A framework for the development of a global standardised marine taxon reference image database (SMarTaR-ID) to support image-based analyses

Video and image data are regularly used in the field of benthic ecology to document biodiversity. However, their use is subject to a number of challenges, principally the identification of taxa within the images without associated physical specimens. The challenge of applying traditional taxonomic keys to the identification of fauna from images has led to the development of personal, group, or institution level reference image catalogues of operational taxonomic units (OTUs) or morphospecies. Lack of standardisation among these reference catalogues has led to problems with observer bias and the inability to combine datasets across studies. In addition, lack of a common reference standard is stifling efforts in the application of artificial intelligence to taxon identification. Using the North Atlantic deep sea as a case study, we propose a database structure to facilitate standardisation of morphospecies image catalogues between research groups and support future use in multiple front-end applications. We also propose a framework for coordination of international efforts to develop reference guides for the identification of marine species from images. The proposed structure maps to the Darwin Core standard to allow integration with existing databases. We suggest a management framework where high-level taxonomic groups are curated by a regional team, consisting of both end users and taxonomic experts. We identify a mechanism by which overall quality of data within a common reference guide could be raised over the next decade. Finally, we discuss the role of a common reference standard in advancing marine ecology and supporting sustainable use of this ecosystem.


Introduction
There is a long history of using images in marine ecological studies. The first underwater photograph was taken in 1856 in UK seas [1] but it took until 1893, on the sunlit Mediterranean seabed, for the first clear images to be produced [2]. Following this, the use of underwater photography became widespread in shallow seas, opening up this environment to a wider public (e.g. [3]). The first deep-sea photograph was taken from the porthole of a bathysphere in the early 1930s [4] and shortly after, the first self-contained deep-sea photographic systems were developed in the 1940s at the Woods Hole Oceanographic Institution [5,6]. Whilst there were many good deep-sea photographs available between this time and the early 1970s [7,8], few biologists studied them, as often no corresponding samples of animals were taken, making identification difficult [9]. The notable exceptions to this [9,10,11,12,13,14] paved the way for photography to become established as an important tool for the study of deep-water environments [15,16,17,18,19]. Today, with the routine use of seafloor cameras, towed camera platforms, remotely operated and autonomous underwater vehicles (ROVs and AUVs), photographic assessment of marine fauna and faunal assemblages is a vital tool for research used by both scientists and industry [20,21,22].
Imaging is an important non-destructive tool for studying marine geology and biodiversity at a wide range of spatial scales (from millimetres to tens of km) [21,23]. It enables a rapid assessment of wide areas while retaining valuable ecological information, such as spatial distribution and associations between organisms and with the landscape. Photographic and video assessment is particularly useful in complex terrain or sensitive areas [24,25], where direct sampling is challenging or undesirable. Imaging is generally used to provide both qualitative and quantitative information on the marine environment (e.g. sediment type [26]; hyperbenthic (living immediately above the seafloor) and midwater organisms [27]; benthic epifauna (the organisms living on the sediment surface [24,28,29]); and faunal activity or behaviour (through visible life traces or video/time-lapse images [30,31,32]). As a nondestructive tool, imaging is also paramount in the identification of Vulnerable Marine Ecosystems (VMEs) [33,34]. It has also been widely used to access the impact of human activities on benthic communities e.g. [35,36] and to evaluate the distribution of marine litter on the seafloor e.g. [37,38]. Imaging has also been applied to detecting and assessing temporal variation [22,39]. Estimates of organism densities from seafloor imagery have proven more accurate than those obtained by physical sampling methods, such as trawling. For instance, densities derived from seafloor imagery provided a 10-50 fold increase in accuracy in comparison to trawling in the Porcupine Abyssal Plain in the North East Atlantic [40]. However, it is likely that diversity is underestimated as a result of difficulties of identification of the taxa to lower taxonomic levels from imagery [21].
The use of images to collect faunal data brings with it the challenge of identifying taxa from image data. Identification of physical specimens is usually achieved using taxonomic keys that have been developed by experts working on specific taxonomic groups. These keys are developed based on thorough study of preserved specimens, incorporating a systematic analysis of characteristic morphological features, followed by the development of a dichotomous key. While traditional taxonomic keys may be useful in the identification of some taxonomic groups from imagery (e.g. fish), many such keys rely on characteristics that are not visible in imagery (e.g. the arrangement of mesenteries in anemones, spicule shape in sponges, sclerite morphology in gorgonians, and the ossicles of holothurians). Therefore, for many taxonomic groups the development of field guides are essential to support taxon identification from image data. Many field guides have been developed for shallow-water marine species for use by SCUBA divers. These rely heavily on image data to show form, function and details of anatomy that can be used for accurate identification e.g. [41,42], but they are rare for depths beyond recreational SCUBA diving capability (>30 m) (hereinafter referred to as deep-water species). Good field guides are usually underpinned by a comprehensive understanding of the species pool for the region of study. For most deep-water regions, this understanding is lacking. Notable exceptions include the Monterey Canyon [43] and the soft sediment (trawlable) habitats of the North Atlantic. The lack of comprehensive field guides for deep-water marine organisms presents a significant challenge to those faced with the interpretation of image data from poorly known regions or habitats, such as seamounts, ridges, or other areas of hard and high-relief substrates that are not conducive to trawling surveys.
