper rift walls. Depth is usually an important variable that predicts the occurrence of VME indicator species and can act as a surrogate for other important variables such as temperature, salinity, oxygen, nutrients, water masses, exported surface production, and aragonite saturation\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e,\u003cspan citationid=\"CR83\" class=\"CitationRef\"\u003e83\u003c/span\u003e,\u003cspan citationid=\"CR84\" class=\"CitationRef\"\u003e84\u003c/span\u003e\u003c/sup\u003e. \u003cem\u003eS. oculata\u003c/em\u003e was collected/observed at 598\u0026ndash;1311 m depth in Cape Verde\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e, at 745 m in east of Miami Terrace, south-west of Bimini\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e, at 900\u0026ndash;790 m in the Vit\u0026oacute;ria Trindade seamounts chain\u003csup\u003e\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e, at 738‒1040 m in RGR by Hajdu el al.\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e, and in our study, from 681 to 1203 m. The similarity in depth ranges in the Atlantic between these records suggests that this sponge occurs in bathymetrically constrained bands. However, it is important to note that this bathymetric range alone is not sufficient to predict the distribution, as the inner SW plateau is within this range, but the models had a low likelihood in this region.\u003c/p\u003e \u003cp\u003eThere are a few global distribution models of VME indicator taxa, namely scleractinians, octocorals, and black corals, that predicted their occurrence on RGR\u003csup\u003e19,85\u0026minus;87\u003c/sup\u003e. Similar models were built using data exclusively from the Brazilian continental margin by Barbosa et al.\u003csup\u003e\u003cspan citationid=\"CR88\" class=\"CitationRef\"\u003e88\u003c/span\u003e\u003c/sup\u003e, who addressed the distribution of scleractinians and octocorals in RGR. Overall, they predicted a high suitability of VME indicator taxa on the plateau of RGR, especially in regions near the rift, somewhat similar to the predicted output of our models. However, these models use coarse resolution data (\u0026thinsp;\u0026gt;\u0026thinsp;=\u0026thinsp;0.0083\u0026deg;, ~\u0026thinsp;1 km) as environmental predictors, and none had access to biological records from RGR to train or evaluate the model. High-resolution bathymetric data can better represent seabed physiographic features and improve regional and local suitability models\u003csup\u003e\u003cspan citationid=\"CR89\" class=\"CitationRef\"\u003e89\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003eImplications for Spatial Management and Conservation\u003c/h2\u003e \u003cp\u003eThere is a growing number of examples in the literature that demonstrate the potential of using species distribution models to predict the occurrence of deep-sea sponges for their conservation and management from impacts caused by anthropogenic activities such as bottom trawling and mining\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. They provide fundamental ecosystem functions, and even in low densities, hexactinellid sponges may create a suitable substrate for colonization and development of several invertebrate taxa, serving as island habitats on the deep-sea floor\u003csup\u003e\u003cspan citationid=\"CR90\" class=\"CitationRef\"\u003e90\u003c/span\u003e\u003c/sup\u003e. Thus, increasing knowledge of their distribution and environmental conditions responsible for their formation and persistence are key factors to ensure the protection of the marine environment, especially from harmful effects resulting from human activities.\u003c/p\u003e \u003cp\u003eThe models performance were excellent or good in most cases, correctly identifying 85.4% of sites for validation and 88.8% for testing. Our study uses a completely different dataset to test the models (Area 2), independent from the data used to train and validate the model (Area 1). This method was intended to simulate how models would perform in case they were used to predict the distribution of \u003cem\u003eS. oculata\u003c/em\u003e in a neighboring, unexplored region compared to the original area where models were built. The high performance models had in the test dataset suggests they could be used to predict the distribution of \u003cem\u003eS. oculata\u003c/em\u003e in unsurveyed areas, at least to some extent. In addition, our study suggests an overlap in the potential distribution of Fe-Mn crusts and \u003cem\u003eS. oculata\u003c/em\u003e, which should be addressed in the management of mining to minimize impacts in this community and diminish the loss of biodiversity in RGR.