Quantification of plant trait data from herbarium scans in the DiSSCo Research Infrastructure

Research Ideas and Outcomes 11: e160367 doi: 10.3897/rio.11.e160367 Reviewed v 1 Research Article Quantification of plant trait data from herbarium scans in the DiSSCo Research Infrastructure  Rajapreethi Rajendran , Claus Weiland , Jonas Grieb , Soulaine Theocharides , Sam Leeflang , Wouter Addink , Sharif Islam ‡ Senckenberg – Leibniz Institution for Biodiversity and Earth System Research, Frankfurt am Main, Germany § Naturalis Biodiversity Center, Leiden, Netherlands | Distributed System of Scientific Collections - DiSSCo, Leiden, Netherlands ¶ DiSSCo, Leiden, Netherlands Corresponding author: Rajapreethi Rajendran ([email protected]) Academic editor: Laurence Livermore Received: 27 May 2025 | Accepted: 30 Oct 2025 | Published: 16 Dec 2025 Citation: Rajendran R, Weiland C, Grieb J, Theocharides S, Leeflang S, Addink W, Islam S (2025) Quantification of plant trait data from herbarium scans in the DiSSCo Research Infrastructure. Research Ideas and Outcomes 11: e160367. https://doi.org/10.3897/rio.11.e160367 Abstract The Distributed System for Scientific Collections (DiSSCo) is a research infrastructure to integrate European natural science collections (NSCs) digitally. The aim is to facilitate and enhance the access, management and analysis of collection assets in one unified digital collection. The Machine Annotation Services (MAS) are essential components of DiSSCo’s Digital Specimen Architecture (DSArch). These services automate the annotation of digital objects to enable labelling and categorisation of NSC's digital assets. To further advance this, a Machine Learning as a Service (MLaaS) approach was developed which provides researchers with the access to pre-trained machine-learning models for complex tasks, such as instance segmentation and morphological analysis of datasets. MLaaS enhances the DiSSCo’s scalability and flexibility and allows the integration of machine-learning tools in close alignment with the FAIR (Findable, Accessible, Interoperable, Reusable) principles. This study employs DiSSCO's MLaaS framework for the quantitative analysis of herbarium specimens. Machine-learning models, such as Mask R-CNN and YOLO11, are ‡ ‡ ‡ § §,| §,| §,¶ © Rajendran R et al. This is an open access article distributed under the terms of the Creative Commons Attribution License (CC BY 4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. comparatively applied to detect and generate the pixel-level masks of plant organs in herbarium sheets. Subsequently, these models are used to reconstruct the scale in the herbarium sheet and to calculate the surface area of identified plant organs. The determination of quantitative characteristics of plant specimens, such as measuring leaf area or the timestamp of the floral transition, opens up herbarium data for reuse in the large prognosis platforms currently developed in the framework of the Common European Data Spaces. In this way, plant trait data mobilised from natural science collections can improve the predictive capability of the vegetation model components of climate-related data spaces. Keywords Digital Specimen Architecture, plant organ detection, quantitative traits, deep learning, DiSSCo, image processing, instance segmentation, Mask R-CNN, YOLO11, Common European Data Spaces Introduction The loss of biodiversity in the anthropocene, intensified by climate change, has a significant impact on human societies by reducing the benefits - designated as Ecosystem Services - that humans derive from ecosystems and environment (Pörtner et al. 2021). To address this critical challenge, natural science collections, including in particular their enriched and annotated digital representations, have a pivotal role for assessment and analysis of the current and future biodiversity loss by providing the fundamental baseline data collection that reflects the actual and past dynamics of marine, freshwater and terrestrial biodiversity. The Distributed System of Scientific Collections (DiSSCo) is a European Research Infrastructure (RI) encompassing over 300 collecting institutions (DiSSCo 2025). A major aim of DiSSCo’s