In the absence of a good knowledge of the taxonomy of many groups and regional field guides, a common practice in the interpretation of image data is the development of a morphospecies reference image dataset (Fig 1) and the use of operational taxonomic unit (OTU) numbers. The OTU numbers are used in place of taxon names for organisms for which a species name has not yet been assigned owing to the lack of physical specimens to corroborate the observation [24,43,44,45,46,47,48,49,50]. These morphospecies reference image catalogues provide a permanent reference of what has been observed in the study. But perhaps more importantly, allow the user to differentiate between taxa below the lowest level of the taxonomic hierarchy to which the observed organism can be identified, using traditional taxonomic features, and thus preserve important information on biodiversity. For example, taxonomic identification of many sponge and soft coral species is impossible from image data alone, since their taxonomy is based on the arrangement, size and shape of microscopic structures in their skeletons. Thus, following traditional methods of sample analysis, all observed species would be assigned the level Porifera or Alcyonacea, resulting in a significant loss of resolution in the data. However, use of a morphospecies reference image catalogue allows the observer to assign morphologically different (and in most cases, likely taxonomically distinct) forms to a unique OTU number, which can then be assigned to the taxon (e.g. Porifera msp. 1, Porifera msp 2 etc.) if needed, thereby retaining taxonomic resolution in the data.
The problem with this approach is that each study or group uses a different naming convention for morphospecies. It then becomes impossible to compare or combine datasets between studies. Morphospecies catalogues are not usually published, making it difficult for researchers to compare data or check identifications. Comparison between research studies or industrygathered data (for example from environmental impact assessments or site monitoring) are also impaired by this issue. In addition, both field guides and morphospecies reference image catalogues fail to document explicitly the visual characteristics used to differentiate taxa. They generally provide little more than a visual idea of what a taxon looks like. This compounds problems of observer biases that are well documented in biological sample analysis [51,52,53). When identifying taxa from image data, it is necessary to use a combination of traditional taxonomic features and ecological data (e.g. depth, location, habitat, knowledge of the local species pool) to arrive at an identification. This skill in 'field identification' is often acquired through an 'oral tradition' with little in the way of formalised training materials provided to new researchers entering the field or new consultants provided with image data to analyse.
Developments in autonomous and robotic technology, and the increased use of them across different fields, are increasing the amount of image-based data that can be collected [54,55,56]. For example, a single 22-hour AUV mission returned over 150,000 seafloor images [40,56]. Manual image analysis is a time-consuming process, which forms the current bottleneck in image-based ecological sampling [21,57,58,59]. As a result, a number of research teams are investigating the use of artificial intelligence (AI) and computer vision (CV) as potential means to accelerate and standardise the interpretation of ecological image data [51, 52, 53, 56,  60]. The most promising of these techniques is supervised machine learning to automatically detect and classify taxa [53,58,61]. However, consistent interpretations by humans are initially required, providing 'gold standard' classifications, with as much data as possible, which can be used to train these algorithms. Moving forward, developments in AI and CV approaches that combine the use of visible morphological characteristics with deep learning, would benefit significantly from the development of a standard image-reference dataset. For those taxonomic groups in which the morphological characteristics commonly used to differentiate taxa are not discernible in images (e.g. sponges, anemones, zoanthids and plexaurid gorgonians), these types of combined approaches will first require development of novel visual multi-access keys, which themselves can only be created from a high-quality reference image dataset and skilful determination of characteristics differentiating taxa.
While there are a variety of on-line open-access databases that are designed to archive biological and ecological information, including genetic data (for example GenBank), species occurrence records (for example the Global Biodiversity Information Facility GBIF) and even images of taxa (for example Morphbank), there are few that provide a reference guide to support the interpretation of image-based datasets. Table 1 provides a list of existing field guides and morphospecies reference image catalogues for deep-water species of the Atlantic Ocean that are currently publicly available. However, many more are un-published or inaccessible to others, and are held as a mixture of printed and electronic materials. Recently there have been attempts to make morphospecies reference image catalogues associated with specific research programmes or projects available to others (for example [43,47,62,63,64,65,66] to mention a few). In addition classification based approaches to this issue have also been developed [67]. While useful, this 'piece-meal' approach will not solve the challenges outlined above.
There is a clear need for the development of a standard reference guide to support the use of image-based sampling. Failure to develop appropriate tools will ultimately hinder progress in marine ecology, particularly in deep-sea marine ecology where images are frequently one of the few collected datasets. In order to improve data quality and comparability, realise the benefits of new technologies in both image data collection and interpretation, and ultimately raise standards of taxonomic identification within academia, government, and industry, we must move towards the use of standard reference guides, quality controlled and curated by experts in both taxonomy and field identification.
