The interpretability of VME model predictions directly depends on the underlying data used for model building, meaning the data-characteristics need to match their application (Guillera-Arroita et al., 2015). provide a clear overview of how data characteristics determine how the predictions from species distribution models should be (but often are not) interpreted. Building on (Guillera-Arroita et al., 2015), Figure 1 illustrates how model interpretation varies depending on the type of estimates yielded by the model, i.e. rank suitability, relative likelihood, true probability of occurrence, or density, which is in turn influenced by the data type. In simple terms, when P_s decreases, our confidence in being able to have reliable Presence-Absence data decreases. Rank suitability estimation provides a ranked order of suitable sites for harbouring a VME, which is not proportional to the actual probability of occurrence, i.e. does not indicate how suitable a site is for harbouring a VME, with P_S unreliable. In contrast to rank suitability, the relative likelihood is proportional to the probability of occurrence P_p. Both rank suitability and relative likelihood can be used to discriminate sites, i.e. distinguish between occupied and unoccupied sites better than random. However, to estimate the relative likelihood is not sufficient for comparing across multiple modelled taxa or habitats and/or regions. In order to estimate the true probability of occurrence for a VME, models need to be fitted on presence-absence data with reliable detectability of taxa P_s (Guillera-Arroita et al., 2015). Further, when abundance data are available, the model can estimate density or percentage cover of indicator taxa.

Fishing gears have generally a low catchability for benthic organisms, especially for small, fragile and brittle VME taxa (Auster et al., 2011; Watling and Auster, 2017). Fishing gears are typically designed to catch or attract motile species, as opposed to providing representative samples of VME indicator taxa: P_S is small and unreliable. Moreover, there is a lack of consistency across fisheries management agencies regarding data collection of VME indicator taxa (Thompson et al., 2016). As a result, taxa may be present (and impacted) by fishing gears, but not all are necessarily caught, brought to the surface and/or reported (Knudby et al., 2013; Watling and Auster, 2017; Du Preez et al., 2020). Therefore, VME data acquired with fishing gear is widely considered as “Presence-only” data (i.e., no “true absence” data, P(0), available) with highly imperfect detection (i.e., P_S << 1). With poor confidence in estimating P_S, we are marginalising over P_S to give a relative measure (see Figure 1 ). Besides fisheries bycatch data, some scientific physical sampling gear have been specifically designed for benthic ecosystems on soft substrates (Lauria et al., 2017 ), such as benthic sleds, yielding “Presence-absence” or “Abundance” data with reliable benthic species detectability P_s (Williams et al., 2015). However, physical sampling gear typically only allows indicator taxa collection, since the identification of habitats generally requires direct observation, e.g. via in situ imagery. Although VME data to date have been largely collected with physical sampling gear, these aforementioned restrictions, both on model predictions and the modelled biotic entity, impact the interpretation against the FAO criteria. Further, because physical sampling methods can permanently impact individuals and assemblages, these methods should be avoided for sampling vulnerable ecosystems (Williams et al., 2015).

Underwater imagery allows sampling the seafloor with a reliable detectability of most taxa P_S (Williams et al., 2015; Long et al., 2020). It is generally considered that “Presence-absence” data can be derived from these images since most VME taxa are sessile or with a limited mobility. This kind of data can be used to create models that are: (i) discriminative (i.e., occupied vs. unoccupied sites), (ii) well-calibrated (i.e., exact probability, not relative), and (iii) comparable among VME entities and/or regions (see Figure 1 ). Importantly, habitat identification and/or accurate VME taxa abundance can be extracted from images (Williams et al., 2015), which is important for addressing the FAO criteria. Underwater imagery allows finer model resolution, e.g. 25 m² in (Rowden et al., 2017). Nevertheless, scientific surveys that utilise imagery are rare in the deep sea which restricts the scale of those datasets compared to the ones derived from fishing vessels. Taxonomic identification from images is also more prone to errors and uncertainties (e.g. taxonomic groupings) than when relying on physical samplings. Additionally, the resolution of environmental predictors needs to match the biological data, and high resolution environmental data has limited availability, which is another limitation to the usability of fine-resolution imagery data (Rengstorf et al., 2012). Overall, the sampling method limitations should be clearly communicated to the model’s users to guarantee a cautious interpretation of the modelling results.