\u003c/p\u003e \u003cp\u003eThere are a few limitations that should be considered during the modeling approach. Species occurrence in different areas can be difficult to model, as SDMs can fail to account for biotic processes, such as competition and predator-prey interactions, and due to shortfalls in the available data, such as sampling bias and lack of key drivers of habitat suitability\u003csup\u003e\u003cspan citationid=\"CR91\" class=\"CitationRef\"\u003e91\u003c/span\u003e\u003c/sup\u003e. The absolute uncertainty of the model prediction is unknown, but the bootstrap procedures provided a measure of internal consistency across the models\u003csup\u003e\u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e\u003c/sup\u003e. Uncertainty maps are a key resource when applying predictions of distribution models to management measures. The uncertainty of the ensemble model was the lowest near the rift border on both sides, and higher below 1,000 m depth or in some areas on top of the plateaus. Predictions in the rift border had more confidence, which could be explained by a higher sampling effort in this region. Thus, it is advised to use our models carefully to predict areas away from the rift. Increase in model performance and reduction in uncertainty may be achieved with (a) a regularly spaced sampling regime that covers the entirety of the environmental conditions observed in the region of interest\u003csup\u003e\u003cspan citationid=\"CR92\" class=\"CitationRef\"\u003e92\u003c/span\u003e\u003c/sup\u003e, (b) inclusion of key environmental drivers for sponges, such as sediment type, current regimes, and nutrients\u003csup\u003e\u003cspan citationid=\"CR76\" class=\"CitationRef\"\u003e76\u003c/span\u003e,\u003cspan citationid=\"CR77\" class=\"CitationRef\"\u003e77\u003c/span\u003e\u003c/sup\u003e, (c) and broader high-resolution surveys, that are still scarce for the area\u003csup\u003e\u003cspan citationid=\"CR71\" class=\"CitationRef\"\u003e71\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eRio Grange Rise had been treated as an area beyond national jurisdictions (ABNJ) and under regulations of the International Seabed Authority (ISA). In December 2018, Brazil presented a partial revised submission to the Commission on Limits of The Continental Shelf (CLCS), which includes RGR as part of its continental margin\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. If approved, RGR will become a region under the jurisdiction and sovereign rights of Brazil, along with its mineral resources. This creates unforeseen implications for the management of RGR, as the regulations that will govern this area are still uncertain. Nevertheless, this issue should not undermine research that produces data and results which could be used to inform management and conservation planning in the area.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003eData availability\u003c/p\u003e\n\u003cp\u003eThe data that support the findings of this study are available from the corresponding author upon reasonable request.\u003c/p\u003e\n\u003cp\u003eCode availability\u003c/p\u003e\n\u003cp\u003eCode is available at the GitHub repositories https://github.com/ correapvf/SDM-RGR-2022 and https://github.com/correapvf/caretSDM.\u003c/p\u003e\n\u003cp\u003eAcknowledgements\u003c/p\u003e\n\u003cp\u003eThis study was funded by the Funda\u0026ccedil;\u0026atilde;o de Amparo a Pesquisa do Estado de S\u0026atilde;o Paulo (FAPESP, BR) grant number 2014/50820-7 and the Natural Environment Research Council (NERC, UK), as part of the Brazil-UK joint project \u0026ldquo;Marine ferromanganese deposits: a major resource of E-tech elements\u0026rdquo;. PVFC was funded by FAPESP grant number 2017/11884-8 and PYGS was funded by CNPq grant number 301554/2019-6. We are grateful to the Captain and Crew Members of the N/Oc. AlphaCrucis and \u003cem\u003eRSS Discovery\u003c/em\u003e for their professionalism and dedication to data acquisition.\u003c/p\u003e\n\u003cp\u003eAuthor contributions\u003c/p\u003e\n\u003cp\u003eThe authors contributed equally for this work.\u003c/p\u003e\n\u003cp\u003eCompeting interests\u003c/p\u003e\n\u003cp\u003eThe authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eBeazley, L. \u003cem\u003eet al.\u003c/em\u003e Predicted distribution of the glass sponge Vazella pourtalesi on the Scotian Shelf and its persistence in the face of climatic variability. PLOS ONE \u003cb\u003e13\u003c/b\u003e, e0205505 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHogg, M. M. \u003cem\u003eet al. Deep-sea Sponge Grounds: Reservoirs of Biodiversity\u003c/em\u003e. (UNEP-WCMC, 2010).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMaldonado, M. \u003cem\u003eet al.\u003c/em\u003e Sponge Grounds as Key Marine Habitats: A Synthetic Review of Types, Structure, Functional Roles, and Conservation Concerns. in \u003cem\u003eMarine Animal Forests: The Ecology of Benthic Biodiversity Hotspots\u003c/em\u003e (eds. Rossi, S., Bramanti, L., Gori, A. \u0026amp; Orejas Saco del Valle, C.) 1\u0026ndash;39 (Springer International Publishing, 2016). