A workflow can be thought of as chain of “atomised and executable components with the relationships between them to clearly define a control flow and a data flow” (Hardisty et al. 2022, p. 324). Further:

Digitization workflows span across human mediated processes through data and computationally intensive automation where software tools and services are the actors and intersect field collection techniques, institutional accession policy, differences in curatorial practice among domains, and involvement of the general public in crowd-sourced methods. (Beaman and Cellinese 2012, p. 12)

In their report for the Australian Museum, Tann and Flemons (2008) investigate various means of data capture from images of insect specimen labels, while studying the viability of allocating some tasks to volunteers. Inserting data into an institutional collection management system (in this case, Axiell EMu for internal access only), the workflow options parallel the decisions being made in other collections (Fig. 2). They conclude that there “are some advantages to taking a photograph of a specimen label before capturing the data from that label … the following transcription process is quicker by about 20%, easier and more convenient” (p. 16).

Two years later, in their conference paper Moen et al. (2010) propose a high-throughput workflow for extracting data from specimen labels at the Botanical Research Institute of Texas (IH herbarium code: BRIT; over 1 million specimens at the time). While the paper focuses on meta-data standards, the workflow is notable for its human-in-the-loop aspects (Fig. 3).

In the same year, Granzow-de la Cerda and Beach (2010) analyse digitisation approaches at the University of Michigan Herbarium (MICH) during a 2002-2008 programme focussed on Mexican and Mesoamerican holdings. Nine key tasks were organised in three workflows (Fig. 4), with timing and costing assessed. The authors advocate for the “clear efficiency benefit of articulating a biological specimen data acquisition workflow into discrete steps, which in turn could be individually optimized” (p. 1830) and reiterate that “the limiting cost for collections computerization is the human labour needed to capture data from specimen labels into a structured database” (p. 1831). This work also demonstrates the benefit of language and geographic familiarity for data transcription efficacy.

By 2012, digitisation within collections was sufficiently established that the ALA published the Digitisation of Heritage Materials guidance (Kalms 2012). The document contains many diagrams, with the workflow for the creation of a data asset being of most interest here (Fig. 5).

A special edition of ZooKeys, with twelve papers, was published in July 2012 (eds. Vladimir Blagoderov and Vincent Smith): ‘No specimen left behind: mass digitization of natural history collections.’ Haston et al. (2012) detail work done at the Royal Botanic Garden Edinburgh (RBGE), which “given the irregular nature of much of the funding available for digitisation” favoured “digitisation workflows [based] on a modular system which has the potential to be scaled up as funding becomes available” (p. 94). The workflow (Fig. 6) is of particular interest for its splitting of the specimen and data workflows, with the intent of separating the imaging from the data capture. Their text is almost unique in its acknowledgement of the resourcing impost of subsequent data management. Tegelberg et al. (2012) describe a commercial digitisation service (previously named Digitarium, now Bioshare Digitization (bioshare.com)), including a conveyor belt system for faster handling. Their workflow (Fig. 7) differs from others in its relationship to a ‘customer’, which is understood to refer to herbarium collections, it is included here as the workflow does cover elements of interest. Nelson et al. (2012) describe five task clusters observed as common to the digitising workflows of the varying biological collections (within USA) surveyed:

1. pre-digitization curation and staging,
2. specimen image capture,
3. specimen image processing,
4. electronic data capture, and
5. georeferencing locality descriptions;

ordered into three workflows (Fig. 8). Tulig et al. (2012) describe digitisation efforts at the New York Botanical Garden (NYBG) and demonstrates how they have changed over time. As expected of a maturing process, the complexity of the three workflows increases as the process evolves (Fig. 9). Four years later, Thiers et al. (2016) provide an update on this work at NYBG in which workflow is textually described without workflow diagrams.

In the 2013 paper by Barber et al. (2013) the focus is on the development of technology (SALIX: Semi-Automatic Label Information eXtraction) to extract data from a specimen label at Arizona State University (ASU) Herbarium. The workflow within which this specific work fits is also shared (Fig. 10).

The Phillips et al. (2014) review of the digitisation practices of eighteen partners of the SYNTHESYS3 initiative (synthesys.info) is an excellent survey of the wider natural history collection landscape. While no diagrams were in the report (likely due to the diverse collection types covered), the authors highlight that “the most common order in which tasks are performed (were all the tasks to be performed)” were those listed below (as per the original text, p. 13):

1. Selection
2. Transfer of material from one area to another
3. Application of barcodes and “other” tasks
4. Full (or partial) data capture
5. Imaging
6. Records management
7. Returning material
8. Quality assurance

The authors also note that “the vast majority of institutions included a full data capture step within their digitisation workflows” (p. 13), and that “the majority of institutions are still capturing full specimen metadata prior to the imaging step in their main digitisation workflows” (p. 18).