Our aims were to develop 1) a database structure to facilitate the standardisation (and ultimately pooling) of morphospecies reference image catalogues between individuals and groups, supporting onward use in multiple applications; and 2) a framework for coordination of international efforts to develop reference guides for the identification of deep-water species from image-based data.

Methods
The initial stages of developing the framework for the database consisted of assessing the requirements of those working with image-based data through end user group discussions, both informally and as part of an international workshop. This included the need for both online and offline databases and printable catalogues for use in making identifications at sea.
We reviewed current relevant databases and database standards. These were focused around the Darwin Core standard, the Ocean Biogeographic Information System (OBIS), and the World Register of Marine Species databases (WoRMS), which are all used regularly by the end user community.
The Darwin Core is an international standard set of terms and definitions that facilitates sharing biodiversity data [84]. The Darwin Core quick reference guide (http://rs.tdwg.org/ , and, thus, is the internationally agreed standard upon which further database development should be based. Darwin Core Archives (DwC-A) comprise a set of text files, including both the dataset (.csv) and a document (.xml) which describes the included files, fields, and their relationships. This offers a standard format used to describe biodiversity data and is being commonly employed to share more complex and structured datasets. OBIS [87] was originally developed as the information management component of the Census of Marine Life (2000-2010) programme. OBIS founder, Dr. J. F. Grassle, articulated the vision of OBIS as "an online, worldwide atlas for accessing, modelling and mapping marine biological data in a multidimensional geographic context". The OBIS database currently consists of over 55 million observations of nearly 124,000 marine species. In 2009, OBIS was adopted as a project by the International Oceanographic Data and Information Exchange (IODE) programme of the Intergovernmental Oceanographic Commission (IOC) of UNESCO. It represents an internationally important archive for species distribution data. OBIS is closely linked with WoRMS, which provides the taxonomic backbone, and geospatial data are provided by the Marine Regions database. Additional functionality includes the taxon match tool for resolving names used by other similar platforms, providing crucial quality control support for taxonomic data among the research community and biodiversity platforms [89].
WoRMS is an authoritative classification and catalogue of marine names including information on synonymy, and is curated by around 400 taxonomists globally, in accordance with best practice [88,89,90]. The content of WoRMS is managed by taxonomic and thematic experts, who are responsible for controlling the quality of the information contained within the database [89]. WoRMS is underpinned by the Aphia platform, which is a Microsoft Structured Query Language (MS SQL) database, containing over 400 fields spread over more than 80 related tables. This infrastructure is designed to capture taxonomic and related data and information. WoRMS is also the basis of the World Register of Deep-Sea Species (WoRDSS), which, through its app, Deep Sea ID [68], represents one of the few existing image-based deepsea species guides (but see Table 1).
The Marine Regions database [91] provides a standard, relational list of geographic names, coupled with information and maps of the geographic location of these features. All geographic objects of the Marine Regions database have a unique ID, called the Marine Regions Geographic Identifier (MRGID). The different geographic objects are determined by a placetype and coordinates. While the coordinates are represented as different vector data types being a point, a line or a polygon, a placetype provides contextual information to the geographic objects, for example a sea, a bay, a ridge, a sandbank or an undersea trench. Following the initial review of relevant databases and database standards, a strawman database architecture, to facilitate the standardisation of morphospecies reference image catalogues between individuals / groups, was proposed and circulated to an international team of end users, database specialists and programmers. An international workshop funded by the Deep-Sea Biology Society was held at Plymouth University, UK, on the 4 th -5 th December 2017, where the draft structure was reviewed and refined. The workshop consisted of a cross section of attendees including major dataset holders, computer scientists, taxonomists, benthic ecologists, and representatives from WoRMS / WoRDSS. Following the workshop, the refined structure was tested by both workshop participants and members of the wider community, who input their existing morphospecies reference image catalogues into the proposed format. This resulted in further minor changes and the development of the final data-sharing structure.
Workshop participants also considered how to coordinate international efforts to develop reference guides to the identification of deep-water species from images. The following questions were considered by the workshop attendees, how can we: 1) merge existing published and unpublished catalogues? 2) manage new submissions to a merged catalogue? 3) improve the scope and quality of the image data within a merged catalogue? and 4) improve and classify the quality of identification from images?

End product needs
Workshop participants, and specifically those engaged in image-based analysis, felt the most critical tools urgently required to support their work were in-situ photo-guides in book format (hard copy or e-book), a standard reference morphospecies taxonomic tree (or annotation scheme) that can be imported into different annotation software, and on-line user-friendly image reference catalogues that include information on characteristics used to classify animals as belonging to a particular OTU. The final database structure must therefore be such that these end-use products can be easily created from the database by a query using purpose-built web-accessible software as part of future developments.

