Abstract
Ground beetles are a highly sensitive and speciose biological indicator, making them vital for monitoring biodiversity. However, they are currently an underutilized resource due to the manual effort required by taxonomic experts to perform challenging species differentiations based on subtle morphological differences, precluding widespread applications. In this paper, we evaluate 12 vision models on taxonomic classification across four diverse, long-tailed datasets spanning over 230 genera and 1769 species, with images ranging from controlled laboratory settings to challenging field-collected (in-situ) photographs. We further explore taxonomic classification in two important real-world contexts: sample efficiency and domain adaptation. Our results show that the Vision and Language Transformer combined with an MLP head is the best performing model, with 97% accuracy at genus and 94% at species level. Sample efficiency analysis shows that we can reduce train data requirements by up to 50% with minimal compromise in performance. The domain adaptation experiments reveal significant challenges when transferring models from lab to in-situ images, highlighting a critical domain gap. Overall, our study lays a foundation for large-scale automated taxonomic classification of beetles, and beyond that, advances sample-efficient learning and cross-domain adaptation for diverse long-tailed ecological datasets.
Introduction
Ground beetles (Carabidae) serve as vital bioindicators; their presence and diversity provide critical insights into ecosystem health. However, with over 40,000 species, many of which exhibit nearly identical morphological features, identification remains a significant challenge. Currently, this process relies on taxonomic experts who must spend hours examining sub-millimeter details like elytral striations or antennal segments under microscopes.
This "taxonomic bottleneck" hinders large-scale biodiversity monitoring. While computer vision offers a potential solution, it faces hurdles such as subtle inter-species differences, high intra-species variation across regions, and a significant "domain gap" between clean laboratory photographs and noisy, in-situ field observations.
Methodology and Datasets
The researchers conducted a systematic evaluation of 12 foundation models using "linear probing." The models included vision-only self-supervised (e.g., DINOv2), vision-only supervised (e.g., ConvNeXt), and vision-language models (e.g., CLIP, BioCLIP). Instead of full fine-tuning, feature embeddings were extracted and used to train a Multi-Layer Perceptron (MLP) head.
The study utilized a massive collection of 100,885 images across four distinct datasets: BeetlePUUM (lab images), BeetlePalooza (ethanol-preserved specimens), NHM-Carabids (museum collection), and I1MC (mixed lab and field photos).
Key Insights and Results
The evaluation revealed several critical findings for automated taxonomic classification:
Model Performance
The Vision and Language Transformer (ViLT) emerged as the top performer, achieving 97.15% accuracy at the genus level and 93.97% at the species level. By leveraging both visual and textual taxonomic concepts during pretraining, ViLT developed a superior understanding of fine-grained biological features. These results significantly outperform previous benchmarks, which typically ranged between 51% and 75% accuracy.
Sample Efficiency
A key practical finding is the "50% efficiency sweet spot." Models trained on only half of the available data performed nearly as well as those trained on the full dataset. This suggests that for long-tailed ecological data, strategic sampling is more impactful than sheer volume.
The Domain Gap and Metadata
A significant performance drop (approximately 37%) occurs when models trained on museum specimens are applied to field-collected (in-situ) images. Environmental factors like leaf litter, lighting, and moisture create a "domain gap" that remains difficult to bridge. Additionally, while metadata (morphological traits and environmental data) helped some models like BioCLIP, it occasionally hindered others at smaller scales, indicating that data fusion techniques require careful tuning.
Conclusion and Future Work
While BeetleVerse covers 1,769 species, this represents less than 5% of known ground beetle species. The "open-set" problem—identifying species the model has never seen—remains a hurdle for real-world deployment. Future research will focus on hierarchical consistency in predictions, improving domain adaptation from lab to field, and utilizing phylogenetic proximity to enhance model uncertainty and decision-making.
Summary updated