In distribution modelling studies, VMEs have most often been considered as either assemblages dominated by a single indicator taxon or co-dominated by multiple indicator taxa. Within these two categories, different modelling approaches have been proposed, each with different implications for identifying VMEs using the FAO criteria. Figure 2 illustrates the different approaches based on how VMEs have been defined by model-builders.

When VMEs are defined as an assemblage dominated by a single indicator taxon, two approaches are usually employed: either by modelling the distribution of the indicator taxon, as a proxy for the associated assemblage, or by modelling the distribution of the assemblage itself (see Figure 2 , top panel). Most studies model the distribution of the VME indicator taxon instead of the associated assemblage. However, the occurrence of a VME indicator taxon alone may not be sufficient to properly identify a VME (FAO, 2009; Howell et al., 2011; Watling and Auster, 2017), as it does not provide evidence of a habitat with important functional significance. The distribution of a VME assemblage often represents only a small subset of the distribution of its associated dominating indicator taxon, especially when the assemblage requires environmental factors that are more restrictive than for the occurrence of the taxon. For instance, Howell et al. (2011) model and compare the distribution of a cold-water coral (Lophelia pertusa, VME indicator taxon) vs. the reef it forms (assemblage forming a VME habitat), and show that the predicted area of high suitability for the taxon is much broader than for the assemblage (7.17% vs. 0.56% of the studied area). To model VMEs using assemblage data is however limited by the scarce availability of this data type, since it generally requires underwater imagery sampling (see section Connections Between the Type of VME Data and Model Use), and by the lack of consensus around the quantitative definition of VME assemblages, e.g. composition or density of taxa (Rowden et al., 2020).

When VMEs are defined as an assemblage of multiple indicator taxa, there are two components to the modelling approach: defining the assemblage (“Assemble”) and making predictions (“Predict”). These generally occur separately and can be in either order, i.e. “Assemble first then Predict” or “Predict first then Assemble”, as defined in (Ferrier and Guisan, 2006), see Figure 2 bottom panel. The “Assemble first then Predict” aims at predicting the occurrence of predefined benthic assemblages, e.g. circalittoral vertical rocky walls with Caryophyllia cyathus and Corallium rubrum in (Torriente et al., 2019). These assemblages can be defined by: (i) prior knowledge, e.g. biotopes known to occur in the area of interest (Bo et al., 2021), (ii) expert judgement using phylogenetic and functional diversity metrics, or (iii) cluster and ordination analysis (Barrio Froján et al., 2016; Torriente et al., 2019). Importantly, this approach cannot provide insights on environmental requirements of each taxon and can lead to an artificially wide niche description (Yesson et al., 2012; Brewin et al., 2020). The “Predict first then Assemble” approach models and predicts the spatial distribution of each VME taxon before assembling model outputs. It allows assemblage characterisation, e.g. species composition and their environmental requirements (Hill et al., 2020). For instance, (Chu et al., 2019) combined predictions from individual models developed for cold-water coral and sponge groups to identify areas as suitable assemblage for multiple biogenic habitat-forming species. In VME studies, the individual taxon predictions have mostly been assembled based on a prior beliefs about how assemblages are structured [e.g. grouping by higher taxonomic level (Yesson et al., 2012) or functionality (Davies and Guinotte, 2011)], which may not reflect real distribution patterns or confer a limited ecological interpretability (Jansen et al., 2018).

To identify a VME in accordance with the FAO guidelines, the five criteria need to be encoded into the modelling approach ( Table 1 ). However, most VME distribution modelling studies focused on predicting suitable locations for VME indicator taxa, which provides an incomplete assessment according to the FAO multi-criteria guidelines. Nevertheless, the translation of the FAO criteria into modelling terms remains unclear, especially for the “functionality” and “structural complexity” criteria. The “functionality” of an ecosystem cannot be easily modelled given the current FAO VME definition and the available knowledge of these remote ecosystems. The “structural complexity” of a habitat can potentially be assessed if abundance data with excellent detectability is available, e.g. by characterising the density and/or the morphologies of the biotic and abiotic features.