Database structure
The final database structure consists of two tables that contain fields that map onto Darwin Core fields, together with additional fields for which no Darwin Core equivalent could be established. Table 2 is the OTU table. It documents the OTU, and primarily maps to fields from the Darwin Core classes "Taxon" and "Identification". Table 3 is the image table. It documents the individual image file and maps onto fields from multiple Darwin Core classes, including "Occurrence", "Identification", "Event", "Location", "Record-level", and "Organism". The two tables are related via the "OTU" field. This structure allows a single OTU (one entry into Table 2) to be related to multiple example images of the OTU (many entries in Table 3).
The OTU table (Table 2) consists of a GUID field "Number", the inclusion of which is standard practice in database tables. The "OTU" field is a unique number given to this taxon and is initially assigned by the user. The subsequent four fields: "scientificName", "scientificNa-meID", "scientificNameAuthorship", "taxonRank", provide the link to the WoRMS database. The link is via the "scientificNameID" field, which requires the user to input the appropriate Life Science Identifier (LSID) for the OTU drawn from the WoRMS database. Each taxon in WoRMS receives a unique and persistent identifier, known as the AphiaID. This AphiaID can be expanded to a LSID. WoRMS has implemented LSIDs for all its taxonomic names and they are displayed on each taxon page. The LSID integrates the AphiaID and so is the preferred option, of the two possible fields, to use as a link. The appropriate LSID for an OTU is the lowest formal taxonomic rank that can be assigned to an image. For some taxa, this may be at the species level; however, for many image-based identifications it will be at a higher taxonomic level, such as Family, Class or Phylum level. Use of the LSID field ensures that the OTU can be linked to standard taxonomic nomenclature and the related taxonomic hierarchy. Using this LSID, the other three fields within the database ("scientificName", "scientificNameAuthorship", "taxonRank") can be auto-populated from WoRMS.
The "Morphospecies" field maps onto the "identificationQualifier" field in Darwin Core and allows the input of extra details distinguishing between different morphotypes; for example, Brisingidae msp1, or in the case of sponges, Porifera encrusting msp1, Porifera branching msp1. Thus, entries into this field will be of the form msp1, msp2, encrusting msp1, branching msp1, etc. The "CombinedNameID" field is then autopopulated by adding the "scientific-Name" and "Morphospecies" fields to give, for example, Brisingidae msp1, Porifera branching msp1. The "CombinedNameID" field can be mapped onto the "taxonconceptID" Darwin Core field. A recommended best practice for the standardisation of entries to the "identifica-tionQualifier" field, specifically related to nomenclatural qualifiers used in image analyses is now in preparation. The "PreviousName" field is not intended to document recombinations of taxonomic nomenclature as this is captured and managed in WoRMS [90]. Rather, this field is to capture changes to the assigned identity of the OTU. For example, where Brisingidae  Global marine taxon reference image database organism is very likely this species/taxon based on investigation (literature search, consultation with outside taxonomic experts, 3 = Certain: the organism has been collected and has been definitively identified by a taxonomic expert. We have modified these categories as follows: 1 = identified from image only, 2 = identified from image and physical specimens sampled from the same region, 3 = identified from image and physical specimen of the actual individual in the image. There are often instances where an organism has been identified from an image and a specimen collected that has not yet been identified. Under these circumstances the quality score would be 1, but the existence of a specimen noted in the "PhysicalSample" field. Once a specimen is identified the quality score for the image could be changed to 2 or 3.
The fields pertaining to the Darwin Core class "Location" concern where the image was taken. We recognise that for older image data archives, exact position data may not have been recorded. However, the importance of location and depth to field identification of taxa cannot be understated. We feel it is important to ensure that the terminology used to define location is consistent with a published standard. In addition, we want to ensure that, in the future, users will be able to construct local morphospecies reference image catalogues based on selection of an area through mapping software. The Marine Regions database [91] is ideally placed to provide this geospatial standard. Its use will also ensure compatibility with OBIS such that this database can share data with OBIS and vice versa. The required fields "locality" and "loca-tionID" provide the link to the Marine Regions database. The user must input the appropriate "locality" and "locationID" for the image drawn from the Marine Regions database. The "loca-tionRemarks" field is an optional free text field that allows users to capture more detailed location information that is not captured by the options available in the Marine Regions database. The fields "minimumDepthInMeters", "maximumDepthInMeters" are also required as species distributions are structured with depth [92] and this characteristic is likely to be important in the development of future field guides. The remaining fields, "decimalLatitude", "decimalLongitude", are optional so as to accommodate older data and / or sensitive data, for example, from industry partners.