Nelson et al. (2015) then refine the 2012 iDigBio information (Nelson et al. 2012) to focus on flat sheets and draw two new visions of possible digitisation workflows: (a) object to data to image, Fig. 11; and (b) object to image to data, Fig. 12. Tasks are separated into fourteen modules (listed below, as per text in Box 1, p. 3) with very detailed descriptions of the considerations involved with each task; available via GitHub (github.com/iDigBioWorkflows/FlatSheetsDigitizationWorkflows).

1. Pre-digitisation curation
2. Selecting components for an imaging station
3. Imaging station setup, camera / copy stand
4. Imaging station setup, light box
5. Imaging station setup, scanner
6. Imaging
7. Image processing
8. Organising and implementing a public participation imaging blitz
9. Image archiving
10. Selecting a database
11. Data capture
12. Organising and implementing a public participation transcription blitz
13. Georeferencing
14. Proactive digitisation

The authors caution that “broad disparities in digitization starting points, institutional infrastructure, curatorial practices, and precise digitization tasks among and within these groups focused on different taxa make the development of a single, consensus object-to-digitized-content workflow impractical” (p. 2).

Around this time the literature appears to shift from the workflows – settling on the object-data-image (Fig. 11) and ­object-image-data (Fig. 12) options (per Nelson et al. 2015) – to focus on the optimisation of specific elements. Many take the image as a starting point and then focus on harvesting data from the specimens or labels (e.g., Triki et al. (2020), White et al. (2020), Owen et al. (2019), Kirchhoff et al. (2018), Haston et al. (2015), Drinkwater et al. (2014), Mononen et al. (2014)); some include streamlining imaging process (e.g., Sweeney et al. 2018, Tegelberg et al. 2014); and others extend the use of the digital images (e.g., Corney et al. (2018), Carranza-Rojas et al. (2017), Unger et al. (2016))

While the excellent paper of Nieva de la Hidalga et al. (2020) focusses on automating image manipulation at the Meise Botanic Garden (MGB), it includes a digitisation workflow (Fig. 13) and detailed task listing. The modular design of that workflow allows for the integration of in-house and outsourced digitisation efforts and focuses on achieving consistent quality standards and recording an audit trail to ensure scalable image production. While a digital image is again the starting point in Walton et al. (2020), the increased modularity of the digitisation process is evident (Fig. 14).

With a few notable exceptions, most digitisation workflows available in the literature are generalised, and understandably so, for this facilitates their uptake and adaptation. Surveys or applications of workflows tend to focus on large institutions or conglomerates; and large-scale processes for flat sheet herbarium specimens appear to have converged on conveyor belt systems with manual transcription, such as that used by the National Herbarium of New South Wales in Australia which incorporates Picturae (digitisation) and Alembo (transcription) (Cox 2022).

In this paper we present the detailed workflow paths for digitisation of MELU collection, as a real-life case study and contribution to the literature for medium-sized institutions. Though, echoing Nelson et al. (2015) (p.2), it is worth noting advice from Barkworth and Murrell (2012) (p. 60):

There is no best approach for digitizing herbaria; there are multiple effective approaches. The needs and resources of large research herbaria with multiple type specimens and collections from many countries and multiple centuries differ from those of small herbaria serving a forest district or a teaching institution. … Adopting theoretically suboptimal procedures for digitization may be the best procedure if the resources needed for adopting a better procedure are not available.

The MELU digitisation workflow outline (Fig. 15) summarises significant detail for the digitisation streams, and groups steps into modules. Using the descriptions from Nelson et al. (2015) (Fig. 11 and Fig. 12 respectively), Stream 1 (left) is effectively an object→data→high-res-image pathway, and Stream 2 (middle) an object→image→data pathway. At MELU, Stream 1 has historically been the established pathway for digitising specimens that were accessioned in the collection but had not yet been digitally catalogued. This stream facilitated digitisation of specimens that were previously recorded in the print catalogue only. Stream 2 was established to increase efficiency of data transcription within the collection. This stream was developed during 2020-2021, within the 12-month period immediately prior to the onset of the Covid-19 pandemic and became the primary pathway for specimen digitisation during the lockdowns enforced as part of the COVID-19 pandemic response in Victoria, Australia. Each stream is further discussed below.