The FAO guidelines do not specify what constitutes a “significant concentration” (Auster et al., 2011; Vierod et al., 2014; Rowden et al., 2020). A quantitative criterion is needed to distinguish dense aggregations of VME taxa that form a “structurally complex habitat” from low frequency occurrences of VME indicator taxa. Most RFMO/As have established operational definitions based on thresholds of VME indicator taxa caught as fisheries bycatch (Bell et al., 2019), which then triggers precautionary management actions, e.g. CCAMLR conservation measure CM 22-07 (CCAMLR, 2013). However, the relevance of these operational encounter thresholds has been questioned (Auster et al., 2011; Ardron et al., 2014; Watling and Auster, 2017; Geange et al., 2020) because they (i) have undergone limited scientific validation and, (ii) are not taxon specific, and therefore do not take into account the differential catchability of taxa and their differential vulnerability, distribution or life history characteristics. In addition, VME extent is operationally defined in many RFMO/As by an arbitrary distance around the location where a certain amount of VME indicator taxa has been encountered, e.g. CCAMLR conservation measure CM 22-07 (CCAMLR, 2013), often without further scientific investigation. Further work is then needed to establish quantitative references, scientifically validated, of what constitutes a VME, both in terms of taxon abundance and VME extent.

With the vast majority of the deep-sea being unsampled (Clark et al., 2016b), distribution modelling has the potential to play an important role in guiding scientific surveys aiming to locate undiscovered VME (Chu et al., 2019; Baco et al., 2020; Long et al., 2020). For this application, models with large spatial scale and coarse resolution can be employed to broadly identify areas that may harbour VME (Vierod et al., 2014). Presence-only data from unstructured and opportunistic datasets can be used to identify the most suitable sites for a particular VME taxon occurrence. For instance, (Yesson et al., 2017) model the rank suitability of black corals using presence-only data with P_S<<1 or inconsistent P_S across the collated datasets (see Figure 1 case 1, “Presence-only P_S<<1”), at a global scale and with a 1 km² resolution. At a smaller scale, distribution modelling can also be used to enhance sampling programs in areas where data availability is limited, e.g. by filling gaps in survey coverage (Kenchington et al., 2014).

Spatial closures are widely recognised as the most effective method for avoiding serious adverse impacts on VME (Bell et al., 2019). Distribution modelling has increasingly been employed to assess their effectiveness and relevance for VME protection, e.g. (Lagasse et al., 2015; Rowden et al., 2017; Buhl-Mortensen et al., 2019; Georgian et al., 2019; Torriente et al., 2019) or the recent protection effectiveness assessment by the Northwest Atlantic Fisheries Organization (Kenchington et al., 2019b). To assess whether the policy driven conservation targets are met or not, a common metric is the percentage of VME spatial distribution included in protected areas. For this application, modelling the relative likelihood of occurrence, e.g. using Presence-absence data with P_S < 1 but constant, can be sufficient, see section Connections Between the Type of VME Data and Model Use. For instance, (Howell et al., 2016) model the relative likelihood of sponge grounds to quantify the percentage of these that are protected by reserves. Ross et al. (2015) study the influence of the model resolution on the estimation of this metric and draw two important conclusions: (i) low resolution models are sufficient for the assessment of marine reserve effectiveness at regional scales, e.g. to appraise the progression towards protection targets, (ii) high resolution models are recommended for detailed reserve design at local scales, because of their better accuracy in detecting key habitat occurrences. Alternatively, another relevant metric for this application is the absolute VME spatial extent contained in the reserve system (Williams et al., 2020), for which the estimation of the true probability of occurrence is needed, see model requirements in section Connections Between the Type of VME Data and Model Use. Even when the data exist to model the true probability of occurrence, Ross et al. (2015) show that predicted VME spatial extents vary with model resolution without consistent direction and proportion across habitats (i.e., different taxa respond differently to a change in model resolution). This means that absolute extent predictions should be interpreted with caution. For a more robust VME extent estimation, we recommend the use of an abundance-based model (instead of probability of occurrence), e.g. using underwater imagery data and the method proposed by Williams et al. (2020).