infrastructure is thus to open up specimen data and make it widely reusable in compliance with the FAIR Principles (Koureas et al. 2023), preparing data for both discovery by humans and for autonomous processing by machines (i.e. machine-actionability, Jacobsen et al. (2020)). To achieve this machine-actionability, DiSSCo developed its core data model, the Digital Specimen, in close alignment with the approach of FAIR Digital Objects (FDOs) to represent the physical specimen also in the digital domain (Islam et al. 2020, Hardisty et al. 2022b): A Digital Specimen encompasses or persistently links to all information artefacts (Ceuster and Smith 2015), which are about this physical specimen, such as sequence data, images, chemical measurements or taxonomic determinations. In this way, an object-centred and machine-interpretable representation of a specimen and its connections is realised, which makes independent operations on this digital object by 2Rajendran R et al machines possible, for example, machine-learning-based annotation of scanned images and extraction of associated data such as information on labels (Hardisty et al. 2022). The wider objective of establishing a “Machine Learning as a Service” (MLaaS, Grieb et al. (2021)) framework for DiSSCo is to mobilise traceable geoand biodiversity data for the large analysis and prognosis infrastructures set up currently in the context of the socalled Common European Data Spaces (CEDS, Scerri et al. (2022)). The CEDS are European Union-led initiatives to enable secure, interoperable data sharing across sectors and borders to enable research, innovation and policy-making. Data Spaces related to the European Green Deal are particularly suited for the integration of biodiversity data, involving the EU-flagship initiative Destination Earth and B-Cubed - Biodiversity Building Blocks for Policy (Penninga et al. 2021, Groom et al. 2023, Puechmaille et al. 2025). A closer alignment of the natural science collections with the CEDS could thus significantly increase discoverability, availability and subsequent reuse of collection data. Building on that, the objective of the present study is the functional integration of a Machine Annotation Service (MAS) that enables quantitative determinations of morphological parameters into DiSSCo’s Digital Specimen infrastructure (DSArch, Leeflang et al. (2022)). The paper is further organised as follows: In the next section, we outline the components developed for the MAS. Afterwards, we present the results achieved using the framework with Senckenberg – Leibniz Institution for Biodiversity and Earth System Research's herbarium collection. The final section concludes the paper detailing directions for further developments. Methods Model selection Building on previous research on detecting and annotating the plant organs from digitised herbarium scans (Younis et al. 2018, Younis et al. 2020b), we present now an extended approach involving an advanced segmentation technique to facilitate detailed quantitative analysis of morphological traits. The aim of the segmentation method is to subdivide images into objects or regions. Fundamentally, there are two types of segmentation: semantic and instance segmentation (Long et al. 2015, He et al. 2017). Semantic segmentation labels pixels with a class without distinguishing them into individual objects, whereas the instance segmentation labels each pixel and differentiates the individual objects of each class. Instance segmentation, used in the present context, enables the model to detect the plant organs and segment each organ individually. This is particularly beneficial for distinguishing closely-positioned or overlapping specimen organs, such as leaves, stems, fruits, seeds and flowers. Within the framework of this study, Mask R-CNN and YOLO11 were chosen as instance segmentation models. These models are based, as further explained below, on different architecture approaches: Mask R-CNN is a two-stage model, while YOLO11 is single stage. In the initial stage of the Mask R-CCN classification, it identifies regions of interest Quantification of plant trait data from herbarium scans in the DiSSCo Research ... 