The fields pertaining to the Darwin Core class "Record-level" focus on ownership and origin of the image. Required information includes the name of the institution that owns the image ("institutionID"), a licence document ("dcterms:license"), the name of the person / institution managing right over the image ("dcterms:rightsHolder"), and the terms of access to the image ("dcterms:accessRights"). It is anticipated that a standard licencing arrangement can be agreed to upon submission of material to the database, whereby image ownership is retained by the organisation / individual submitting but use for scientific purposes is freely granted. Use of images for commercial gain would be prohibited. There are existing licencing models for WoRDSS and these can be replicated here. Optional fields allow the identification ("collectionID") and citation ("bibliographicCitation") of any previously published or inhouse morphospecies reference image catalogues from which the image data have been drawn. The modified field is autopopulated and is the most recent date-time on which the resource was changed.
There are just two fields that relate to the image collection event via the Darwin Core class "Event". These are the fields "RawImage", which maps to the Darwin Core "eventID" field, and "habitat". It is not the intention of this database to capture details of the research cruises, ROV dives, etc., on which the organism images were taken. These details are not overly important to the creation of a field guide. However, should this information be viewed as important in the future, we suggest that images are given the name of the original image from which the organism was cropped, and that this name be extended to consist of the following elements: CruiseNumber_StationNumber_timestamp_imagename. The "habitat" field is able to capture the geomorphological setting in which the organism was observed, e.g. seamount, canyon, mid-ocean ridge. We felt this information might be useful in the development of a field guide. The ideal situation would be to use standardised terms to describe these settings. We suggest the use of Greene et al. [93] as a standard reference; however, the European Nature Information System (EUNIS) [94,95] or other classification systems may also provide a reasonable standard and the standard used could be indicated when data are submitted. One final field maps to the Darwin Core class "Organism" and is used to capture previous names that have been assigned to the organism in the image ("previousIdentifications"). As with the "Previous-Name" field in the OTU table, this field is not used to capture taxonomic name changes, which are well recorded by WoRMS. It is used to capture changes in opinion on the identity of the organism in the image.
The remaining fields in the Image table do not map to Darwin Core fields but do provide additional information that is important to record. The "SubstrateType" field allows details of the substrate on which the organism was observed to be logged. Substrate is an important environmental factor that determines the distribution of species and can play a role in the field identification of taxa. As always though, it is preferable to use standard terminology to record substrate and there are many standards available. Among workshop participants, there was no consensus on methods of substrate interpretation from image data, and the terminology standards used. Some use EUNIS [94,95], some a modified Folk [96] classification and others percentage of Wentworth [97] sediment size classes. The "SubstrateMethod" field allows the user to indicate the standard they have followed. The "Size" field, standardised to centimetres, is self-explanatory and may be useful in the future development of a field guide. The "Project-Name" field offers the opportunity to credit specific projects with provision of imagery, while the "Link to external database" field enables links to be made to source on-line morphospecies reference image catalogues.
The images are not stored within the table itself but should be provided as separate image files. Those with existing morphospecies reference image catalogues have tended to either paste images into Word or Power Point files, organise their data as Apple ibooks, or organise their images into Phylum or Class level folders. While this is useful at an individual level, and provides the end product required, it limits onward use and is not the appropriate format for a database.

A framework for coordination
While the database structure outlined above provides the means to archive data, the development of a unified morphospecies reference image catalogue requires a management structure to curate the database and manage new data submissions. The WoRMS database provides a model that can be adapted for use with this database. WoRMS is curated by teams who are responsible for different taxonomic groups. Each team is led by an editor who takes overall responsibility for that group. We suggest that the morphospecies reference image database is similarly managed by teams focused at the taxonomic grouping level. The appropriate taxonomic grouping will vary depending on variety represented by each phylogenetic level of the group, and expertise available. For example, Hexacorallia may have separate teams grouped at the Order level (e.g. Scleractinia, Actiniaria, Antipatharia), whereas Echinodermata may have separate teams grouped at Class level (e.g. Asteroidea, Echinoidea, etc.). Each team will consist of experts in taxonomy of the group plus ecologists engaged in field identification of organisms from imagery. We felt it was important to have both taxonomists and field ecologists working together, to ensure that the final database considers both taxonomic rigor and the practical use of the images. Each team will have a nominated lead, and leads will come together, as a steering committee, to ensure that a standard approach to data organisation and curation is achieved across the entire database.
We anticipate a two-stage process whereby an initial effort is made to collate and compile existing morphospecies reference image catalogues at a regional level using the new database structure described above as a data transfer format. In this format not all fields will need to be populated by those submitting data, for example the GUID fields, and other autopopulated fields. This initial effort to compile existing catalogues would be followed by new and on-going submissions of data, including from those encountering new organisms not in the existing database, and from those with higher quality images of organisms already listed in the database (Fig 2). We have committed to stage 1 of this process and morphospecies reference image databases held by all authors have been entered into this new database format and submitted to a central repository. Curation teams are now bringing these data submissions together into a single database. Stage 2 of this process will involve the effort of the global community and could potentially be a focus for the up-coming UN Decade of Ocean Science for Sustainable Development (2021)(2022)(2023)(2024)(2025)(2026)(2027)(2028)(2029)(2030). This could be a very light-touch involvement, where end users simply submit images of new organisms not currently present in the database to the database for inclusion (Fig 2). Or it could be a more targeted and active involvement aimed at raising the quality of the data already in the database. For example, principal investigators of research cruises could actively help to move taxa from "identificationVerificationStatus" 1 to level 3 by targeted insitu imaging and collection of organisms on an opportunistic basis. Raising the quality of the data in the database should be a priority over the next decade, and with that the recognition of the importance of skills in taxonomy and species identification. A concerted effort to ground truth species identified from imagery only, should be made in order to provide robust tools with which to monitor ecosystems.