Structural and functional connectivity among protected areas is an important aspect of conservation management and planning. Often used as a proxy for resilience, the analysis of population connectivity is especially valuable for VME indicator taxa because of their late maturity, low fecundity and with uncertain recruitment sources (Kenchington et al., 2019a; Rowden et al., 2019). For this application, it is important to consider the entire VME indicator taxon distribution (i.e., all species occurrences, not only the associated VME habitat distribution) since the connectivity between habitats (e.g., coral reefs) is likely to be maintained by the wider species distribution (Howell et al., 2011). Taxa distribution and dispersal models have been recently compared to investigate connectivity between VME, in order to study how many interbreeding populations are present within a given metapopulation (Ross et al., 2019) or to analyse oceanographic features that modulate larval dispersal connectivity (Kenchington et al., 2019a). Although highly relevant for conservation assessments, the feasibility of these studies remains hampered by our limited knowledge on ecological processes influencing these rare and remote communities (Parker and Bowden, 2010).

Distribution modelling can be used to describe the community composition of VME, e.g. the diversity and abundance of VME indicator taxa or the association with other macrofauna. Community characterisation is highly relevant during site prioritisation of management planning. In the early stages of management planning, models estimating relative likelihood of occurrence can be valuable to identify sites that are entirely unsuitable for VME (where risks from a proposed activity are minimal), or to rank vulnerable sites within the same modelled region (Baker et al., 2021). However, to prioritize sites with a high VME taxa biodiversity, e.g. based on species richness mapping, the use of probability or abundance-based models is recommended (Gonzalez-Mirelis et al., 2021), see requirements in section Connections Between the Type of VME Data and Model Use. Besides the VME protection, some studies examine spatial correlations between VME and other macrofauna presence or abundance, e.g. (Stone, 2006; Tissot et al., 2006; Kenchington et al., 2013). For instance, Ashford et al. (2019) map the relative biomass of several VME indicator taxa to analyse the community structure and their influence on precacarid assemblages. More community-oriented studies are needed to better understand the functionality of VME taxa, such as their association with commercially important deep-sea fishes.

The reduction of significant adverse impacts from human activities on VME, as urged by the UNGA 61/105 resolution (UNGA, 2006), is reliant upon a comprehensive impact assessment, for which distribution modelling can play a key role (Gianni et al., 2016; Bell et al., 2019). For this purpose, we recommend comparing the predicted distribution of VME taxa with fishing footprint, as done by Brewin et al. (2020) which assessed the impact of fishing activities on VME across adjacent areas with different management policies. Using species distribution modelling with environmental and fisheries variables, Lauria et al. (2017) showed that two VME indicator taxa have a very different vulnerability to fishing trawlers, since the predicted distribution of one studied taxon overlapped with fishing activities while the other is estimated to only occurs where the fishing is low or absent. Besides fisheries pressure, distribution modelling can also be used to assess impact of natural disasters on VME (Georgian et al., 2020). Ideally, the damage or threat needs to be characterised against different factors, such as intensity, timing and duration of the impact, its spatial extent, and the VME’s ability to recover (Penney and Guinotte, 2013). From these characteristics, the ecosystem’s intactness (i.e. naturalness: lack of or low level of human induced disturbance or degradation) can be derived by adjusting the model outputs with the recorded human activity, e.g. by discounting the VME predictions as per the bottom-trawling impact (Ardron et al., 2014). VME intactness mapping can guide decision-makers towards maximising the protection of VME that are still intact but fall within fishing depths, or monitoring the recovery of VME that have already been significantly damaged.

Besides the spatial distribution, some user groups are also interested in understanding the relationship of VME with environmental conditions. For this application, distribution modelling is mainly used to synthesize the available knowledge (Etnoyer et al., 2018 ; Lauria et al., 2021), for instance to define what are the primary drivers of a functional VME unit, e.g. (Williams et al., 2020). Importantly, VME characterisation in terms of terrain parameters is highly scale and resolution dependent (Davies and Guinotte, 2011; Rengstorf et al., 2012), which emphasizes the importance of correct spatial matching between response data and predictors for this application (Ross et al., 2015).

