deepearth/california-flourishing-pollination

California Flourishing & Pollination

DeepEarth × UC Berkeley QED Lab — a self-supervised spatial-feature dataset of every iNaturalist Research-grade observation of every California-native plant and every California-observed flying pollinator, encoded with DINOv3 ViT-L/16 plus PhenoVision flowering/fruiting probabilities.

Maintained by Ecological Intelligence, Inc. (Lance Legel, PI) in collaboration with the Quantitative Ecosystem Dynamics Lab at UC Berkeley (Trevor Keenan, PI). The data-collection / DINOv3-inference / publication pipeline (with full scientific provenance) lives in legel/california_flourishing_pollination. The downstream DeepEarth models trained on this dataset live in legel/deepearth/models/flowering and legel/deepearth/models/pollination.

🌼 Interactive viewer: deepearth/california-flourishing-pollination Space

At a glance

Images embedded	10,273,298 (99.77 % of 10.30 M URL manifest)
Observations	5,244,656 iNaturalist Research-grade
Species	16,446 (6,383 California-native plants + 10,063 flying pollinators)
Geographic scope	California (US state) — Research-grade only, coordinates required
Per-image features	DINOv3 ViT-L/16 CLS (1024,) + 14×14 spatial patches (1024-dim, fp16) + PhenoVision flowering & fruiting probabilities (sigmoid)
Storage	~4.15 TB across 1,275 parquet shards (embeddings/embeddings_*.parquet)

Why this dataset exists

California is a global biodiversity hotspot whose native flora and pollinator fauna are under accelerating stress from climate change, land-use change, and invasive species. There is no single, openly-licensed, AI-ready dataset that binds plant flowering to its pollinator visitation at scale across the state. This dataset is a first step: it precomputes Meta AI's DINOv3 ViT-L/16 spatial features and PhenoVision flowering/fruiting probabilities for every Research-grade iNaturalist observation of every California-native plant and every flying pollinator with at least one California record, freezing those features as a reusable scientific asset.

What it contains

• species/plants_california_native.parquet — every CA-native plant taxon from the CNPS Calscape canonical export (8,507 taxa; 6,383 with photos in this dataset).
• species/pollinators_california_flying.parquet — animal taxa documented in GloBi as pollinating or visiting flowers of a CA native AND with ≥1 Research-grade CA iNat observation, gated by a curated flight-ability rule table (1,275 GloBi-confirmed core). The image dataset is broader — see below.
• interactions/globi_ca_plant_pollinator.parquet — every GloBi pollination interaction (RO_0002455 / RO_0002456 / RO_0002622 / RO_0002623) between a CA-native plant and an animal, geographically scoped to California (45,805 rows).
• manifests/image_manifest.parquet — image-grain manifest of every photo we embedded: gbif_occurrence_id, image_url_large, taxon_name (clean binomial, no authority), taxon_name_verbatim (GBIF scientificName with authority), gbif_taxon_key, family, license, rights_holder, creator, observed_on, decimal_latitude, decimal_longitude, locality, dataset_role, kingdom. We do not redistribute the photos; iNaturalist remains the source of record.
• embeddings/embeddings_*.parquet — 1,275 shards, ~10.27 M rows, ~4.15 TB:
  • DINOv3 CLS token (1024,) in cls_fp16 + cls_shape
  • DINOv3 spatial patches (14, 14, 1024) in patches_fp16 + patches_shape
  • PhenoVision phenovision_flowering_prob, phenovision_fruiting_prob ∈ [0,1] (sigmoid)
  • Full row metadata (license, rights_holder, taxon_name, lat/lng, observed_on, …)
• lookups/photo_attribution.parquet (236 MB) — per-photo CC license + rights_holder + creator from GBIF DwC-A multimedia.txt, keyed by (gbif_occurrence_id, image_url_large). Use for retroactive consumer-side join with any shard.
• lookups/taxon_clean_names.parquet (1.4 MB) — canonical clean taxon name (no authority) + rank per gbif_taxon_key.
• lookups/shard_index.parquet (111 MB) — image_url_large → shard_path for the Space viewer to fetch single-row embeddings without scanning all shards.
• lookups/umap_numpy.npz (206 MB) — pretrained UMAP(1024→3) extracted as numpy arrays (training data + embeddings + channel ranges) for cross-image-consistent RGB visualization. Loads under any Python version.
• lookups/umap_encoder.joblib (1.46 GB) — original UMAP encoder (joblib pickle; works under Python 3.10–3.11 only).
• lookups/global_pca.npz (18 KB) — top-3 global PCA components for a tiny alternative projection.
• provenance/*.jsonl — every API query, every file hash, every snapshot timestamp, every model checkpoint, copied verbatim from the pipeline run.
• PROVENANCE.md — human-readable provenance with all four GBIF DOIs and citation chain.

Pollinator scope (broader than the GloBi-confirmed core)

The image dataset includes Research-grade CA observations of any taxon in:

• Insecta (~8,500 species)
• Trochilidae (hummingbirds)
• Chiroptera (bats)
• Ptiliogonatidae, Mimidae, Icteridae, Parulidae, Cardinalidae, Bombycillidae (added 2026-05-23 to broaden flower-visiting bird coverage)

minus Formicidae (ants — flightless workers, per project scope).

Total: 10,063 pollinator species with photos.

