A continental-scale dataset of ground beetles with high-resolution images and validated morphological trait measurements

AI summary

Overview

• Research area: A dataset paper at the intersection of computer vision, imageomics, and ecology — high-resolution imaging and automated morphological trait measurement of ground beetles (Coleoptera: Carabidae).
• Technical level: Intermediate. The biological framing is accessible, but the paper assumes familiarity with object detection (Grounding DINO), annotation platforms (CVAT, TORAS), and morphometric trait extraction from images.
• Scope (1 sentence): The paper describes a publicly released multimodal dataset of over 13,200 ground beetle specimens from 30 National Ecological Observatory Network (NEON) sites across the continental United States and Hawaii, with high-resolution images and digitally measured, manually validated elytral traits.

What This Paper Is About

Global trait databases are heavily biased toward vertebrates and plants, leaving hyper-diverse invertebrate groups such as carabid beetles poorly represented, while NEON's extensive carabid collection exists mainly as physical specimens that are hard to access and analyze at scale. The authors digitize over 13,200 NEON carabids from 30 sites through standardized high-resolution imaging and measure elytral traits digitally, validating those measurements against manual caliper measurements. The goal is to remove the access and standardization barriers that currently block large-scale trait-based and AI-driven research on North American carabids.

Key Contributions

1. A continental-scale carabid image dataset. Over 13,200 NEON carabid specimens from 30 sites spanning the continental US and Hawaii were imaged at high resolution, split into two specimen types: pinned specimens photographed at the NEON facility in the Pu'u Maka'ala (PUUM) Natural Area Reserve in Hilo, Hawaii, and vial specimens obtained from the NEON Biorepository for the sample year 2018.
2. Digitally measured, manually validated trait data. Elytral length and width were measured digitally for each specimen, with the digital pipeline validated against manual caliper-based measurements and reported to achieve sub-millimeter precision, giving a foundation for automated trait extraction.
3. Two public FAIR-compliant repositories. Images, metadata, and trait annotations are released on the HuggingFace datasets hub as imageomics/Hawaii-beetles (pinned; CC-BY-4.0) and imageomics/2018-NEON-beetles (vial; CC BY-SA 4.0), with standardized directory layouts, documented naming conventions, and metadata linking specimens to NEON identifiers.
4. A hybrid human-plus-AI annotation and segmentation workflow. Trait annotation used the Toronto Annotation Suite (TORAS) for pinned specimens and the Notes from Nature platform for vial specimens, while individual beetle segmentation combined the open-vocabulary detector Grounding DINO with manual review and correction in CVAT.

Main Findings

• Dataset composition (pinned repository): The imageomics/Hawaii-beetles repository contains 162 original group images, 1,614 individually cropped pinned specimen images, an images_metadata.csv file with 1,614 specimen-level records, and a trait_annotations.csv file with 1,579 specimen-level trait measurements, all under CC-BY-4.0.
• Two distinct imaging protocols were needed. Vial specimens were shot with a NIKON D500 and a 60 mm f/2.8G ED macro lens at 1/13s, f8.0, ISO 100 under a shadowbox with no flash, while pinned specimens were shot with a Canon EOS 7D and a 24–105 mm macro lens at 1/100s, f8.0, ISO 100 using two diffused flashes.
• Trait definitions differ by specimen type. Pinned specimens were annotated for three traits (elytral length, basal pronotum width, maximum elytral width) using TORAS, whereas vial specimens were annotated for two traits (elytral length and elytral width) via Notes from Nature.
• Vial elytral width was unreliable and excluded from final analyses. Inter-annotator agreement was high for vial elytral length, but elytral width showed greater variability because the mesothoracic leg landmarks were sometimes partially occluded; both length and width values remain in the dataset.
• Scale of the citizen-science annotation effort. Vial trait annotation covered 577 group images, originally captured at 5568 × 3712 pixels, which exceeded the 1 MB per-image upload limit of Notes from Nature, so resized images were used for annotation.
• Measurement reliability was assessed with repeated annotation. A subset of vial specimens was annotated independently by up to three contributors to evaluate inter-annotator agreement, and pinned measurements used predefined annotation "recipes" and standardized landmark definitions to limit inter-observer variability.
• Specific numeric error rates are not reported in the content provided. The paper states that digital trait extraction achieves sub-millimeter precision with transparent error quantification, but the truncated content does not include the numerical validation figures (these appear in the validation and inter-annotator agreement sections, which are referenced but not shown in full).

Methodology in Plain English

The authors started from specimens NEON already collects. NEON runs 81 field sites (47 terrestrial, 34 aquatic) across 20 ecoclimatic domains, and at each terrestrial site ground beetles are caught in pitfall traps: 16 oz deli containers holding 150–250 ml of propylene glycol, placed at the south, east, and west positions of a 40 × 40 m plot, collected roughly every two weeks (about 14 days of deployment) for up to 13 bouts per season, typically May through October. Estimates of trapping effort were standardized across sites. Field staff transfer specimens into 95% ethanol within 24 hours, and NEON parataxonomists sort carabids from bycatch and identify them. Specimens with uncertain identifications are pinned for expert review; confirmed specimens are pinned if ten or fewer individuals of a species appear in a bout and otherwise stored in ethanol vials. Plot counts began at 10 per site, reduced to 6 starting in 2023, and the north trap position was discontinued in 2018 (four traps to three).

Imaging was done with two separate calibrated setups. Pinned beetles were photographed in groups of 3–5 on white backgrounds with all pins (including the scale bar pin) at a uniform height, which produced a single focal plane and avoided focus stacking. Identification labels were intentionally rotated so text appears upside down, keeping text away from the beetles for cleaner downstream segmentation. Vial beetles were dried on absorbent paper after ethanol removal, arranged on wooden stakes for a consistent dorsal view and focal plane, and imaged together with the source vial and its barcode positioned in the left of the frame, plus a scale bar in the upper left. Every image was checked for focus, lighting, feature visibility, label legibility, scale bar placement, and color fidelity, and failures were retaken on the spot; passing images were backed up in triplicate before any post-processing.

