TreeOfLife-200M

What TreeOfLife-200M is

TreeOfLife-200M is a large-scale multimodal dataset designed for training computer vision models in organismal biology, ecology, and conservation. Curated by researchers from the Imageomics Institute and collaborating institutions, it aggregates 233 million organism images spanning 933,798 unique taxa across the tree of life. It serves as the primary training corpus for hierarchical contrastive learning foundation models such as BioCLIP 2 and BioCLIP 2.5 Huge.

The dataset integrates organism imagery across a wide range of observational contexts, including museum specimen photography, automated camera trap records, and citizen science observations. By combining records from four major biodiversity repositories—GBIF, EOL, BIOSCAN-5M, and FathomNet—it broadens taxonomic breadth across kingdoms such as Animalia, Plantae, Fungi, and Chromista, covering an estimated 70.38% of species listed under threatened categories by the International Union for Conservation of Nature (IUCN).

To address label inconsistencies and data noise typical of biological aggregators, records are resolved through automated taxonomic normalisation and image validation pipelines. The compilation includes tabular metadata, pre-computed model text embeddings, and provenance tracking, enabling reproducible training and evaluation for fine-grained biological image recognition.

How it was made

Images and occurrence records are gathered from four core data providers: the Global Biodiversity Information Facility (GBIF), Encyclopedia of Life (EOL), BIOSCAN-5M, and FathomNet. Bounding boxes are applied to isolate marine organisms in FathomNet images. Taxonomic labels are mapped to a consistent seven-rank Linnaean hierarchy (kingdom to species) using the TaxonoPy algorithm, which queries the GBIF Backbone Taxonomy, Catalogue of Life, and OpenTree hierarchies in sequential order.

Data curation enforces image quality, privacy, and deduplication across all sources. Images must have a minimum dimension of 224 pixels and are constrained to a maximum of 720 pixels on their longest side. Non-organismal museum specimen records (such as specimen tags and collection drawers) are removed using a nearest-centroid classifier. Empty camera trap bursts are identified and discarded using MegaDetector. In citizen science observations, inconsistent photos are filtered by analysing pairwise BioCLIP embedding distances, while MTCNN removes identifiable human faces. Exact duplicate images and potential benchmark leaks are removed using MD5 and perceptual PDQ hashing.

Key facts

What you can use it for

Foundation model pre-training

Supplies pre-processed image-text pairs with hierarchical taxonomic text descriptions to pre-train vision-language models like BioCLIP 2.

Zero-shot and few-shot species classification

Enables taxonomic classification of rare, unseen, or under-sampled organisms using pre-computed model text embeddings across 933,798 taxa.

Biodiversity monitoring across contexts

Provides diverse training distributions encompassing field camera traps, macro insect captures, marine ROV imagery, and herbarium specimens.

Taxonomic representation and hierarchy benchmarks

Facilitates evolutionary and morphological feature analysis across taxonomic ranks, including dedicated sub-evaluations such as Darwin's finches.

Explore the data

The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.

Strengths and limits

Questions

How to cite it

Gu, J., Stevens, S., Campolongo, E. G., Thompson, M. J., Zhang, N., Wu, J., Kopanev, A., Mai, Z., White, A. E., Balhoff, J., Dahdul, W. M., Rubenstein, D., Lapp, H., Berger-Wolf, T., Chao, W.-L., & Su, Y. (2025). BioCLIP 2: Emergent Properties from Scaling Hierarchical Contrastive Learning. Advances in Neural Information Processing Systems, 38, 102778–102811.