3 in an image and, in the second stage, it utilises local features around these proposed regions to determine contained objects. On the other hand, single-stage detectors, like YOLO, employ a fully convolutional approach to gridded regions of an image for the simultaneous prediction of bounding boxes and segmentation in a single pass (Redmon et al. 2016). Plant organ segmentation dataset preparation As in the aforementioned previous studies (Younis et al. 2020b), the dataset used for training plant organ segmentation consists of 652 images of herbarium scans from the Muséum national d’Histoire naturelle's (MNHN) vascular plant collection (Le Bras et al. 2017, Younis et al. 2020a). These images are openly accessible through the Global Biodiversity Information Facility (GBIF) portal (MNHN and Chagnoux 2020). Out of 652 images, 497 images were used for training and 155 images were used for testing. As shown in Table 1, the images were annotated into six different categories: leaf, stem, flower, fruit, root and seed. These images were previously annotated for object detection in plant organs. The training subset has 15486 annotations and the testing subset has 4137 annotations. Category Training subset (497 images) Testing subset (155 images) Complete dataset (652 images) Leaf 7865 2051 9916 Stem 3315 961 4276 Flower 3179 763 3942 Fruit 1045 296 1341 Root 78 60 138 Seed 4 6 10 Total 15486 4137 19623 For the instance segmentation task, the annotations from the previous study were further refined to generate detailed pixel-level masks for each plant organ using the Segment Anything Model (SAM, Kirillov et al. (2023)). The files containing the annotations were parsed to extract the bounding box information, which was then passed to the SAM model along with the corresponding original images. The SAM model iteratively segmented the plant organs for each bounding box and the resulting segmentations were combined for each image. Next, the masks were used to prepare instance segmentation annotations by uniquely identifying and labelling each organ type across all images (Fig. 1). This approach facilitates the reuse of the existing dataset and annotations were produced suitable for MASK R-CNN training. Additionally, Table 1. The number of annotated bounding boxes and segmentation masks for each plant organ category is presented for both the training and testing subsets. 4Rajendran R et al the dataset was reformatted to comply with YOLO11's input specifications and allows precise segmentation of individual features in the scans and supports more detailed morphological analysis. Scale training dataset preparation The model was further independently trained to detect the scale in the digitised herbarium sheets and to calculate the surface area of plant organs in the digital herbarium sheet. The training dataset consists of 163 annotated images. Amongst these, 32 images are sourced from the Senckenberg herbarium dataset (Younis et al. 2020a), 24 images from the vascular plants collection at the Herbarium of the Muséum national d'Histoire naturelle, Paris (MNHN and Chagnoux 2020) and an additional 107 images from the GBIF to increase variability in the dataset for the training for scale detection. Of the total, 124 images are used for training and 39 for testing. A dataset containing the subset of images included from the Senckenberg Herbarium is available (Rajendran et al. 2025). Figure 1. Mask generated by the SAM (Segment Anything Model) for a herbarium specimen scan of Rubus pottianus H.E. Weber. The figure shows the segmentation mask output produced by the SAM model for a digitised herbarium sheet labelled FR-0030810 (CETAF ID: https:// id.senckenberg.de/object/sesam-353465) from the training dataset (Rajendran et al. 2025).  Quantification of plant trait data from herbarium scans in the DiSSCo Research ... 