Invasive alien plants drive biodiversity loss and ecosystem disruption, costing billions globally. Roadsides faciliate spread, yet large-scale monitoring is difficult. Traditional methods like object detection and instance segmentation require costly, time-consuming annotations, impractical for thousands of species. Citizen science platforms like Pl@ntNet use efficient classification models for global identification, yet they struggle with high-resolution multi-species imagery. In this work we adapted the Pl@ntNet global model, trained on millions of images, to detect multiple invasive species in roadside imagery (Espitalier et al. 2025). Pl@ntNet uses a Vision Transformer trained on 6.6 million images covering 43,683 species. Optimized for single-species classification from fixed-resolution (384×384) close-ups of plant organs, it fails on high-resolution complex, multi-scale roadside scenes where plant size varies. We explored two strategies to bridge this domain gap. VaMIS (Variable Model Input Size), forces the model to process the whole roadside scene at once at a larger but non optimal resolution (1024×768). Multi-scale tiling acts like a scanner, decomposing the high-resolution image into hundreds of overlapping patches at multiple zoom levels. Each patch is analyzed independently at the model's native optimal resolution (384×384). If a species is identified with high confidence in any patch, it is considered present in the full image. Both were tested on the Danish Road Dataset (Dyrmann et al. 2021), containing 14,808 high-resolution (4024×3036) vehicle-mounted images for evaluating the detection of six invasive taxa: Cytisus scoparius , Lupinus polyphyllus , Pastinaca sativa , Reynoutria spp., Rosa rugosa , and Solidago spp. 44% of images contain no invasive species, and few include multiple taxa. Performance was measured using balanced accuracy (average recall of presence/absence classes, mitigating imbalance) and AUC (Area Under the ROC Curve, assessing global discrimination of True vs. False Positive rates). F1 and Jaccard index were used on a multi-label test set of 206 images. The tiling approach demonstrated strong generalization, achieving 84.9% balanced accuracy and 91.6% AUC, despite requiring no additional annotations, whereas VaMIS underperformed. To reduce domain shift and re-optimize internal parameters, essential for VaMIS to handle full scenes, we fine-tuned the whole transformer (VaMIS) or the classification head (tiling). Fine-tuning improved both: VaMIS rose from 66.8 to 91.8% accuracy, and tiling from 84.9 to 92.1%. On the multi-species test set, fine-tuned tiling achieved the best results (F1 = 0.85, Jaccard = 0.76), outperforming fine-tuned VaMIS and baseline CNN or YOLO, both models representing former classification and object detection baselines reproduced from previous work (Dyrmann et al. 2021) for comparison. Performance varied by species, but detection of Rosa rugosa consistently exceeded 99% AUC even without fine-tuning. Tiling was about 34x slower but far less resource-intensive to fine-tune (20 min vs 25 h). These results highlight complementary strategies. Tiling offers a generic, training-free pipeline, while VaMIS is faster but needs fine-tuning and memory. Reusing Pl@ntNet models cuts annotation costs, enabling scalable monitoring. Together, these strategies extend biodiversity monitoring to realistic roadside imagery, aligning with FAIR principles.