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/978-3-319-17001-5_24-1\u003c/span\u003e\u003cspan address=\"10.1007/978-3-319-17001-5_24-1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCook, S. E., Conway, K. W. \u0026amp; Burd, B. Status of the glass sponge reefs in the Georgia Basin. Marine Environmental Research \u003cb\u003e66\u003c/b\u003e, S80\u0026ndash;S86 (2008).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKlitgaard, A. B. The fauna associated with outer shelf and upper slope sponges (Porifera, Demospongiae) at the Faroe Islands, northeastern Atlantic. Sarsia \u003cb\u003e80\u003c/b\u003e, 1\u0026ndash;22 (1995).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBeazley, L., Kenchington, E., Yashayaev, I. \u0026amp; Murillo, F. J. Drivers of epibenthic megafaunal composition in the sponge grounds of the Sackville Spur, northwest Atlantic. Deep Sea Research Part I: Oceanographic Research Papers \u003cb\u003e98\u003c/b\u003e, 102\u0026ndash;114 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBett, B. J. \u0026amp; Rice, A. L. The influenceof hexactinellid sponge (Pheronema carpenteri) spicules on the patchy distribution of macrobenthos in the porcupine seabight (bathyal ne atlantic). Ophelia \u003cb\u003e36\u003c/b\u003e, 217\u0026ndash;226 (1992).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBo, M. \u003cem\u003eet al.\u003c/em\u003e Role of deep sponge grounds in the Mediterranean Sea: a case study in southern Italy. in \u003cem\u003eAncient Animals, New Challenges: Developments in Sponge Research\u003c/em\u003e (eds. Maldonado, M., Turon, X., Becerro, M. \u0026amp; Jes\u0026uacute;s Uriz, M.) 163\u0026ndash;177 (Springer Netherlands, 2012). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/978-94-007-4688-6_14\u003c/span\u003e\u003cspan address=\"10.1007/978-94-007-4688-6_14\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMarliave, J. B., Conway, K. W., Gibbs, D. M., Lamb, A. \u0026amp; Gibbs, C. Biodiversity and rockfish recruitment in sponge gardens and bioherms of southern British Columbia, Canada. Mar Biol \u003cb\u003e156\u003c/b\u003e, 2247\u0026ndash;2254 (2009).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCathalot, C. \u003cem\u003eet al.\u003c/em\u003e Cold-water coral reefs and adjacent sponge grounds: hotspots of benthic respiration and organic carbon cycling in the deep sea. Frontiers in Marine Science \u003cb\u003e2\u003c/b\u003e, (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMaldonado, M., Navarro, L., Grasa, A., Gonzalez, A. \u0026amp; Vaquerizo, I. Silicon uptake by sponges: a twist to understanding nutrient cycling on continental margins. Sci Rep \u003cb\u003e1\u003c/b\u003e, 30 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePile, A. J. \u0026amp; Young, C. M. The natural diet of a hexactinellid sponge: Benthic\u0026ndash;pelagic coupling in a deep-sea microbial food web. Deep Sea Research Part I: Oceanographic Research Papers \u003cb\u003e53\u003c/b\u003e, 1148\u0026ndash;1156 (2006).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRossi, S. The destruction of the \u0026lsquo;animal forests\u0026rsquo; in the oceans: Towards an over-simplification of the benthic ecosystems. Ocean \u0026amp; Coastal Management \u003cb\u003e84\u003c/b\u003e, 77\u0026ndash;85 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRamirez-Llodra, E. \u003cem\u003eet al.\u003c/em\u003e Man and the Last Great Wilderness: Human Impact on the Deep Sea. PLOS ONE \u003cb\u003e6\u003c/b\u003e, e22588 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRowden, A. A. \u003cem\u003eet al.\u003c/em\u003e High-Resolution Habitat Suitability Models for the Conservation and Management of Vulnerable Marine Ecosystems on the Louisville Seamount Chain, South Pacific Ocean. Front. Mar. Sci. \u003cb\u003e4\u003c/b\u003e, 335 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e\u003cem\u003eInternational Guidelines for the Management of Deep-Sea Fisheries in the High-Seas\u003c/em\u003e. (Food and Agriculture Organization of the United Nations, 2009).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMohriak, W. Genesis and evolution of the South Atlantic volcanic islands offshore Brazil. Geo-Mar Lett \u003cb\u003e40\u003c/b\u003e, 1\u0026ndash;33 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCavalcanti, J. A. D., Santos, R. V., Lacasse, C. M., Rojas, J. N. L. \u0026amp; Nobrega, M. Potential Mineral Resources of Phosphates and Trace Elements on the Rio Grande Rise, South Atlantic Ocean. \u003cem\u003e44th Underwater Mining Conference\u003c/em\u003e (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMontserrat, F. \u003cem\u003eet al.