5 Model training and testing The Mask R-CNN model (He et al. 2018) is trained using PyTorch (Paszke et al. 2019b) to perform instance segmentation on digitised herbarium scans. Out of 652 images, 497 annotated images were used for training and 155 images were used for testing. Data transformations, such as random horizontal and vertical flips, colour jitter, scaling and rotation, were applied during training to introduce variability to improve the model’s generalisation. The Mask R-CNN model used a ResNet-50 backbone to balance feature extraction quality and computational efficiency for the segmentation of high-resolution herbarium images. Training strategies, such as early stopping, were implemented by comparing mean Average Precision (mAP) of the current epoch and previous epoch, involving a patience threshold to avoid overfitting (Ying 2019). Additionally, the model was customised with a modified Region Proposal Network (RPN) anchor generator and adjusted Region Of Interest (ROI) heads to improve performance on the dataset (He et al. 2017). The model performance was evaluated with metrics, such as precision and recall, for object detection and segmentation. To test the model generalisation, inference was conducted on the images from the Frankfurt Senckenberg Herbarium dataset (Younis et al. 2020a). The YOLO 11x-seg (Jocher et al. 2023) segmentation model was trained with the same dataset used for MASK R-CNN training (Jocher et al. 2023). The default input image size for the YOLO model training is 640 × 640 pixels, but in this study, a resolution of 1064 × 1064 pixels was used, since it was the largest resolution that could fit into the GPU memory during training. The training strategies, such as early stopping of training, were applied in the same way as in the previous case. Detection of scale and organ surface area calculation The measurement of surface area of plant organs is an essential component of morphometric analysis in biodiversity studies (Hodač et al. 2024). To calculate the surface area of a plant organ in the herbarium sheet, the scales need to be detected and the surface area in pixels needs to be calculated and converted into absolute values, which are units of measurement such as square millimetres or square centimetres. This is accomplished through a combination of deep learning techniques for scale detection and Optical Character Recognition (OCR) for extraction of numerical values. Scale detection The Mask R-CNN and YOLO11 models were again employed for the detection of scales in the herbarium images. The surface area of plant organs was subsequently calculated. We used 167 images, of them 124 images for training and 43 for testing. 6Rajendran R et al Once the scale was detected by the model, Tesseract OCR was employed to extract the numerical values from the scale present in the images (Smith 2007). Numerical extraction using Tesseract OCR Tesseract (Smith 2007) is a widely used optical character recognition tool to identify text in the images. Image preprocessing techniques, such as rotation of scales to correct orientation, greyscale conversion and bilateral filtering, are applied to enhance the text visibility. Following the preprocessing, the Tesseract OCR was utilised on the scales to extract the numerical values from the scale. The digits with a confidence level above 75% and in the range of numbers from 0 to 10 are filtered and sorted to identify the consecutive sequences. The pixel distances between adjacent digits are calculated and averaged with the detected numbers to determine the number of pixels corresponding to 1 cm. In cases where the consecutive digits are not detected, the closest pair of digits is used to estimate the pixel distance. The character recognition accuracy is computed by comparing the detected digits with a ground truth reference. Once the pixel distances between consecutive digits on the scale have been calculated, the next step is to establish the conversion factor that relates the relative pixel measurements to absolute measurement in centimetres. The pixel-to-centimetre conversion factor is computed as follows: where •Pixel Distance Between Digit A and Digit B is the pixel distance between the detected digits; •Difference Between Digit A and Digit B is the difference between the actual values of the digits. Surface area calculation of detected plant organs With the pixel-to-centimetre conversion factor established, the next step is to calculate the surface area of the detected plant organs. The plant organ surface area is calculated as follows: where •The sum of pixels in the segmented organs refers to the total number of pixels in the segmented organ regions; Quantification of plant trait data from herbarium scans in the DiSSCo Research ... 