The details for Stream 1 digitisation workflow are included below (Fig. 17; using shapes defined in Fig. 16). The first task (1A) involves: collection management and curation tasks of assigning an accession number to each specimen; setting up a data collection template (using Specify Workbench, which is an environment in Specify allowing for temporary holding of data prior to formal uploading into the CMS); the most recent application of a taxon name on the specimen is checked to ensure that the taxon name is current (as recognised in the Australian Plant Census; anbg.gov.au/cpbr/program/hc/hc-APC.html); all data is input into the collection template. Data capture into the Specify Workbench enables the data entry person to receive feedback about data quality from the functionality built into Specify (including indication of taxa, agents, and locations that are not currently recorded). As all data are captured in Specify, Stream 1 has the benefit of not requiring the transfer of files among software, therefore limiting version control requirements and the risks inherent with these file transfers.

Next, (1B; Fig. 17) quality assurance (QA) of data in the Specify Workbench is undertaken by the curation officer or curator (note this is not undertaken by the trained volunteer entering the data), prior to uploading to the Specify CMS. Labels are then printed (1C; Fig. 17) and either affixed to the specimen or paperclipped if specimen conservation is required prior to imaging.

There may be a gap in time between the data collection in Stream 1 and taking the high-resolution image/s (section H; Fig. 18). Steps H1 and H2 include the image capture and quality control (MELU uses Leaf Aptus imaging hardware and Leaf Capture software), and actions are required to archive the image/s for backup and then to connect the image/s to the data in Specify CMS. Following this, the next step (step P; Fig. 18) is the scripted provision of data/images in Specify to the online MELU collection portal and ALA.

The initial task (2A) in Stream 2 (Fig. 19) involves steps similar to 1A, but with only skeletal data being collected (MELU accession number, the current taxon name, the number of specimen sheets associated with the collection object, and the number of images for each collection object). The data is collected using a .CSV file (not Specify Workbench), and the current taxonomic name is only checked and recorded if transcription will be completed by volunteers who are not affiliated with MELU (e.g., citizen scientists).

Next, (2B; Fig. 19) data is temporarily input into a test (duplicate) Specify CMS, to enable a label containing the MELU accession number, or any annotation labels to be printed, if required. Then (2C; Fig. 19) an image is taken, using a fixed camera mount, mobile phone camera, plus wireless camera remote control. The time required for lower-resolution image generation is significantly less than that required for generation of the high-resolution images. While these initial modules require physical interaction with the collection specimens, it is no more than required for that of Stream 1. The images are then reformatted into the file size required by the software used to view the images for transcription.

The next steps are to set up for and engage in the manual transcription of the data from the specimen images (modules 2D and 2E, Fig. 20). After module 2C, MELU’s workflow can be undertaken entirely remotely if necessary or preferred (right side stream), either by internal (trained volunteers and students) or external (DigiVol; digivol.ala.org.au) transcribers. The potential for remote completion of workflow steps increases flexibility within this stream and reduces the number of times specimens are handled, the latter potentially reducing the average overall time required for progression through this transcription stream (recall the observation in Tann and Flemons 2008). The option of transcribing the data from the specimens in person in the herbarium remains because experience and surveys suggest “a clear preference … for working with physical specimens” (Drinkwater et al. 2014, p. 27). Where data are transcribed externally (i.e., by DigiVol citizen scientists), the data file (a .CSV file) is downloaded from the online interface through which the data are captured for upload into Specify CMS.

The penultimate task in Stream 2 is 2F (Fig. 20), quality assurance and data upload. Quality control of the specimen data can reference the specimen images, so the data can be uploaded into Specify remotely. The final task in this stream (2G; Fig. 20) is printing and attaching any new labels that are required, either full institutional labels where specimens were lacking that label, or annotation labels for accepted name changes. In the remote pathway, the taxon name is checked and an annotation label printed and affixed prior to imaging (2B). If the transcription was undertaken by internally trained volunteers or students the taxon name was checked immediately prior to data transcription (2C), therefore annotation labels are printed and affixed after that step.

At the end of Stream 2, all collection data are uploaded into Specify CMS. The lower-resolution image/s are retained in collection records but are not uploaded into the CMS and are not publicly shared. In this way, this stream does not always immediately complete the entire digitisation workflow, as high-resolution images may not have yet been generated. The final decision regarding generation of high-resolution images suitable for online sharing is made based on collection curation priorities and staff/volunteer availability.

Increasingly, data is entering herbarium databases soon after collection via digital records kept by the collector. Stream 3 (Fig. 21) outlines this emerging protocol; noting this is for material new to the MELU collection and MELU advises the collector about data format and structure. Data is entered by the collector into a .CSV template provided by the digitisation coordinator. This is more commonly done by referencing field notebooks, though the possibility of using dried specimens is included in the workflow. Following data collection, the file is ready for future integration into Stream 1 at step 1B (Fig. 17).