\u003c/em\u003e Deep-sea mining on the Rio Grande Rise (Southwestern Atlantic): A review on environmental baseline, ecosystem services and potential impacts. Deep Sea Research Part I: Oceanographic Research Papers \u003cb\u003e145\u003c/b\u003e, 31\u0026ndash;58 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eda Silva, A. P. Brazil advances over the Area: The inclusion of the Rio Grande Rise within the Brazilian outer continental shelf and its consequences for other states and for the common heritage of mankind. Marine Policy \u003cb\u003e125\u003c/b\u003e, 104399 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBenites, M., Hein, J. R., Mizell, K., Blackburn, T. \u0026amp; Jovane, L. Genesis and Evolution of Ferromanganese Crusts from the Summit of Rio Grande Rise, Southwest Atlantic Ocean. Minerals \u003cb\u003e10\u003c/b\u003e, 349 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSousa, I. M. C. \u003cem\u003eet al.\u003c/em\u003e Mineralogy and chemical composition of ferromanganese crusts from the Cruzeiro do Sul Lineament - Rio Grande Rise, South Atlantic. Journal of South American Earth Sciences \u003cb\u003e108\u003c/b\u003e, 103207 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChristiansen, B., Denda, A. \u0026amp; Christiansen, S. Potential effects of deep seabed mining on pelagic and benthopelagic biota. Marine Policy \u003cb\u003e114\u003c/b\u003e, 103442 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLevin, L. A. \u003cem\u003eet al.\u003c/em\u003e Defining \u0026ldquo;serious harm\u0026rdquo; to the marine environment in the context of deep-seabed mining. Marine Policy \u003cb\u003e74\u003c/b\u003e, 245\u0026ndash;259 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchmidt, C. W. Going Deep: Cautious Steps toward Seabed Mining. Environmental Health Perspectives \u003cb\u003e123\u003c/b\u003e, A234\u0026ndash;A241 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDunn, D. C. \u003cem\u003eet al.\u003c/em\u003e A strategy for the conservation of biodiversity on mid-ocean ridges from deep-sea mining. Science Advances \u003cb\u003e4\u003c/b\u003e, eaar4313 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJones, D. O. B., Ardron, J. A., Cola\u0026ccedil;o, A. \u0026amp; Durden, J. M. Environmental considerations for impact and preservation reference zones for deep-sea polymetallic nodule mining. Marine Policy \u003cb\u003e118\u003c/b\u003e, (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLodge, M. W. \u0026amp; Verlaan, P. A. Deep-Sea Mining: International Regulatory Challenges and Responses. Elements \u003cb\u003e14\u003c/b\u003e, 331\u0026ndash;336 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMiller, K. A., Thompson, K. F., Johnston, P. \u0026amp; Santillo, D. An Overview of Seabed Mining Including the Current State of Development, Environmental Impacts, and Knowledge Gaps. Front. Mar. Sci. \u003cb\u003e4\u003c/b\u003e, 418 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGravier, Ch. Note sur une Actinie (\u003cem\u003eThoracactis\u003c/em\u003e n. g., \u003cem\u003eTopsenti\u003c/em\u003e n. sp.) et un Ann\u0026eacute;lide Polych\u0026egrave;te (\u003cem\u003eHermadion Faiweli\u003c/em\u003e n. sp.), commensaux d\u0026rsquo;une \u0026Eacute;ponge siliceuse (\u003cem\u003eSarostegia oculata\u003c/em\u003e Topsent). \u003cem\u003eBulletin du Mus\u0026eacute;e Oc\u0026eacute;anographique de Monaco\u003c/em\u003e \u003cb\u003e344\u003c/b\u003e, 1\u0026ndash;20 (1918).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHajdu, E., Castello-Branco, C., Lopes, D. A., Sumida, P. Y. G. \u0026amp; Perez, J. A. A. Deep-sea dives reveal an unexpected hexactinellid sponge garden on the Rio Grande Rise (SW Atlantic). A mimicking habitat? Deep Sea Research Part II: Topical Studies in Oceanography \u003cb\u003e146\u003c/b\u003e, 93\u0026ndash;100 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTopsent, E. \u003cem\u003eSarostegia oculata\u003c/em\u003e, Hexactinellide nouvelle des \u0026icirc;les du Cap-Vert. Bulletin du Mus\u0026eacute;e Oc\u0026eacute;anographique de Monaco \u003cb\u003e10\u003c/b\u003e, 1\u0026ndash;8 (1904).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDohrmann, M. \u003cem\u003eet al.\u003c/em\u003e Systematics and spicule evolution in dictyonal sponges (Hexactinellida: Sceptrulophora) with description of two new species. Zoological Journal of the Linnean Society \u003cb\u003e163\u003c/b\u003e, 1003\u0026ndash;1025 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTabachnick, K. R., Menshenina, L. L., Lopes, D. A. \u0026amp; Hajdu, E. Two new \u003cem\u003eHyalonema\u003c/em\u003e species (Hyalonematidae: Amphidiscosida) from eastern and south-eastern Brazil, and further Hexactinellida (Porifera) collected from seamounts off south-eastern Brazil by the RV \u0026lsquo;Marion Dufresne\u0026rsquo; MD55 expedition. J. Mar. Biol. Ass. \u003cb\u003e89\u003c/b\u003e, 1243\u0026ndash;1250 (2009).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eReiswig, H. M. Family Farreidae Gray, 1872. in \u003cem\u003eSystema Porifera: a guide to the classification of sponges\u003c/em\u003e (eds. Hooper, J. N. A., Van Soest, R. W. M. \u0026amp; Willenz, P.) 1332\u0026ndash;1340 (Springer, 2002). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/978-1-4615-0747-5_136\u003c/span\u003e\u003cspan address=\"10.1007/978-1-4615-0747-5_136\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHowell, K.-L., Piechaud, N., Downie, A.-L. \u0026amp; Kenny, A. The distribution of deep-sea sponge aggregations in the North Atlantic and implications for their effective spatial management. Deep Sea Research Part I: Oceanographic Research Papers \u003cb\u003e115\u003c/b\u003e, 309\u0026ndash;320 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMarshall, C. E., Glegg, G. A. \u0026amp; Howell, K. L. Species distribution modelling to support marine conservation planning: The next steps. Marine Policy \u003cb\u003e45\u003c/b\u003e, 330\u0026ndash;332 (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eReiss, H. \u003cem\u003eet al.\u003c/em\u003e Benthos distribution modelling and its relevance for marine ecosystem management. ICES Journal of Marine Science \u003cb\u003e72\u003c/b\u003e, 297\u0026ndash;315 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuilhon, M., Montserrat, F. \u0026amp; Turra, A. Recognition of ecosystem-based management principles in key documents of the seabed mining regime: implications and further recommendations. ICES Journal of Marine Science \u003cb\u003e78\u003c/b\u003e, 884\u0026ndash;899 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWedding, L. M. \u003cem\u003eet al.\u003c/em\u003e Managing mining of the deep seabed. Science \u003cb\u003e349\u003c/b\u003e, 144\u0026ndash;145 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWedding, L. M. \u003cem\u003eet al.\u003c/em\u003e From principles to practice: a spatial approach to systematic conservation planning in the deep sea. \u003cem\u003eProceedings of the Royal Society B: Biological Sciences\u003c/em\u003e \u003cb\u003e280\u003c/b\u003e, 20131684\u0026ndash;20131684 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKenchington, E. L. \u003cem\u003eet al.\u003c/em\u003e Use of Species Distribution Modeling in the Deep Sea. \u003cem\u003eCanadian Technical Report of Fisheries and Aquatic Sciences\u003c/em\u003e \u003cb\u003e3296\u003c/b\u003e, 1\u0026ndash;76 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRamiro-S\u0026aacute;nchez, B. \u003cem\u003eet al.\u003c/em\u003e Characterization and Mapping of a Deep-Sea Sponge Ground on the Tropic Seamount (Northeast Tropical Atlantic): Implications for Spatial Management in the High Seas. Frontiers in Marine Science \u003cb\u003e6\u003c/b\u003e, (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWalbridge, S., Slocum, N., Marjean Pobuda \u0026amp; Wright, D. Unified Geomorphological Analysis Workflows with Benthic Terrain Modeler. Geosciences \u003cb\u003e8\u003c/b\u003e, 94 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuisan, A., Thuiller, W. \u0026amp; Zimmermann, N. E. \u003cem\u003eHabitat Suitability and Distribution Models: With Applications in R\u003c/em\u003e. (Cambridge University Press, 2017). doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1017/9781139028271\u003c/span\u003e\u003cspan address=\"10.1017/9781139028271\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNaimi, B., Hamm, N. A. S., Groen, T. A., Skidmore, A. K. \u0026amp; Toxopeus, A. G. Where is positional uncertainty a problem for species distribution modelling? Ecography \u003cb\u003e37\u003c/b\u003e, 191\u0026ndash;203 (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMerow, C. \u003cem\u003eet al.\u003c/em\u003e What do we gain from simplicity versus complexity in species distribution models? Ecography \u003cb\u003e37\u003c/b\u003e, 1267\u0026ndash;1281 (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eR Core Team. \u003cem\u003eR: A Language and Environment for Statistical Computing\u003c/em\u003e. (R Foundation for Statistical Computing, 2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKuhn, M. Building Predictive Models in \u003cem\u003eR\u003c/em\u003e Using the \u003cb\u003ecaret\u003c/b\u003e Package. J. Stat. Soft. \u003cb\u003e28\u003c/b\u003e, (2008).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStrobl, C., Malley, J. \u0026amp; Tutz, G. An introduction to recursive partitioning: Rationale, application, and characteristics of classification and regression trees, bagging, and random forests. Psychological Methods \u003cb\u003e14\u003c/b\u003e, 323\u0026ndash;348 (2009).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eElith, J., Leathwick, J. R. \u0026amp; Hastie, T. A working guide to boosted regression trees. Journal of Animal Ecology \u003cb\u003e77\u003c/b\u003e, 802\u0026ndash;813 (2008).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eElith, J. \u003cem\u003eet al.\u003c/em\u003e A statistical explanation of MaxEnt for ecologists: Statistical explanation of MaxEnt. Diversity and Distributions \u003cb\u003e17\u003c/b\u003e, 43\u0026ndash;57 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePhillips, S. J., Anderson, R. P. \u0026amp; Schapire, R. E. Maximum entropy modeling of species geographic distributions. Ecological Modelling \u003cb\u003e190\u003c/b\u003e, 231\u0026ndash;259 (2006).