Distribution modelling can also deepen our understanding on how climate change is leading to shifts in VME distributions (Morato et al., 2020; Tittensor et al., 2010), which is critically important when developing management plans (Levin and Bris, 2015; Pecl et al., 2017; Sweetman et al., 2017; Johnson et al., 2018). To achieve this, VME distribution models can be fitted with current environmental conditions, and be used to infer distribution changes under future climate projections. If the model-users are only interested in relative changes, models estimating the relative likelihood of VME occurrence can be suitable, see requirements in section Connections Between the Type of VME Data and Model Use. In contrast, models able to estimate the true probability of occurrence are required in order to compare the VME taxa persistence or the changes in VME spatial extent across habitats. Further, the “Assemble first then Predict” approach is not suitable for this application since taxa may be expected to respond differently to future environmental conditions. As an example for this application, (Morato et al., 2020) recently inferred the rank suitability of six cold-water coral species occurrence under the business-as-usual emissions trajectory RCP8.5 for the period 2081–2100, aiming to identify key refugia areas for VME taxa. Their results suggest an alarming decrease in suitable habitat, ranging from 28% to 100% depending on the VME taxa. For these analyses, we however recommend acknowledging the various sources of uncertainty that can limit their accuracy, such as species evolutionary changes and acclimation (Morato et al., 2020).

The FAO VME identification guidelines are policy-oriented and their translation into modelling is not straightforward (see section Lack of Quantitative Standards of What Constitutes a VME). There is a need for VME identification criteria which are both model-oriented and user-specific. Therefore, we recommend early and iterative communication between model-users and -builders to converge on quantitative VME identification criteria that are suitable to the user’s need. To find a multi-country and multi-stakeholder consensus is often challenging. Therefore, the involvement of as many parties as possible during the different steps of the modelling process increases the likelihood of models’ acceptance and of a successful spatial management planning outcome (Ardron et al., 2014; Rowden et al., 2019).

Unlike the occurrence or abundance of a species which are absolute measures, vulnerability is relative and qualitative, i.e. it is a spectrum, from hardy and resilient to fragile and vulnerable, and dependent on a specific threat, specific region, etc. To ensure an adequate and precise interpretation of the model’s outputs, VME distribution models should be based on criteria that are specific and a priori clearly communicated to model-users. Vulnerability criteria should be defined relative to specific threat (e.g., fishing vs. deep sea mining) and to the characteristics of the VME data available for modelling (e.g., model cut-off values should consider the detectability of the sampling method used to collect the training data). For regional models, the VME policy-oriented criteria defined by some RFMO/As for their area of jurisdiction constitute a relevant starting point. For instance, CCAMLR considers the motility and the larval dispersal as additional VME criteria, as relevant aspects for the Southern Ocean context and the longline fishery (CCAMLR, 2009). To date, VMEs have been mainly modelled in regard to the fishing threat. For applications not only focusing on fisheries management, the seven criteria developed by the Convention on Biological Diversity to identify Ecologically or Biologically Significant Areas (EBSA) should be taken into consideration (Dunn et al., 2014). As shown by (Ardron et al., 2014), the EBSA identification criteria complement the FAO VME criteria with the assessment of the area naturalness (or ecosystem intactness, i.e., lack of or low level of human induced disturbance or degradation), the biological productivity and diversity [see (Gonzalez-Mirelis et al., 2021)].

Establishing quantitative-based criteria of what constitutes a VME is needed to impede subjectivity in VME identification. To reach a consensual definition of an ecosystem vulnerability, the establishment of cut-off values, data driven and scientifically validated, is needed (Kenchington et al., 2014; Geange et al., 2020; Rowden et al., 2020). To this end, Rowden et al. (2020) model the relationships between the density of a habitat-forming taxon and the number of other epifauna taxa, i.e. correlation between the amount of structurally complex habitat and the associated species richness. This analysis facilitates the translation of the subjective “Structural complexity” FAO criterion into a quantitative model-oriented criterion by determining percentage cover thresholds that are representative of structurally complex habitats. Such methodology has the potential to operationalise FAO’s criteria while providing distribution models with a data-driven and scientifically validated reference to identify VMEs.

While most VME modelling studies focus on the distribution of a single indicator taxon, epibenthic community composition analyses would provide a more relevant assessment of both the ecological functionality and structural complexity of an ecosystem, as per key FAO VME criteria. Therefore, we recommend modelling the full list of VME taxa instead of focusing on a single taxon or habitat, and to assess the conservation value of the predicted hotspots by characterizing the whole epibenthic community in relation to the predicted VME taxa distributions.