Many digitisation workflow diagrams observed in the literature do not explicitly distinguish manual or human-mediated versus automated or scripted workflow steps. The MELU outline (and subsequent detailed maps) makes explicit the human-mediated steps in digitisation workflow, particularly the regular and iterative handling of the physical botanical specimens. In the outline map (Fig. 15), human-mediated tasks are highlighted by being marked with a green icon (with a 'p' or 'o' in the box), and by the green hexagonal and upside-down wedge (and scattered rectangle) elements in the detailed maps (Figs 17, 18, 19, 20, 21). This human element is made even more apparent by the proliferation of the green elements in the more complex detailed maps, from which the outline map was distilled. This can be fully appreciated when task text is removed from the overall detailed map for streams 1 and 2, and the green elements of specimen handling are highlighted (Fig. 22).

Specimen handling events in the digitisation workflow include tasks such as selection and collation of specimens for digitisation, generation of lower- or high-resolution images, affixing label, and refiling specimens. The placement of these human-mediated steps, inferred, but rarely annotated as such in workflow landscapes, has significant implications in terms of efficiency in the digitisation workflows. Specimen handling tasks are typically labour-intensive steps and many, such as specimen selection and refiling specimens into the collection cannot be eliminated. However, reducing the number of times the specimens are handled during the digitisation curation workflow, for example by reducing the requirement for (re-)sorting or (re-)filing specimens, introduces efficiencies and time savings to the overall workflow. Good examples of the timesaving offered by reduction/s in specimen handling are the transition from a data-to-image to an image-to-data workflow and reliance on lower-resolution images, rather than high-resolution images, as a source for specimen transcription. The time requirement for generation of lower-resolution images is significantly less than that required for the generation of high-resolution images. Generation of digital images, albeit lower-resolution images, early in the digitisation workflow, enables sorting and searching of digital images, which can result in a significant timesaving over sorting and searching physical specimens. Generation of high-resolution specimen images is typically still a desired component of the digitisation workflow. Where lower-resolution specimen images are used for data curation, decisions regarding the allocation of resources and time to the generation of high-resolution specimen images can follow collection imaging priorities rather than data capture priorities. Where curation and digitisation resources are finite, as is the case in all collections, such efficiencies in the workflow can release staff time for other essential curation tasks.

The coronavirus (COVID) pandemic, during 2020 and 2021, provided the impetus for a pivot to 'remote' curation and taxonomic work (e.g., Phang et al. (2022)) that few would have anticipated was possible prior to that event. As a result, collection managers were required to identify those curation tasks that could be conducted remotely and, conversely, those that could only be completed when workers were on-site in the collection. Both technology and protocol changes were amongst significant adaptations to curation tasks and workflows, in collections globally to ensure ongoing productivity while collection staff were working primarily remotely (Baker 2020). At MELU, some of those changes were in the planning stages and were fast-tracked in response to the immediate need to increase the capacity for remote curation work. While the CMS for MELU was not cloud-based, placement onto a virtual machine made remote access to collection records possible for Herbarium staff and student volunteers. The increased use of an image-to-data workflow (Stream 2) also enabled remote curation work. During the extended lockdowns enforced in Victoria, Australia in 2020 and 2021, a single staff member could safely work in person in the collection, completing specimen pre-processing, generating the lower-resolution images, and undertaking the initial (shorter) preparation steps necessary to enable subsequent, potentially remote, transcription efforts. In additional to transcription of primary label data, georeference data can be generated, and data quality can be checked and errors corrected based on available images (Baker 2020). In this way, MELU staff and student volunteers, like their colleagues in collections globally, were able to continue their digitisation work during what could have been, and was for some, a time of reduced productivity.

Any diagram, by its nature, is an abstraction of reality and may appear to imply that work simply flows from one task to the next. Representations necessarily omit detail and seem to suggest that connections are seamless. But it can be these very transitions, between modules and between tasks within modules, that can be the most difficult part of a digitisation project, for they often involve data format transformations, transfers between storage locations or software, which are time-consuming and may be points of highest risk for data loss. Additionally, workflows that require multiple format conversions between input and output data files are often not very resilient to workflow adjustments, which can limit the ease of maintenance and evolution of these workflows over time (Dou et al. 2012).