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRenner, I. W. \u0026amp; Warton, D. I. Equivalence of MAXENT and Poisson Point Process Models for Species Distribution Modeling in Ecology. Biometrics \u003cb\u003e69\u003c/b\u003e, 274\u0026ndash;281 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWood, S. N. Fast stable restricted maximum likelihood and marginal likelihood estimation of semiparametric generalized linear models. Journal of the Royal Statistical Society: Series B (Statistical Methodology) \u003cb\u003e73\u003c/b\u003e, 3\u0026ndash;36 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBasheer, I. A. \u0026amp; Hajmeer, M. Artificial neural networks: fundamentals, computing, design, and application. Journal of Microbiological Methods \u003cb\u003e43\u003c/b\u003e, 3\u0026ndash;31 (2000).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eElith, J., Ferrier, S., Huettmann, F. \u0026amp; Leathwick, J. The evaluation strip: A new and robust method for plotting predicted responses from species distribution models. Ecological Modelling \u003cb\u003e186\u003c/b\u003e, 280\u0026ndash;289 (2005).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThuiller, W., Lafourcade, B., Engler, R. \u0026amp; Ara\u0026uacute;jo, M. B. BIOMOD \u0026ndash; a platform for ensemble forecasting of species distributions. Ecography \u003cb\u003e32\u003c/b\u003e, 369\u0026ndash;373 (2009).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDormann, C. F. \u003cem\u003eet al.\u003c/em\u003e Model averaging in ecology: a review of Bayesian, information-theoretic, and tactical approaches for predictive inference. Ecological Monographs \u003cb\u003e88\u003c/b\u003e, 485\u0026ndash;504 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAra\u0026uacute;jo, M. B. \u0026amp; New, M. Ensemble forecasting of species distributions. Trends in Ecology \u0026amp; Evolution \u003cb\u003e22\u003c/b\u003e, 42\u0026ndash;47 (2007).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAnderson, O. F. \u003cem\u003eet al.\u003c/em\u003e Habitat suitability models for predicting the occurrence of vulnerable marine ecosystems in the seas around New Zealand. Deep Sea Research Part I: Oceanographic Research Papers \u003cb\u003e115\u003c/b\u003e, 265\u0026ndash;292 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFourcade, Y., Besnard, A. G. \u0026amp; Secondi, J. Paintings predict the distribution of species, or the challenge of selecting environmental predictors and evaluation statistics. Global Ecology and Biogeography \u003cb\u003e27\u003c/b\u003e, 245\u0026ndash;256 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu, C., Berry, P. M., Dawson, T. P. \u0026amp; Pearson, R. G. Selecting thresholds of occurrence in the prediction of species distributions. Ecography \u003cb\u003e28\u003c/b\u003e, 385\u0026ndash;393 (2005).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eValavi, R., Guillera-Arroita, G., Lahoz-Monfort, J. J. \u0026amp; Elith, J. Predictive performance of presence-only species distribution models: a benchmark study with reproducible code. Ecological Monographs \u003cb\u003e92\u003c/b\u003e, e01486 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSaito, T. \u0026amp; Rehmsmeier, M. The Precision-Recall Plot Is More Informative than the ROC Plot When Evaluating Binary Classifiers on Imbalanced Datasets. PLOS ONE \u003cb\u003e10\u003c/b\u003e, e0118432 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFlach, P. \u0026amp; Kull, M. Precision-Recall-Gain Curves: PR Analysis Done Right. in \u003cem\u003eAdvances in Neural Information Processing Systems\u003c/em\u003e (eds. Cortes, C., Lawrence, N., Lee, D., Sugiyama, M. \u0026amp; Garnett, R.) vol.\u0026nbsp;28 838\u0026ndash;846 (Massachusetts Institute of Technology (MIT) Press, 2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAllouche, O., Tsoar, A. \u0026amp; Kadmon, R. Assessing the accuracy of species distribution models: prevalence, kappa and the true skill statistic (TSS): Assessing the accuracy of distribution models. Journal of Applied Ecology \u003cb\u003e43\u003c/b\u003e, 1223\u0026ndash;1232 (2006).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePhillips, S. J. \u0026amp; Elith, J. POC plots: calibrating species distribution models with presence-only data. Ecology \u003cb\u003e91\u003c/b\u003e, 2476\u0026ndash;2484 (2010).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGarc\u0026iacute;a, S., Fern\u0026aacute;ndez, A., Luengo, J. \u0026amp; Herrera, F. Advanced nonparametric tests for multiple comparisons in the design of experiments in computational intelligence and data mining: Experimental analysis of power. Information Sciences \u003cb\u003e180\u003c/b\u003e, 2044\u0026ndash;2064 (2010).