The urgent call to reverse biodiversity loss has led to unprecedented efforts to improve biodiversity monitoring, databases, harmonize indicator frameworks, and promote open access to data (Lindenmayer et al. 2011). Global initiatives such as the Kunming-Montreal Global Biodiversity Framework (GBF) (Stephens 2023) emphasize the need for robust, interoperable biodiversity data to support evidence-based decision-making across scales. In this context, more data and better technical standards are often seen as key enablers of effective biodiversity governance. However, the assumption that increasing data volume and standardization will automatically lead to better decisions does not always hold true (e.g., Porto de Albuquerque et al. 2021)—particularly at the local level, where conservation action actually takes place. Local decision-makers often face practical, institutional, personal and epistemic barriers that limit their ability to access, interpret, or apply available data in ways that support their mandates and contexts. Meanwhile, local knowledge and plural values are rarely integrated into higher-level biodiversity data infrastructures. ENABLElocal*1 is a transdisciplinary mixed-method research project working in Czechia, Germany, and Sweden, which explores how biodiversity data is understood, accessed, and used by stakeholders involved in land use and conservation management. It highlights the importance of governance, context, and stakeholder engagement in determining whether biodiversity data becomes usable—and ultimately useful—for transformative action on the ground. The ENABLElocal project aims to bridge the persistent gap between biodiversity data—including monitoring datasets, species records, and environmental variables—and local-scale decision-making in conservation and land use. While large volumes of ecological data are increasingly accessible through national and international infrastructures, their uptake in local governance remains limited (Güntsch et al. 2024). ENABLElocal addresses this across three case studies—Germany, Sweden, and Czechia—each situated in distinct governance systems and biodiversity contexts. At CzechGlobe, the focus is on identifying institutional, social, and technical barriers that hinder effective data use in local decisions. More broadly, the project examines how biodiversity data supply-and-demand align (or not), and how governance conditions shape perceived usability. The project applies a living lab approach to foster iterative knowledge co-production with stakeholders in local biodiversity governance. Such integration of qualitative social science methods into biodiversity monitoring remains rare but essential for making data meaningful and actionable. The first round of workshops (autumn 2024) mapped stakeholder objectives, decision processes, and data types used or needed—from formal datasets to local knowledge. To examine actual data use in context, 39 semi-structured interviews were conducted (summer 2025) with local actors, exploring barriers to access, use, and relevance. A complementary survey, run across all case studies, identified commonly used datasets, unmet needs, and preferences. All activities followed harmonized protocols; interviews were coded using MaxQDA,*2 and user stories helped ground data needs in stakeholder perspectives. Preliminary results point to consistent challenges in using biodiversity data locally. Workshops revealed diverse stakeholder objectives, even within regions—complicating one-size-fits-all tools and reinforcing the need for context-sensitive approaches. Interviews highlight poor communication between data providers and users as a key institutional barrier. Many stakeholders lack spaces to express data needs or feedback on usability. Although analysis is ongoing, findings suggest governance reform may be as critical as technical improvements. Overall, the results underline ENABLElocal’s core insight: more data and better standards do not automatically lead to better decisions. Trust, communication, and participatory processes are essential to make biodiversity data actionable. ENABLElocal shows that the barriers to effective biodiversity data use are often not technical—but social, institutional, and relational. Even as global frameworks promote standardization and open data, local actors struggle to translate datasets into decisions. Addressing this requires more than better tools; it demands better relationships. Our work suggests that communication, co-production, and context-awareness should be treated as core components of biodiversity data infrastructure. Without them, even FAIR (Findability, Accessibility, Interoperability, and Reusability) data may fail to support fair outcomes (Wilkinson et al. 2016). As ENABLElocal continues, upcoming workshops (scheduled for January 2026) and interview analysis will deepen our understanding of how local governance systems can become active co-creators—not just passive users—of biodiversity knowledge.