Trait measurement diverged by specimen type. Pinned specimens were measured in the Toronto Annotation Suite (TORAS) using AI-assisted polyline drawing, with standardized "recipes" specifying landmarks, allowed point counts, handling of difficult cases (for example, open elytra required a three-segment polyline across the left elytron, the midline gap, and the right elytron), and a required annotation order ending with the scale bar. A Python script converted polyline pixel lengths to millimeters using the scale bar. Vial specimens were annotated by contributors on Notes from Nature using a protocol aligned with the TORAS definitions.

To isolate individual beetles from group images, the authors ran Grounding DINO with the prompt "a beetle" to produce candidate bounding boxes, then imported all candidates into CVAT for human review, where missing beetles were added, boxes were expanded to include structures such as antennae, and false positives were removed. The final cropped specimen images were saved from the corrected boxes. Metadata and naming conventions were designed so that each cropped image links back to its parent group image and to the NEON individualID.

Why This Matters

• It addresses the invertebrate gap in trait databases. Trait databases are dominated by vertebrates and plants, so this dataset adds a large, standardized carabid resource for a group with over 40,000 known species that functions as a bioindicator of ecosystem health.
• It makes a physical collection computationally usable. NEON's carabids previously existed mainly as physical specimens without comprehensive images or trait measurements; digitization enables large-scale analysis without requiring researchers to visit collections.
• It supplies a methodologically transparent benchmark. Because digital measurements were validated against manual caliper measurements and reported to reach sub-millimeter precision, the dataset gives computer vision researchers a reference for automated trait extraction rather than an unvalidated output.

Real-world applications:

• Automated species identification and trait extraction. The images and validated measurements support training and evaluation of models for identifying carabids and measuring body size traits directly from images.
• Biodiversity monitoring and conservation planning. Carabid trait values relate to dispersal capacity, habitat preference, and sensitivity to disturbance, which supports indicator-species analysis and habitat assessment.
• Trait–environment and cross-taxon analysis. Because specimens carry NEON identifiers, the data can be joined with NEON environmental, climatic, and remote sensing data, and compared with NEON's plant, small mammal, and fish trait datasets.
• Long-term ecological forecasting. Integrating with NEON infrastructure supports repeated, standardized measurements that could feed into monitoring and forecasting pipelines.

Industry relevance: the dataset is directly relevant to organizations building computer vision systems for biodiversity and specimen digitization — natural history museums, biodiversity data platforms, environmental monitoring agencies, and agricultural and biosecurity groups interested in insect identification. It also serves as a testbed for the practical constraints of ecological AI: ecological datasets are typically small and heterogeneous, which limits conventional computer vision approaches that assume large standardized training sets, and the paper explicitly frames this as a motivating gap.

Future Directions

• Extending the trait set beyond elytral length and width. The paper frames the current traits as a foundation for automated trait extraction; adding traits such as body size proxies, dispersal-related measurements, and finer morphological characters is the natural next step, and the pinned pipeline already includes basal pronotum width and maximum elytral width.
• Recovering reliable vial elytral width measurements. Because mesothoracic leg landmarks were sometimes occluded, vial elytral width was excluded from the final analyses; better landmark definitions or imaging could bring this trait back into use. The values are already released for studies that want them.
• Quantifying and reducing intraspecific variation uncertainty. The authors state that the degree of intraspecific variation across geographic clines, life stages, and environmental conditions remains unknown for most carabid species, which the dataset could begin to address.
• Scaling coverage across taxa, sites, and years. The current work covers 30 imaged sites and, for vial specimens, the 2018 sample year, with pinned specimens drawn from the on-site PUUM facility collection; expanding across NEON's 81 sites, additional years, and comparing with the Bishop Museum collection in Honolulu are open opportunities.

Target Audience

• Computer vision and machine learning researchers working on fine-grained recognition, open-vocabulary detection, segmentation, and automated morphometry, who need standardized, validated ecological image data.
• Ecologists and macroecologists studying functional diversity, community assembly, trait–environment interactions, and cross-taxon comparisons, especially those already using NEON data products.
• Biodiversity informatics and natural history collection staff interested in FAIR-compliant digitization workflows, specimen imaging protocols, and metadata standards.
• Citizen science and annotation platform practitioners, since the paper documents both expert-driven (TORAS) and volunteer-driven (Notes from Nature) annotation pipelines and their reliability trade-offs.

Authors’ abstract

Despite the ecological significance of invertebrates, global trait databases remain heavily biased toward vertebrates and plants, limiting comprehensive ecological analyses of high-diversity groups like ground beetles. Ground beetles (Coleoptera: Carabidae) serve as critical bioindicators of ecosystem health, providing valuable insights into biodiversity shifts driven by environmental changes. While the National Ecological Observatory Network (NEON) maintains an extensive collection of carabid specimens from across the United States, these primarily exist as physical collections, restricting widespread research access and large-scale analysis. To address these gaps, we present a multimodal dataset digitizing over 13,200 NEON carabids from 30 sites spanning the continental US and Hawaii through high-resolution imaging, enabling broader access and computational analysis. The dataset includes digitally measured elytra length and width of each specimen, establishing a foundation for automated trait extraction using AI. Validated against manual measurements, our digital trait extraction achieves sub-millimeter precision, ensuring reliability for ecological and computational studies. By addressing invertebrate under-representation in trait databases, this work supports AI-driven tools for automated species identification and trait-based research, fostering advancements in biodiversity monitoring and conservation.