Detailed workflow maps permit the inspection of the technology required for each digitisation task and, subsequently, requirements for data transfer among software and storage or archival location/s. For example, the complexity of the CMS infrastructure, the software components, and the resulting curation steps involved in the MELU workflow around tasks H1 and H2 are detailed in Fig. 23. For MELU, like many herbaria, the digitisation workflow works within a complex data management system infrastructure; spanning individual objects in the collection, the digital records held in the CMS, image storage on physical hard drives, image servers and archives, and including interactions (data provision) with institutional and external repositories. Understanding the component technologies, and their input and output requirements, are essential for streamlining the digitisation workflow to ensure format cross-compatibility among software in the workflow series and that standards for storage and archiving of images and their metadata are being met. While this example figure is specific to technology MELU currently uses, it is hoped that other collections can use the workflow maps to facilitate identification of tools enabling each task. Such clarity presents the opportunity to review where the process may benefit from software and/or hardware substitution, because best practices require that these workflows are both sustainable for maintenance in perpetuity, including for database management and migration over time (Thomer et al. 2019).

What is evident from these landscape maps is the complexity of data handling requirements for all digitisation workflows. Comprehensively mapped workflows, as provided here, clearly illustrate the complexity and labour-intensity of managing not only the collection objects and their primary data, but also any derived objects and metadata, while also maintaining the links among these entities (e.g., “digital-extended specimens” (Hardisty et al. 2022)). Best practice for voucher-enabled biodiversity research data includes the retention of links among specimens in herbaria or museums and third-party repositories, to maintain an accurate taxonomic context for the derived research data. These links must be actively curated in perpetuity to maintain the significant value of the derived research data for reuse. While the workflow for born-digital specimen data (Nelson et al. 2012, p. 20), is more efficient as transcription of these data is not required, the challenge for these born-digital data is in subsequent required curation of the links among these specimens and their derived objects (e.g., microscope slides created from tissue destructively sampled from specimens or resulting micrographs of these preparations) or their derived data (e.g., DNA sequence or trait measurement data). Well documented and intuitive curation workflows are necessary for the efficient mobilisation and management of specimen-associated digital objects or data to achieve the Digital Extended Specimen as a fully interconnected network of digital objects on the internet (Hardisty et al. 2022).

While this case-study has focused on digitisation of specimens and their primary data, well-documented digitisation landscapes such as those presented here can provide the necessary framework for subsequent mapping of workflow/s for digitisation of derived specimen objects and for in-house curation of specimen-derived research-associated data that are typically provided by researchers to third-party global repositories (e.g., GenBank, MorphoBank).

Data management represents an increasingly labour-intensive task for curators, which is a challenge for all collections, and in particular for small and medium size collections with limited curation staff. Mapping what currently remains a predominantly manual workflow enabled identification of the steps that hold potential for automation (e.g., H1; the scripted upload of images into the CMS and into the image storage archive). The increased availability of openly available workflows and software architectures with standardised interfaces that meet the information technology and archive requirements of natural history collections and that are customisable to meet the diverse needs of collections (e.g., Kurator, Dou et al. 2012; StanDAP-Herb, Kirchhoff et al. 2018), will be a valuable resource for small and medium size collections to potentially increase the ease and efficiency of these data management requirements.

For the time being MELU still relies on manual transcription of data from specimen sheets. As has already been noted, even the largest organisations (with arguably more funding) also appear to continue to invest in manual transcription. The ongoing engagement of citizen scientists for ‘remote’ elements of Stream 2 has been important for expanding opportunities for ongoing digitisation outputs at MELU. This approach is by no means intended as a replacement for mass-digitisation pathways seen in large institutions, but it is a ‘lightweight’ approach to transcribing from an image and making progress toward digitisation goals. It is suitable for small and medium collections because of this simplicity, and it is cost-effective to apply with minimal tool or technology changes. MELU is currently exploring what efficiencies may be introduced via machine-learning and -reading, believing that “information extraction from specimen labels [is] among the digitization workflow activities which can benefit from greater automation” (de la Hidalga et al. 2022, p. 2). These investigations include optical character recognition (OCR) and applying object-detection machine-learning models to read label data (Thompson et al., forthcoming; Turnbull et al., in draft).

Integration of the archival requirements of the vast amounts of data and digital files that are the result of digitisation efforts is also necessary to ensure these resources are curated and accessible across the research data life cycle. While current efforts continue to focus on the generation of the first set of digital files associated with physical specimens, ongoing study of those physical specimens may require the addition of an annotation label, for example, to denote a change in the taxonomic identity of the specimen or the sampling and removal of material from the specimen. Version control of images and data becomes even more complex when both specimen data and images are shared with global repositories. New functionality will be required to enable curation and version control of digital extended specimens given the dispersion of objects and their data into multiple databases and repositories and ensure that current and consistent versions of those objects are accessible for curation and use globally (Hedrick et al. 2020).