The digitisation of natural history collections at scale raises a range of logistical and curatorial challenges. One concern is the physical expansion of storage infrastructure. As part of the Distributed System of Scientific Collections UK (Smith et al. 2022), the Natural History Museum London (NHM) will rehouse a significant proportion of its 36 million pinned entomological specimens (~130,000 drawers). Digitisation involves imaging every specimen and attaching a physical, unique, human- and machine-readable identifier (a small card label with a Data Matrix barcode). These barcodes must be readable from above, and their addition during digitisation can increase drawer occupancy if they exceed the existing footprint. Manual estimation of the footprint increase is impractical at this scale, given the range of circumstances associated with different specimens and drawers. To address this, the NHM has developed an AI-driven approach to estimate specimen drawer expansion and support resource planning. The deep learning pipeline can automatically detect drawer objects, referred to here as classes, including specimens, labels, barcodes, notes, unit trays, and drawers, from high-resolution images. Our dataset includes 11,090 digitised pinned Coleoptera drawers from the Index Lot collection (Natural History Museum 2014), representative of typical entomological collections. The AI pipeline supports both "pre-" and "post-" digitisation analysis by calculating the specimen bounding areas, handling overlaps, and estimating the net footprint change. Our object detection models achieved high performance under real-world conditions, with a mean average precision (mAP*1) of 85.45% across all classes. Barcode detection reached 86.63% mAP, while the standard unit tray detection and unit tray type classification model achieved 99.5% accuracy. (An example of the detection outputs is shown in Fig. 1) Three different calculation methods were tested and evaluated to estimate the drawer expansion area, corresponding to three specimen expansion rates for specimens with Data Matrix barcodes attached. Rate 1 (Per-Drawer Area via Polygon Masks*2): calculates expansion rates per drawer by measuring the total area occupied by specimens and their labels, ensuring overlaps are counted only once using polygon masks, as shown in Fig. 2. Rate 2 (Total-Based Expansion Rate): calculates expansion rates by averaging the occupancy areas of barcode-attached specimens across the entire dataset. Rate 3 (Per-Specimen Expansion Rate): averages the expansion rate per specimen using only bounding box*3 areas. Every specimen contributes equally, regardless of its size, to the final value. Rate 1 (Per-Drawer Area via Polygon Masks*2): calculates expansion rates per drawer by measuring the total area occupied by specimens and their labels, ensuring overlaps are counted only once using polygon masks, as shown in Fig. 2. Rate 2 (Total-Based Expansion Rate): calculates expansion rates by averaging the occupancy areas of barcode-attached specimens across the entire dataset. Rate 3 (Per-Specimen Expansion Rate): averages the expansion rate per specimen using only bounding box*3 areas. Every specimen contributes equally, regardless of its size, to the final value. Table 1 compares different expansion rates. Both per-drawer Rate 1 and per-specimen Rate 3 have confidence intervals. Fig. 3 and Fig. 4 show their distributions: Rate 1 is right-skewed, indicating modest area increases from overlap, while Rate 3 is bimodal*4, with a small peak near 0 (large specimens unchanged) and another around 0.3 (barcode additions on smaller specimens). The global total-based Rate 2 is the mean of the ratios without confidence intervals, providing a macro-level view. In addition to estimating drawer expansion and supporting budget planning, this tool has been designed in a modular fashion, allowing individual components to be used independently at different stages of the digitisation and curation workflow. For example, specific models can be used to flag missing labels or barcodes during digitisation, assist in tracking specimen relocation, and support downstream re-curation decisions. Importantly, the pipeline is designed to be reusable across other entomological collections, with all code to be made openly available alongside a forthcoming publication, promoting scalability within the community. Our findings demonstrate that automated spatial analysis not only improves accuracy and speed in collection management but also lays the groundwork for predictive infrastructure modelling across large-scale digitisation efforts.
Herbarium collections serve as irreplaceable archives for taxonomic, ecological, and evolutionary research. Recent advances in deep learning now make it possible to design AI models capable of automatically annotating herbarium specimens (Ariouat et al. 2025, Sklab et al. 2025a, Sklab et al. 2025b), to develop intelligent systems for querying, visualizing, and exploring them. In previous work Sklab et al. 2025a, state-of-the-art Computer Vision (CV) models failed to generalize to unseen specimens, relying on superficial cues like labels and mounting paper rather than learning true plant morphology. Meanwhile, rich botanical texts remain disconnected from specimen images: CV models lack semantic depth, and LLMs extract traits but can’t ground them visually. Vision–Language Models (VLMs) emerge as a natural solution, combining the strengths of both to close this gap . By jointly reasoning across images and text, they have the potential to align visual perception with botanical semantics. However, most existing VLMs are trained on general-purpose