7 •Calculated pixels present in 1 cm is the number of pixels that represent 1 cm on the image, estimated using the detected scale bar. This additional functionality of calculating surface area supports detailed morphometric analysis of herbarium specimens. Machine Learning as a Service (MLaaS) Both models for plant organ segmentation and surface area calculation are then hosted as a single Machine Annotation Service (MAS) to streamline the process of extracting and analysing the herbarium data within the DiSSCo platform. Fig. 2 further illustrates the process flow and communication between the DiSSCo Digital Specimen architecture and the plant organ segmentation MAS. In DSArch, a service request for a MAS is requested through the user front-end DiSSCover (https://sandbox.dissco.tech) on the digital media. This request adds a message in DiSSCover’s Message Broker, which triggers scheduling of a MAS. The corresponding MLaaS APIs are hosted via Uvicorn (Trylesinski and Christie 2019), a high-performance asynchronous python web server designed for efficient handling of API requests on a remote virtual machine (VM) at Senckenberg. Within the scope of our specific use case, the API receives image URLs as the input from the Message Broker through the HTTP POST requests. The API then adds the message to the processing queue, which forwards the message to a web socket service, hosted on a server for performing plant organ segmentation. Subsequently, the model performs plant organ segmentation and scale detection on the image and sends the output information comprising bounding box coordinates, class labels, confidence scores and area in pixels. Both the pixel-to-centimetre conversion ratio Figure 2. Schematic overview illustrating the information flow between DiSSCo core architecture and the MAS workflow deployed at Senckenberg. (i) Message Broker, which handles asynchronous communication; (ii) MLaaS (Machine Learning as a Service), API hosted via Uvicorn, serving as the inference interface; and (iii) Machine Annotation Service (MAS) modules responsible for task orchestration. The architecture diagram highlights how herbarium image data and metadata are processed, annotated and returned to the DiSSCo system in a scalable and modular fashion.  8Rajendran R et al and an area calculation in cm² are returned to the Uvicorn API, which relays the results to the DiSSCo infrastructure. These processed data are then structured into an annotation event complying to DiSSCo's open Digital Specimen (openDS, Addink and Hardisty (2020)) specification. The message is published back to DiSSCo’s Core architecture by placing it on the Message Broker, making it in this way available for downstream applications. Fig. 3 provides a detailed visualisation of the annotations which allows users to explore the extracted information by hovering over specific segments. This includes details, such as the type of organ, confidence score, segmented polygon coordinates, area in pixels, pixel-to-centimetre conversion ratios and calculated areas in square centimetres for each segmented region. Fig. 3 demonstrates the integration of MAS within the DiSSCover platform highlighting its ability to generate enriched and standardised annotations in compliance with openDS . Results MASK R-CNN Mask R-CNN was employed on 203 herbarium images from the Senckenberg collection (Younis et al. 2020a). Fig. 4 shows an example of object detection inference performed by Mask RCNN on a herbarium sheet. Figure 3. Annotated herbarium specimen sheet processed through the DiSSCover platform. The figure presents an annotated herbarium image processed using the DiSSCover pipeline, which includes detection of plant organs and segmentation. The visual overlays include bounding boxes, class labels, confidence scores from the prediction model, area in pixels, pixel-tocentimetre conversions and polygon coordinates for each detected organ.  Quantification of plant trait data from herbarium scans in the DiSSCo Research ... 