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCalvo, B. \u0026amp; Santaf\u0026eacute;, G. scmamp: Statistical Comparison of Multiple Algorithms in Multiple Problems. The R Journal \u003cb\u003e8\u003c/b\u003e, 248\u0026ndash;256 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJovane, L. \u003cem\u003eet al.\u003c/em\u003e Multidisciplinary Scientific Cruise to the Rio Grande Rise. Front. Mar. Sci. \u003cb\u003e6\u003c/b\u003e, 252 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGollner, S. \u003cem\u003eet al.\u003c/em\u003e Resilience of benthic deep-sea fauna to mining activities. Marine Environmental Research \u003cb\u003e129\u003c/b\u003e, 76\u0026ndash;101 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAbdul Wahab, M. A., Fromont, J., Gomez, O., Fisher, R. \u0026amp; Jones, R. Comparisons of benthic filter feeder communities before and after a large-scale capital dredging program. Marine Pollution Bulletin \u003cb\u003e122\u003c/b\u003e, 176\u0026ndash;193 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDolan, M. F. J., Grehan, A. J., Guinan, J. C. \u0026amp; Brown, C. Modelling the local distribution of cold-water corals in relation to bathymetric variables: Adding spatial context to deep-sea video data. Deep Sea Research Part I: Oceanographic Research Papers \u003cb\u003e55\u003c/b\u003e, 1564\u0026ndash;1579 (2008).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHowell, K. L., Holt, R., Endrino, I. P. \u0026amp; Stewart, H. When the species is also a habitat: Comparing the predictively modelled distributions of Lophelia pertusa and the reef habitat it forms. Biological Conservation \u003cb\u003e144\u003c/b\u003e, 2656\u0026ndash;2665 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRengstorf, A. M., Mohn, C., Brown, C., Wisz, M. S. \u0026amp; Grehan, A. J. Predicting the distribution of deep-sea vulnerable marine ecosystems using high-resolution data: Considerations and novel approaches. Deep Sea Research Part I: Oceanographic Research Papers \u003cb\u003e93\u003c/b\u003e, 72\u0026ndash;82 (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRengstorf, A. M., Yesson, C., Brown, C. \u0026amp; Grehan, A. J. High-resolution habitat suitability modelling can improve conservation of vulnerable marine ecosystems in the deep sea. Journal of Biogeography \u003cb\u003e40\u003c/b\u003e, 1702\u0026ndash;1714 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRobert, K., Jones, D. O. B., Roberts, J. M. \u0026amp; Huvenne, V. A. I. Improving predictive mapping of deep-water habitats: Considering multiple model outputs and ensemble techniques. Deep Sea Research Part I: Oceanographic Research Papers \u003cb\u003e113\u003c/b\u003e, 80\u0026ndash;89 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRodr\u0026iacute;guez-Basalo, A. \u003cem\u003eet al.\u003c/em\u003e High Resolution Spatial Distribution for the Hexactinellid Sponges Asconema setubalense and Pheronema carpenteri in the Central Cantabrian Sea. Front. Mar. Sci. \u003cb\u003e8\u003c/b\u003e, 612761 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTong, R., Purser, A., Guinan, J. \u0026amp; Unnithan, V. Modeling the habitat suitability for deep-water gorgonian corals based on terrain variables. Ecological Informatics \u003cb\u003e13\u003c/b\u003e, 123\u0026ndash;132 (2013).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYeo, I. A. \u003cem\u003eet al.\u003c/em\u003e Distribution of and hydrographic controls on ferromanganese crusts: Tropic Seamount, Atlantic. Ore Geology Reviews \u003cb\u003e114\u003c/b\u003e, 103131 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKlitgaard, A. B. \u0026amp; Tendal, O. S. Distribution and species composition of mass occurrences of large-sized sponges in the northeast Atlantic. Progress in Oceanography \u003cb\u003e61\u003c/b\u003e, 57\u0026ndash;98 (2004).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeathwick, J. R., Elith, J., Francis, M. P., Hastie, T. \u0026amp; Taylor, P. Variation in demersal fish species richness in the oceans surrounding New Zealand: an analysis using boosted regression trees. Marine Ecology Progress Series \u003cb\u003e321\u003c/b\u003e, 267\u0026ndash;281 (2006).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThresher, R. \u003cem\u003eet al.\u003c/em\u003e Strong Depth-Related Zonation of Megabenthos on a Rocky Continental Margin (\u0026sim;700\u0026ndash;4000 m) off Southern Tasmania, Australia. PLOS ONE \u003cb\u003e9\u003c/b\u003e, e85872 (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDavies, A. J. \u0026amp; Guinotte, J. M. Global Habitat Suitability for Framework-Forming Cold-Water Corals. PLoS ONE \u003cb\u003e6\u003c/b\u003e, e18483 (2011).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYesson, C., Bedford, F., Rogers, A. D. \u0026amp; Taylor, M. L. The global distribution of deep-water Antipatharia habitat. Deep Sea Research Part II: Topical Studies in Oceanography (2015) doi:\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.dsr2.2015.12.004\u003c/span\u003e\u003cspan address=\"10.1016/j.dsr2.2015.12.004\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYesson, C. \u003cem\u003eet al.