To model the distribution of multiple species, approaches with joint responses among the species, such as Hierarchical Modelling of Species Communities (Tikhonov et al., 2020), Species Archetype Model (Dunstan et al., 2011; Hui et al., 2013) or Regions of Common Profiles model (Foster et al., 2013), should be investigated as they can directly reveal the VME’ community structure (Chu et al., 2019). Besides the assemblage approaches commonly used in the VME literature (see bottom panel of Figure 2 ), future studies could consider the “Analyse simultaneously” approach which concurrently correlates biological with environmental data to generate regions containing relative homogenous and distinct assemblages within a single model (Hill et al., 2020).

Once hotspots of VME taxa are identified, the ecological value of a site can be evaluated in respect to the biodiversity and productivity of the entire epibenthic community, e.g. by computing the total epibenthic megafauna richness and abundance (Dunstan et al., 2012; Gonzalez-Mirelis et al., 2021). The total epibenthic megafauna richness and abundance can arguably be used as relevant quantitative indicators to assess the “structural complexity” and “functionality” FAO criteria (Gonzalez-Mirelis et al., 2021). Further, since biodiversity and productivity are part of the EBSA identification process, their quantification in VME distribution studies supports a better incorporation of VME protection within management tools across sectoral authorities and international agreements.

The use of abundance data extracted from underwater imagery is preferable for conducting the community-oriented modelling described above. Imagery-based abundance data provides a significantly more reliable, comprehensive and non-destructive characterization of benthic communities than other sampling methods (see section Connections Between the Type of VME Data and Model Use). Besides, the visual assessment allowed by images is more compelling for non-specialists such as policymakers. Rapid advances in imaging technologies have recently reduced costs and thus increased underwater imagery coverage (Morato et al., 2018). Further, there has been a substantial global effort to make these images publicly available, e.g. MAREANO video database (Buhl-Mortensen et al., 2015). To increase the spatial and environmental coverage of the training dataset, we encourage model-builders to collate all available data that have been already acquired across the studied region, even when not annotated yet (Etnoyer et al., 2018 ; Lauria et al., 2021). Noteworthy recent efforts have been made to standardize protocols for VME identification (Baco-Taylor et al., 2020) and marine reference catalogues (Howell et al., 2019; Horton et al., 2021), both from underwater imagery. While indicator taxa are often defined with a low taxonomic resolution by RFMO/As, the use of images opens the door to higher resolution identifications, which can provide a finer characterization of ecosystems. In addition, taxonomic groups can be split based on morphological characteristics, e.g. using morpho-taxonomic labels (Gonzalez-Mirelis et al., 2021), or functional traits in order to further appraise the ecosystem vulnerability, e.g. “Fragility” FAO criterion. All in all, the increase of both quantity and quality of abundance data is likely to underpin the next generation VME distribution models.

Most published VME distribution models have focused on mapping the occurrence of indicator taxa, which only provides a partial assessment of the entire ecosystem vulnerability. VMEs should instead be identified on the basis of the full suite of vulnerability criteria outlined in the FAO Guidelines (Gianni et al., 2016). Therefore, we recommend future studies to model the vulnerability of an area, rather than its suitability to harbour an indicator taxon. To map the vulnerability of an area, Morato et al. (2018) propose weighting the VME taxa abundance data by taxa-specific scores for each of the five FAO criteria. This multi-criteria assessment method results in the mapping of a single composite metric, the “VME index”, which represents where a particular location lies on the vulnerability spectrum. Some drawbacks are however inherent to such composite score calculation (Ardron et al., 2014), such as sometimes leading to prioritizing sites harbouring a large number of taxa with low vulnerability, over a habitat dominated by a highly vulnerable taxon. There is also no global consensus on what would be the optimal integration method for the criteria to provide a single conservation value to an area (Yamakita et al., 2017). Nevertheless, this methodology has been adopted by the International Council for the Exploration of the Sea (ICES) in the North-East Atlantic (ICES, 2019), and has been recently used to generate the “North Atlantic Ocean basin-scale VME index dataset” (Morato et al., 2020). Building on Morato et al. (2018) method, Burgos et al. (2020) map the “VME index” beyond observed sites using the predictions of VME taxa distribution models. Avenues of improvement include: (i) to use joint-distribution modelling approaches, as detailed in section Appraise Spatial Closures Effectiveness for VME Protection, and (ii) to derive vulnerability scores for each taxon from imagery-based analyses that are less prone to expert subjectivity because data-driven and scientifically validated, see section Locate Undiscovered VME to Guide Scientific Surveys.