9 Acknowledgements We thank Anke Penzlin, Andreas Allspach, Moritz Sonnewald, Alexander Knorrn, André Freiwald, Kristina Hopf, Stefan Dressler (†) and Marco Schmidt for the provision of data from Senckenberg's collections. Funding program Funded by the European Union. Views and opinions expressed are, however, those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Executive Agency. Neither the European Union nor the granting authority can be held responsible for them. DiSSCo Transition grant agreement ID: 101130121, https://doi.org/10.3030/101130121. Conflicts of interest The authors have declared that no competing interests exist. References • Addink W, Hardisty A (2020) ‘openDS’ – Progress on the New Standard for Digital Specimens. Biodiversity Information Science and Standards 4 https://doi.org/10.3897/ biss.4.59338 • Ceuster W, Smith B (2015) Aboutness: towards foundations for the Information Artifact Ontology. In: Ceuster W, Smith B (Eds) International Conference on Biomedical Ontology. URL: https://api.semanticscholar.org/CorpusID:15812859 • Díaz S, Kattge J, Cornelissen JC, Wright I, Lavorel S, Dray S, Reu B, Kleyer M, Wirth C, Colin Prentice I, Garnier E, Bönisch G, Westoby M, Poorter H, Reich P, Moles A, Dickie J, Gillison A, Zanne A, Chave J, Joseph Wright S, Sheremet’ev S, Jactel H, Baraloto C, Cerabolini B, Pierce S, Shipley B, Kirkup D, Casanoves F, Joswig J, Günther A, Falczuk V, Rüger N, Mahecha M, Gorné L (2015) The global spectrum of plant form and function. Nature 529 (7585): 167‑171. https://doi.org/10.1038/nature16489 • DiSSCo (2025) Distributed System of Scientific Collections. https://www.dissco.eu/. Accessed on: 2025-8-01. • Grieb J, Weiland C, Hardisty A, Addink W, Islam S, Younis S, Schmidt M (2021) Machine Learning as a Service for DiSSCo’s Digital Specimen Architecture. Biodiversity Information Science and Standards 5 https://doi.org/10.3897/biss.5.75634 • Groom Q, Abraham L, Adriaens T, Breugelmans L, Clarke D, Fernández M, Hendrickx L, Hui C, Kumschick S, Martini M, McGeoch M, Metodiev T, Miller J, Oldoni D, Pereira H, Preda C, Robertson T, Rocchini D, Seebens H, Teixeira H, Trekels M, Wilson JR, Yovcheva N, Zengeya T, Desmet P (2023) B-Cubed: Leveraging Analysis-Ready Biodiversity Datasets and Cloud Computing for Timely and Actionable Biodiversity Monitoring. Biodiversity Information Science and Standards 7 https://doi.org/10.3897/biss. 7.110734 16 Rajendran R et al • Hardisty A, Brack P, Goble C, Livermore L, Scott B, Groom Q, Owen S, Soiland-Reyes S (2022a) The Specimen Data Refinery: A Canonical Workflow Framework and FAIR Digital Object Approach to Speeding up Digital Mobilisation of Natural History Collections. Data Intelligence 4 (2): 320‑341. https://doi.org/10.1162/dint_a_00134 • Hardisty A, Ellwood E, Nelson G, Zimkus B, Buschbom J, Addink W, Rabeler R, Bates J, Bentley A, José AB, Hansen S, Macklin J, Mast A, Miller J, Monfils A, Paul D, Wallis E, Webster M (2022b) Digital Extended Specimens. Enabling an Extensible Network of Biodiversity Data Records as Integrated Digital Objects on the Internet, BioScience 72 (ue 10). https://doi.org/10.1093/biosci/biac060 • He K, Gkioxari G, Dollar P, Girshick R (2017) Mask R-CNN. 2017 IEEE International Conference on Computer Vision (ICCV) https://doi.org/10.1109/iccv.2017.322 • He K, Gkioxari G, Dollar P, Girshick R (2018) Mask R-CNN. IEEE transactions on pattern analysis and machine intelligence 42 (2): 386‑397. https://doi.org/10.1109/TPAMI. 2018.2844175 • Hodač L, Karbstein K, Kösters L, Rzanny M, Wittich HC, Boho D, Šubrt D, Mäder P, Wäldchen J (2024) Deep learning to capture leaf shape in plant images: Validation by geometric morphometrics. The Plant Journal 120 (4): 1343‑1357. https://doi.org/10.1111/ tpj.17053 • Islam S, Hardisty A, Addink W, Weiland C, Glöckler F (2020) Incorporating RDA Outputs in the Design of a European Research Infrastructure for Natural Science Collections. Data Science Journal 19 https://doi.org/10.5334/dsj-2020-050 • Jacobsen A, de Miranda Azevedo R, Juty N, Batista D, Coles S, Cornet R, Courtot M, Crosas M, Dumontier M, Evelo C, Goble C, Guizzardi G, Hansen KK, Hasnain A, Hettne K, Heringa J, Hooft RW, Imming M, Jeffery K, Kaliyaperumal R, Kersloot M, Kirkpatrick