Ultimately, it is not the database per-se that end users require, but the end products (photoguides in book format, taxonomic tree for annotation software, etc.) that can be pulled from the database. This will require the development of a web interface that draws on the underlying database to produce multiple end use formats (Fig 2). This aspect of the project represents the next stage of development and is anticipated to take place over the next two years.

Immediate advances enabled by the development of a common reference standard
We have proposed a common structure for a database from which a morphospecies reference image catalogue can be built. Our initial development is focused on the North Atlantic deepsea benthos as a case study. However, the structure developed is applicable to any marine region or habitat, and may also be used for terrestrial ecosystems. Individuals need only adopt the structure and populate the tables with their own data. The Standardised Marine Taxon Reference Image Database (SMarTaR-ID) will enable different researchers to bring their data together in a common morphospecies reference image catalogue at an appropriate time. Within the North Atlantic deep sea that time is now. The implementation of coherent monitoring programmes to assess biological biodiversity in marine waters are mandatory under the EU Marine Strategy Framework Directive (MSFD 2008/56/EC), and all European nations are required to monitor sites of community importance every six years. An image catalogue, such as the one herein proposed, will be a powerful instrument to support monitoring efforts, particularly in poorly surveyed regions. We have outlined a framework by which data can be brought together, curated, and new submissions managed going forward, which follows a successful model already applied by WoRMS.
We anticipate the introduction of a common reference standard for the deep sea to enhance significantly our understanding of megafaunal biodiversity by enabling multiple researchers to combine existing datasets to address long-standing ecological questions. This is particularly the case for hard substrate habitats that dominate features, such as seamounts, ridges, banks, abyssal hills, canyons, and areas of the continental slope, and for which image-based techniques remain the only effective means of survey. Past exploration of the deep-sea epibenthic megafauna generated many paradigms, but these were largely built on data obtained using trawls and sledges. Video and still image-based tools have facilitated quantitative sampling of previously inaccessible habitats; and the resulting new findings are challenging the prevailing view of deep-sea ecosystems [98]. However, these new datasets are often limited to individual features or feature types (e.g. seamounts: [99,100], abyssal hills: [101] slopes: [66,102,103] canyons: [48,64,104,105,106]; ridges: [107], fracture zones: [100], and hydrothermal vents: [108]) and thus limit our ability to generalise findings. In their review of major outstanding questions in deep-sea biogeography [109] concluded, among other things, that an integrated biogeographic framework of hard-substrate areas of the deep sea was required to yield more realistic estimates of endemism/cosmopolitanism. It has been repeatedly argued that concerted efforts to link existing independent data streams together to examine long-standing questions of deep-sea diversity are very much needed in order to move the field forward [109,110]. The proposed database will facilitate these advances.
We anticipate that this common reference standard will provide an invaluable tool for environmental managers, industry and wider stakeholders. For environmental managers, it will, for example, enable the development of clearer descriptions and definitions of habitats of conservation concern. For example, deep-sea sponge aggregations potentially qualify as Vulnerable Marine Ecosystems (VME) under the United Nations General Assembly (UNGA) Resolution 61/105. They are also classed as a threatened and declining ecosystem under Annex V of the Oslo-Paris (OSPAR) Convention for the Protection of the Marine Environment of the North East Atlantic. However, comprehensive descriptions of deep-sea sponge aggregations, and specifically the component taxa that compose different types of aggregation, are lacking. In addition, basin-wide data on the distribution of sponge VME indicator taxa are only available for those species / genera whose appearance both in-situ and ex-situ are similar (e.g. Geodia, Hyalonema, Pheronema). For many sponge species, the lack of taxonomic resolution possible when identifying sponges from image data hinders progress in management and conservation of these taxa by limiting our ability to 1) effectively describe sponge VME composition and diversity, and 2) pool data to determine basin-wide distributions. A common morphospecies reference image catalogue will provide a standard reference to use in VME descriptions in the absence of confirmed taxonomic identification of species from physical samples. It will also facilitate the production of basin-wide models of the distribution of habitat forming sponge taxa to support spatial management decisions [111].