How to use it

from datasets import load_dataset
import numpy as np

ds = load_dataset(
    "deepearth/california-flourishing-pollination",
    data_files="embeddings/embeddings_*.parquet",
    streaming=True,
)
row = next(iter(ds["train"]))
cls = np.frombuffer(row["cls_fp16"], dtype=np.float16).reshape(row["cls_shape"])
patches = np.frombuffer(row["patches_fp16"], dtype=np.float16).reshape(row["patches_shape"])
flowering_prob = row["phenovision_flowering_prob"]
fruiting_prob = row["phenovision_fruiting_prob"]

For cross-image-consistent RGB visualization (same flower-vs-leaf patch concept → same color across all observations), see the companion Space or use lookups/umap_numpy.npz:

import numpy as np
from sklearn.neighbors import NearestNeighbors
from huggingface_hub import hf_hub_download

z = np.load(hf_hub_download("deepearth/california-flourishing-pollination",
                             "lookups/umap_numpy.npz", repo_type="dataset"))
nn = NearestNeighbors(n_neighbors=int(z["n_neighbors"]), metric=str(z["metric"]))
nn.fit(z["training_data"].astype(np.float32))

def patches_to_rgb(patches_hwd: np.ndarray) -> np.ndarray:
    h, w, d = patches_hwd.shape
    flat = patches_hwd.reshape(-1, d).astype(np.float32)
    dists, idxs = nn.kneighbors(flat)
    w_ = 1.0 / (dists + 1e-6); w_ = w_ / w_.sum(axis=1, keepdims=True)
    proj = (w_[..., None] * z["training_embeddings"][idxs]).sum(axis=1)
    proj = np.clip((proj - z["channel_min"]) / (z["channel_max"] - z["channel_min"] + 1e-8), 0, 1)
    return (proj.reshape(h, w, 3) * 255).astype(np.uint8)

Known issues / Phase 1.5 backlog

• 64 of the 1,275 shards (the very earliest, pre-combined-extractor era) have no PhenoVision columns — they only have DINOv3 CLS + patches. Phase 1.5 task: re-fetch those images and run PhenoVision inference to add the columns.
• 22 iNat photo URLs in the manifest 404 at fetch time (deleted by their uploader after the GBIF snapshot). They appear in the manifest but won't appear in any embedding shard.
• ~24K bad/corrupt JPEGs that nvJPEG could not decode. They appear in the download checkpoint but not in any embedding shard. The original iNat photo is sometimes a truncated/malformed file; these are unrecoverable without re-fetching from iNat with a different decoder.
• PhenoVision label swap fix (resolved): shards uploaded before 2026-05-24T07:09 originally had the column names correct but the values of flowering and fruiting swapped (per vendor/phenovision/inference.py: class_names = ['fruiting', 'flowering']). All 1,200 affected shards were retroactively swap-fixed via scripts/fix_phenovision_swap_on_hf.py. 1,211 / 1,275 shards now correct (95 %; the 64 legacy shards have no PhenoVision at all).

Licensing

Component	License
This dataset (DINOv3 embeddings + manifests + species lists + interactions + lookups + code)	MIT
Source iNaturalist photos (we store URL + per-photo license string + creator only — never the photo bytes)	per-photo CC license recorded in every row — 83 % CC BY-NC 4.0, 11 % CC BY 4.0, 4.2 % CC0, 1.7 % other CC variants
DINOv3 model weights	per Meta's DINOv3 license (gated on HF)
PhenoVision model weights	MIT (Dinnage 2025)
GloBi interaction data	CC0 (per concept DOI 10.5281/zenodo.3950589)

DINOv3 spatial features are transformative derivatives of the source photos — the embeddings cannot be reversed to recover the image. Each row of the embedding shards carries the original image_url_large plus the per-photo license + rights_holder + creator strings (recovered directly from the GBIF DwC-A multimedia.txt), so downstream consumers re-fetch photos under each photo's own terms.

Citation

@dataset{legel_keenan_2026_cfp,
  title = {California Flourishing \& Pollination: a multi-modal AI dataset for ecological forecasting},
  author = {Legel, Lance and Keenan, Trevor},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/deepearth/california-flourishing-pollination},
}

Also cite the four GBIF Occurrence Downloads underpinning this dataset:

• 10.15468/dl.pbgs4h — plants × Calscape canonical (3.6 M records)
• 10.15468/dl.cvbfp4 — broad pollinators (1.5 M records)
• 10.15468/dl.gphfhs — expanded bird pollinators (0.3 M records)
• 10.15468/dl.nbe8dt — plant cultivar/variety recovery (4.6 M records)

Plus the upstream sources:

• PhenoVision: Dinnage, R., et al. (2025). PhenoVision: A framework for automating and delivering research-ready plant phenology data from field images. Methods in Ecology and Evolution 16(8):1763–1780. https://doi.org/10.1111/2041-210X.70081
• DINOv3: Siméoni, O., et al. (2025). DINOv3. arXiv:2508.10104.
• GloBi: Poelen, J. H., Simons, J. D., & Mungall, C. J. (2014). Global Biotic Interactions. Ecological Informatics 24:148-159.
• iNaturalist Research-grade observations via GBIF dataset key 50c9509d-22c7-4a22-a47d-8c48425ef4a7.
• CNPS Calscape: California Native Plant Society (2026). Calscape: Native Plants for California. https://calscape.org/.

Contact

Lance Legel — lance@ecological.dev · @deepearth on HF · github.com/legel/deepearth