C, Kuhn T, Labastida I, Magagna B, McQuilton P, Meyers N, Montesanti A, van Reisen M, Rocca-Serra P, Pergl R, Sansone S, da Silva Santos LOB, Schneider J, Strawn G, Thompson M, Waagmeester A, Weigel T, Wilkinson M, Willighagen E, Wittenburg P, Roos M, Mons B, Schultes E (2020) FAIR Principles: Interpretations and Implementation Considerations. Data Intelligence 2: 10‑29. https://doi.org/10.1162/dint_r_00024 • Jocher G, Qiu J, Chaurasia A (2023) Ultralytics YOLO. Version 8.0.0. URL: https:// github.com/ultralytics/ultralytics • Kim E, Lee J, Lee J, Lee J, Choo J (2021) Learning Debiased Representation via Disentangled Feature Augmentation. Neural Information Processing Systems https:// doi.org/10.48550/arXiv.2107.01372 • Kirillov A, Mintun E, Ravi N, Mao H, Rolland C, Gustafson L, Xiao T, Whitehead S, Berg A, Lo W, Dollár P, Girshick R (2023) Segment Anything. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) https://doi.org/10.1109/iccv51070.2023.00371 • Koureas D, Livermore L, Alonso E, Addink W, Alves MJ, Casino A, Curral L, Enghoff H, Guiraud M, Hardy H, Hoffmann J, Landel S, Paleco C, Petersen M, Scory S, Smith V, Weiland C, Wesche K, Woodburn M (2023) DiSSCo Prepare Project: Increasing the Implementation Readiness Levels of the European Research Infrastructure. Research Ideas and Outcomes 9 https://doi.org/10.3897/rio.9.e113906 • Koureas D, Livermore L, Addink W, Alonso E, Alonso J, Casino A, Dusoulier F, Ferreira V, Grieb J, Groom Q, Islam S, Kõljalg U, Lymer G, Marhold K, Paleco C, Pijls S, Scory S, Scott B, Weiland C, Worley K (2024) DiSSCo Transition Abridged Grant Proposal. Research Ideas and Outcomes 10 https://doi.org/10.3897/rio.10.e118241 Quantification of plant trait data from herbarium scans in the DiSSCo Research ... 17 • Le Bras G, Pignal M, Jeanson M, Muller S, Aupic C, Carré B, Flament G, Gaudeul M, Gonçalves C, Invernón V, Jabbour F, Lerat E, Lowry P, Offroy B, Pimparé EP, Poncy O, Rouhan G, Haevermans T (2017) The French Muséum national d’histoire naturelle vascular plant herbarium collection dataset. Scientific Data 4 (1). https://doi.org/10.1038/ sdata.2017.16 • Leeflang S, Addink W, Theocharides S (2022) Human and Machine Working Together towards High Quality Specimen Data: Annotation and Curation of the Digital Specimen. Biodiversity Information Science and Standards 6 https://doi.org/10.3897/biss.6.90987 • Long J, Shelhamer E, Darrell T (2015) Fully convolutional networks for semantic segmentation. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) https://doi.org/10.1109/cvpr.2015.7298965 • MNHN, Chagnoux S (2020) The vascular plants collection (P) at the Herbarium of the Muséum national d'Histoire naturelle (MNHN - Paris). 69.186. GBIF. URL: https://doi.org/ 10.15468/nc6rxy • Paszke A, Gross S, Massa F, Lerer A, Bradbury J, Chanan G, Killeen T, Lin Z, Gimelshein N, Antiga L, Desmaison A, Köpf A, Yang E, DeVito Z, Raison M, Tejani A, Chilamkurthy S, Steiner B, Fang L, Bai J, Chintala S (2019) PyTorch: An Imperative Style, High-Performance Deep Learning Library. arXiv https://doi.org/10.48550/arxiv. 1912.01703 • Penninga F, Lutz M, Minghini M, Cetl V, Robbrecht J (2021) INSPIRE, a public sector contribution to the European green deal data space - A vision for the technological evolution of Europe’s spatial data infrastructures for 2030. Publications Office of the European Union URL: https://data.europa.eu/doi/10.2760/8563 • Perez T, Rodriguez J, Mason Heberling J (2020) Herbarium‐based measurements reliably estimate three functional traits. American Journal of Botany 107 (10): 1457‑1464. https://doi.org/10.1002/ajb2.1535 • Poppenwimer T, Mayrose I, DeMalach N (2023) Revising the global biogeography of annual and perennial plants. Nature 624 (7990): 109‑114. https://doi.org/10.1038/ s41586-023-06644-x • Pörtner H, Scholes R, Agard J, Archer E, Arneth A, Bai X, Barnes D, Burrows M, Chan L, Cheung WL(, Diamond S, Donatti C, Duarte C, Eisenhauer N, Foden W, Gasalla M, Handa C, Hickler T, Hoegh-Guldberg O, Ichii K, Jacob U, Insarov G, Kiessling W, Leadley P, Leemans R, Levin L, Lim M, Maharaj S, Managi S, Marquet P, McElwee P, Midgley G, Oberdorff T, Obura D, Osman Elasha B, Pandit R, Pascual U, Pires AF, Popp A, Reyes-García