datasets and fail to capture the specialized features inherent to herbarium specimens. To address this limitation, we investigate how adapting a VLM to the botanical domain enables it to answer expert-level questions about individual herbarium specimens. We adapted the PaliGemma-2 model (Andreas 2024), leveraging its Visual Question Answering (VQA) capabilities to align large language model knowledge with the observable content of herbarium images. In this setup, the digitized specimen is provided as visual context, and the user interacts with the model through targeted questions. The adapted model grounds its answers in visible features of the specimen—morphology, phenology, and context—rather than relying on generic species-level knowledge. By doing so, the system aims to exploit the semantic richness encoded in modern large language models while constraining it to the biological reality captured in a single herbarium sheet — a snapshot of the plant at the moment it was collected, rather than an idealized description of its entire life history. Our approach comprises three stages (Fig. 1): i) domain-specific dataset construction; ii) parameter-efficient fine-tuning; and iii) evaluation. We curated a dataset from herbarium specimens, combining digitized images with expert-annotations (28 traits). To translate structured trait annotations into a linguistically and semantically rich training corpus, we employed the LLaMA 4 (Meta AI 2025) language model to automatically generate over 50,000 question–answer pairs grounded in botanical semantics. To achieve domain adaptation under limited computational resources, we employed Quantization-aware Low-Rank Adaptation (QLoRA) , which enables efficient fine-tuning by introducing trainable low-rank adapter layers into a quantized pre-trained backbone. We evaluated the fine-tuned model against two baselines, its un-fine-tuned version and the larger PaliGemma-28B model, using five metrics (Fig. 2) to quantify both semantic similarity and lexical fidelity with respect to expert annotations. Fine-tuning yielded substantial improvements across all evaluation metrics. The adapted PaliGemma-3B achieved a 40% increase in SBERT similarity (Fig. 2) relative to its base counterpart, reflecting enhanced semantic understanding of botanical traits. Likewise, gains in ROUGE-L and METEOR demonstrate the model’s capacity to employ precise botanical terminology rather than generic descriptive language. This highlights the model’s potential to ground linguistic outputs in visual evidence, enabling more accurate interpretation of herbarium specimens.
It is estimated that there are nearly 400 million herbarium specimens held across approximately 3,500 herbaria worldwide (Davis 2023). Over the past decade, many institutions have embarked on large-scale digitisation initiatives to robustly and efficiently digitise herbarium sheet collections. While these processes can be highly productive, reaching daily rates of 600–1,000 herbarium sheets in some programmes, there remains a risk and cost associated with re-digitisation, often caused by complications during the imaging stage. To address these challenges, the Natural History Museum, London (NHM) partnered with the Royal Botanic Garden Edinburgh (RBGE) to develop an Artificial Intelligence (AI)-based workflow aimed at providing enhanced quality control of herbarium sheet digitisation. This initiative was supported by the UK Government’s Department for Culture, Media and Sport (DCMS) through the NHM–DCMS AI Pilot Programme (Poon et al. 2026). The aim of this project was to develop an AI-assisted quality control framework (Zhang et al. 2025). Trained on collections at RBGE, the tool enables automated quality checks of digitised herbarium sheets. This is done by detecting missing objects (such as colour reference cards or barcodes), identifying cropping errors, and highlighting focus issues (Fig. 1). The pipeline integrates several computational techniques, including object detection, decision trees, classical statistics, and deep learning methods. In addition to post-digitisation analysis, a user-friendly interface was also developed to support in situ quality checks during the imaging stages. Model evaluation on an independent test set demonstrated the high accuracy and robustness of the AI-assisted framework (Zhang et al. 2025). The object detection model achieved an overall precision of 0.98, recall of 0.98, and mean Average Precision scores of 0.99 (mAP50) and 0.94 (mAP50–95) across all object categories. Key reference elements, including colour reference cards, barcodes, and rulers, were detected with near-perfect accuracy, ensuring precise object localisation. We note that as model development was done on RBGE collections alone, this meant that reference elements were similar across many herbarium sheets in the sample. Lastly, the issue detection models also performed strongly, achieving a precision of approximately 1.00 for out-of-focus detection, 0.98 for over-cropping, and 0.93 for under-cropping. The proposed workflow underwent further evaluation with herbarium digitisers at RBGE, which will inform refinements to the model, software interface, and the overall pipeline. The next step is to expand on this pilot project by testing the model with collections across different institutions, leading to more refinements for generalisability. This can also help us evaluate and estimate the total proportions of digitised herbarium sheets with issues. Finally, our global aim is to provide an AI-based quality control service that can be integrated into digitisation workflows not only at RBGE but across institutions worldwide.