For industry, implementation of a standard approach to referencing morphospecies between industry and regulators will facilitate a much more effective impact assessment associated with licensing and consent processes, as well as subsequent monitoring approaches. Often in industry, a range of sub-contractors are used for routine survey and monitoring work by the various industry bodies. Therefore, morphotypes are produced per project with no consistency between sub-contractor or between years in long-term monitoring as data are rarely shared. This standardisation would increase industry and regulatory comparison across applications and across industries to facilitate cumulative impact assessments, thus allowing better understanding of impact at feature and site levels, as required in nature conservation legislation. For industry, this could also decrease levels of risk associated with the assessments as well as decreased analysis time and costs for survey data, and would be a particularly powerful tool if industry could include their own data in the database and play an active role in providing images and survey data.
The need for a standard approach in industry was recently highlighted by the development of the deep-sea mining industry in the Clarion Clipperton Zone (CCZ) of the central Pacific.
Here, baseline data collection is taking place, commonly including seabed imaging-based assessments of megafauna [29,47,65,112,113]. Without a consistent morphospecies reference image catalogue it is difficult to compare studies and generate regional syntheses. This greatly hampers conservation and management efforts, which commonly rely on information on biodiversity, species ranges and behaviour-ecological properties that are difficult to assess without good quality and consistent identifications. Recent work to document megafaunal diversity will help (e.g. 47, 65), but widely adopted and regularly updated catalogues will be vital for improving scientific understanding and effective environmental management.
A common morphospecies reference image catalogue may also serve as a tool to support the identification of taxa from fisheries bycatch by fisheries observers (e.g. [114]. While our proposed database focuses on in-situ images of taxa, we advocate, and have provided for within the proposed database structure, the collection of ex-situ images of taxa. There are a number of existing image guides designed for use by fisheries observers that provide ex-situ images of VME indicator taxa (Table 1). This database could supplement existing guides by providing additional imagery. Interestingly, it may also provide a link between in-situ and ex-situ taxon identification, which may ultimately allow fisheries bycatch data to be pooled with in-situ image data, again broadening our understanding of species distributions (e.g. [115]).
Finally, the simple act of combining multiple existing morphospecies reference image catalogues will advance the overall quality of current identifications. It is important to remember that image data is no substitute for a physical sample, and a long term goal in this endeavour is that all image reference material has been ground truthed by a physical specimen identified by a professional taxonomist. At present different research groups have images of different species for which the "identificationVerificationStatus" level is 3 (the highest level, confirmed by physical specimen). By bringing these reference image sets together, we will collectively have more species that can be identified by reference to images in which we have the highest level of confidence of the animal's identification.

Future advances enabled by the development of a common reference standard
The development of a common reference standard has the potential to advance significantly the field of offshore and deep-sea marine ecology. The ability to pool datasets across time and space will allow us to address a greater range of questions about the offshore and deep-sea benthic ecosystem than is currently possible. Critically, it will enable us to raise standards of identification from image data through the development of training materials and quality control measures. Efforts to develop such tools for shallow water have been undertaken by the UK's National Marine Biological Association Quality Control scheme (NMBAQC). This programme is steered by a range of academic and governmental organisations, and provides guidance on best practice, as well as identification guides, taxonomic workshops, training exercises and quality control ring tests.
There will remain some potential shortcomings on the use of such catalogues related to uncertainties in species identification due to the method of image collection and scale. The ability to zoom-in on specific features of species with ROV cameras means ROVs may provide better imagery for identification than AUVs or drop-down cameras, particularly in cases where species look remarkably similar and occupy overlapping environmental niches. For example, the octocorals Acanthogorgia armata and Acanthogorgia hirsuta can only be distinguished if close up images of the polyps are taken, otherwise identifications have to be left at genus level. Nevertheless, the development of a common reference standard will expose these limitations to a wider audience, and help develop agreed international guidance around the taxonomic levels to which it is appropriate to identify when interpreting image data.
In the longer term, regional field keys are required for use in survey and monitoring of the deep-sea ecosystem. The construction of tools that allow others to identify taxa reliably and consistently in the field is perhaps one of the most underappreciated roles for taxonomists. It is also one of the most challenging roles as taxonomists are often not engaged in field identification, and therefore a gap exists between the generator and end user of a key [116]. The starting point for the development of any key is a standard reference against which to compare new observations. In traditional taxonomy, this is the type specimen, a physical specimen from which a species is described, that is subsequently archived in a museum. The development of dichotomous or polytomous keys is then achieved by measuring the variability in observable characteristics within examples of a taxon and between taxa, then selecting characteristics that best discriminate between taxa for a given region / group. These characteristics are then organised into pathways of character state choices (steps) that lead to identifications.