V, Sankaran M, Settele J, Shin Y, Sintayehu D, Smith P, Steiner N, Strassburg B, Sukumar R, Trisos C, Val A, Wu J, Aldrian E, Parmesan C, Pichs-Madruga R, Roberts D, Rogers A, Díaz S, Fischer M, Hashimoto S, Lavorel S, Wu N, Ngo H (2021) Scientific outcome of the IPBES-IPCC co-sponsored workshop on biodiversity and climate change. Zenodo https://doi.org/10.5281/zenodo.4659158 • Puechmaille D, Schick M, Saulyak B, Dillmann M, Wolf L (2025) Destination Earth Data Lake unlocking Big Earth Data processing. EGU General Assembly 2024. https://doi.org/ 10.5194/egusphere-egu24-11115 • Rajendran R, Marco S, Claus W, Jonas G (2025) Plant organ segmentation images and annotations on digitized herbarium scans. Senckenberg Society for Nature Research https://doi.org/10.12761/fj4m-zr97 18 Rajendran R et al • Redmon J, Divvala S, Girshick R, Farhadi A (2016) You Only Look Once: Unified, RealTime Object Detection. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)779‑788. https://doi.org/10.1109/cvpr.2016.91 • Sapkota R, Ahmed D, Karkee M (2024) Comparing YOLOv8 and Mask R-CNN for instance segmentation in complex orchard environments. Artificial Intelligence in Agriculture 13: 84‑99. https://doi.org/10.1016/j.aiia.2024.07.001 • Scerri S, Tuikka T, de Vallejo IL, Curry E (2022) Common European Data Spaces: Challenges and Opportunities. Data Spaces337‑357. https://doi.org/ 10.1007/978-3-030-98636-0_16 • Smith R (2007) An Overview of the Tesseract OCR Engine. Ninth International Conference on Document Analysis and Recognition (ICDAR 2007) Vol 2629‑633. https:// doi.org/10.1109/icdar.2007.4376991 • Trylesinski M, Christie T (2019) Uvicorn. URL: https://github.com/encode/uvicorn • Ying X (2019) An Overview of Overfitting and its Solutions. Journal of Physics: Conference Series 1168 https://doi.org/10.1088/1742-6596/1168/2/022022 • Younis S, Weiland C, Hoehndorf R, Dressler S, Hickler T, Seeger B, Schmidt M (2018) Taxon and trait recognition from digitized herbarium specimens using deep convolutional neural networks. Botany Letters 165: 377‑383. https://doi.org/ 10.1080/23818107.2018.1446357 • Younis S, Schmidt M, Dressler S (2020a) Plant organ detections and annotations on digitized herbarium scans [dataset]. PANGAEA. Release date: 2020-8-27. URL: https:// doi.org/10.1594/PANGAEA.920895 • Younis S, Schmidt M, Weiland C, Dressler S, Seeger B, Hickler T (2020b) Detection and annotation of plant organs from digitised herbarium scans using deep learning. Biodiversity Data Journal 8 https://doi.org/10.3897/bdj.8.e57090 Quantification of plant trait data from herbarium scans in the DiSSCo Research ... 19 Supplementary material Suppl. material 1: Scale and Structure Recognition Beyond Herbarium Specimens Authors: Rajapreethi Rajendran, Jonas Grieb, Claus Weiland, Anke Penzlin, Andreas Allspach, Moritz Sonnewald, Alexander Knorrn, André Freiwald, Kristina Hopf Data type: Image Brief description: The top figure represents Galeoides decadactylus (Bloch, 1795) with Senckenberg catalogue number SMF-39828 (this specimen is currently in process of ingestion into Senckenberg's Collection Management System), the scale was detected and all digits within the scales were successfully recognised. However, no plant organ like structure was detected in the image and, consequently, no area was calculated. The second figure represents Syngnathus acus (Linnaeus, 1758) associated with CETAF ID: https://id.senckenberg.de/object/sesam-1710769. In this case, the specimen was incorrectly identified as the stem due to its structural similarity and the scale was detected and digits '6' and '8' within the scale regions were identified through optical character recognition (OCR). Pixel-wise segmentation is performed and area measurement was successfully computed. In the third figure, the specimen shows Pisa tetraodon (Pennant, 1777), associated with CETAF ID: https://id.senckenberg.de/object/sesam-1710770. In this case, the specimen was incorrectly identified as flower due to its structural similarity and the scale was detected. Although scale was detected, no numerical digits were recognised by OCR and, as a result, area calculation was not conducted. Download file (11.01 MB) 20 Rajendran R et al