\u003c/em\u003e Global habitat suitability of cold-water octocorals: Global distribution of deep-sea octocorals. Journal of Biogeography \u003cb\u003e39\u003c/b\u003e, 1278\u0026ndash;1292 (2012).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBarbosa, R. V., Davies, A. J. \u0026amp; Sumida, P. Y. G. Habitat suitability and environmental niche comparison of cold-water coral species along the Brazilian continental margin. Deep Sea Research Part I: Oceanographic Research Papers \u003cb\u003e155\u003c/b\u003e, 103147 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRoss, L. K., Ross, R. E., Stewart, H. A. \u0026amp; Howell, K. L. The Influence of Data Resolution on Predicted Distribution and Estimates of Extent of Current Protection of Three \u0026lsquo;Listed\u0026rsquo; Deep-Sea Habitats. PLoS ONE \u003cb\u003e10\u003c/b\u003e, e0140061 (2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBeaulieu, S. E. Life on glass houses: sponge stalk communities in the deep sea. Marine Biology \u003cb\u003e138\u003c/b\u003e, 803\u0026ndash;817 (2001).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVierod, A. D. T., Guinotte, J. M. \u0026amp; Davies, A. J. Predicting the distribution of vulnerable marine ecosystems in the deep sea using presence-background models. Deep Sea Research Part II: Topical Studies in Oceanography \u003cb\u003e99\u003c/b\u003e, 6\u0026ndash;18 (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHirzel, A. \u0026amp; Guisan, A. Which is the optimal sampling strategy for habitat suitability modelling. Ecological Modelling \u003cb\u003e157\u003c/b\u003e, 331\u0026ndash;341 (2002).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Appendix D","content":"\u003cp\u003eAppendix D is not available with this version.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"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":"nature-portfolio","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"","title":"Nature Portfolio","twitterHandle":"","acdcEnabled":false,"dfaEnabled":false,"editorialSystem":"ejp","reportingPortfolio":"","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-1663413/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-1663413/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe mining of ferromanganese (Fe-Mn) crusts in the deep sea have gained more attention in the last decade due to increased demand for rare earth elements that are critical for low-carbon technologies, which makes exploitation of this resource feasible and profitable. The Rio Grande Rise (RGR) is a distinct feature located in the South Atlantic and has become a region of great commercial and scientific interest because of its potential for mining Fe-Mn crusts. Extraction of Fe-Mn deposits may cause irreversible changes by removal of substrate, creation of sediment plumes, among others. Here, we use species distribution models (SDMs) to predict the occurrence of \u003cem\u003eSarostegia oculata\u003c/em\u003e, a branched hexactinellid that mimics the 3D skeletal framework of actual corals. It is the dominant organism in areas rich in Fe-Mn crusts and has relevant ecological importance in RGR. The models had excellent or good performance statistics and a high discrimination power between presence and absence sites. Our results support the relationship between \u003cem\u003eS. oculata\u003c/em\u003e, the Fe-Mn crusts, and the rift at RGR, and may help to create a management plan and preserve unique marine habitats and biodiversity from the deep sea.\u003c/p\u003e","manuscriptTitle":"Distribution models of the branched hexactinellid Sarostegia oculata in Rio Grande Rise (SW Atlantic)","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2022-06-14 13:25:45","doi":"10.21203/rs.3.rs-1663413/v1","editorialEvents":[],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"communications-earth-and-environment","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"commsenv","sideBox":"Learn more about [Communications Earth and Environment](https://www.nature.com/commsenv/)","snPcode":"","submissionUrl":"","title":"Communications Earth \u0026 Environment","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"Communications Series","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"9c5a5d97-5242-4f59-9faf-373dd1bd6ef5","owner":[],"postedDate":"June 14th, 2022","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2022-10-12T14:52:51+00:00","versionOfRecord":[],"versionCreatedAt":"2022-06-14 13:25:45","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-1663413","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-1663413","identity":"rs-1663413","version":["v1"]},"buildId":"7rjqhiLT3MXkJMwkYKINL","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}