In order to move forward with the development of much-needed field keys to deep-water taxa, we must first develop an appropriate standard reference against which to assess new observations. This reference point remains the holotype specimen. Our proposed database will establish an image 'similitype' (or a series of images that contribute to the similitype) to accompany a physical specimen that has been identified with reference to the holotype of a species, or through matching DNA sequences to other specimens identified with reference to the holotype of a species, and thus link traditional taxonomy to field identification. This approach will provide a much-needed strategy to advance the taxonomic description of species based on multisource information collected by both ecologists and taxonomists [117,118]. If researchers use and contribute to this common reference standard, a library of images with examples of each taxon will be built up over time. This library of image examples can then be used to understand both the observable characteristics within a species or higher taxonomic level grouping, and the variability in these characteristics in image-based data. Where possible it is desirable for these characteristics to be those used in traditional taxonomic keys. However, this will not be possible for all groups to all levels of the taxonomic hierarchy. For example, while it is possible to use traditional taxonomic features to determine the order of some coral taxa from image data, it is not possible to do this for anemone taxa, which rely on internal characteristics for positive identification. It is likely that novel characteristics, combinations of characteristics, as well as the use of circumstantial information (e.g. environmental characteristics), will be required to enable reliable and consistent field identification of organisms, but even then some morphotypes will remain as morphotypes.
Multi-access keys (also known as matrix based or free-access keys) may be more appropriate than dichotomous or polytomous keys (also known as single-access keys) for use with image data as they, by their nature, have multiple access points [116,119]. Single-access keys place a logical order on the use of characteristics, with each step in the decision tree taking the user along a predefined pathway that progressively narrows the number of possibilities for the identification of the animal. If a characteristic is not visible at any step along this pathway, the choice required by the user is unanswerable and further progression is not possible. Views of organisms in in-situ image data can be highly variable, and it is likely that in any one image only some features will be visible. This may limit the utility of single-access keys with image data. Multi-access keys enable the user to determine the sequence of choices where the user can select from the list of characteristics offered in order to arrive at an identification. In the context of image data, this would allow the user to employ all visible characteristics (and potentially environmental information) to arrive at an identification. Multi-access keys are more suitable for computer-aided identification tools [116,119]. This also makes them a promising tool to use with image data where analysis is computer based.
If we are to move forward with the application of AI and CV to the identification of taxa, we must have a common reference standard. Our proposed database aims to meet this need through future development that will enable the database to interface with image annotation software, such as Squidle [120] and BIIGLE 2.0 [121,122]. These annotation softwares enable users to mark the x,y position of organisms within an image and attribute this point / polygon with a taxon identification. This process of image annotation is the means by which ecologists extract semantic data from an image in order to then apply numerical and statistical analysis to these data and answer ecological questions. This annotated dataset is also the base data needed in the development of AI and CV algorithms. These algorithms require large numbers of images to "learn" the features that distinguish the different OTUs to which they have been exposed, and which of these features are characteristic of each OTU [123,124]. If researchers are able to use a common reference standard, thus extracting the same information from an image regardless of who is annotating it, then collated datasets from various origins could reach the size needed to train and test CV algorithms (acknowledging challenges of observer bias). Their use within the field of deep-sea benthic ecology, will then increase exponentially through accumulation of data, skill and experience. This can only serve to facilitate the development of CV and bring us closer to automation of image annotation and data extraction.
Ultimately, standardisation of tools and methods is central to long-term monitoring and assessments of ocean health. Woodall et al. [125] recognised this and produced GOSSIP (General Ocean Survey and Sampling Iterative Protocol), which outlines a framework of 20 biological, chemical, physical, and socioeconomic parameters that allow marine scientists to generate comparable data on the function, health and resilience of the ocean. There are several international efforts underway to try and harmonise ocean observing in the areas of biology and ecology, including the efforts of the Group on Earth Observation-Biodiversity Observing Network (GEO-BON) and the Global Ocean Observing System (GOOS) panel on biology and ecosystem-Essential Ocean Variables (EOVs). These efforts are also being informed by international efforts, such as the Deep Ocean Observing Strategy (http://www.deepoceanobserving.org/), which is adding deep ocean context to GOOS EOV specifications. There are more than a dozen regional alliances internationally, which are implementing the GOOS vision with international coordination by the IOC. Together these organisations are forming a means for efforts from individual observers, as well as local to international bodies, to join together to realise the power of 'big data' in observing and understanding change. National-level data management and communications groups affiliated with GOOS are now working to include tools, such as automated image classification, into their information technology systems. The common reference image standard described above will therefore contribute to global efforts under GOOS.

Conclusions
We have developed a database structure (and data transfer format) to facilitate the standardisation of morphospecies image catalogues between individuals, research groups, and nations. We have also proposed a framework for coordination of international efforts to develop reference guides for the identification of deep-sea species from images. We have highlighted the potential gains to be made through the use of this database structure by the deep-sea community in: increasing the quality and quantity of data available to researchers, improvement of overall understanding of the deep-sea ecosystem, more effective management and monitoring by statutory bodies and industry alike, and realising the potential benefits of emerging AI and CV approaches. To make these gains it is critical there is now uptake of this database format by the community, and additional funding is found to contribute to stage two development.