---
license: mit
task_categories:
- zero-shot-classification
language:
- en
tags:
- plants
- GBIF
- iNaturalist
- Ecology
---
# **gbif-plants-raw**

A large-scale dataset of **96.1 million research-grade plant observations** sourced from **iNaturalist Open Data** and aligned with **GBIF taxonomy**.
Each row contains species metadata, taxonomic identifiers, geolocation, event timing, dataset source info, and a direct **image URL**.

This dataset is designed for large-scale image classification, biodiversity modelling, and pretraining work.

---

## **Dataset Summary**

This dataset aggregates all research-grade *Plantae* observations from iNaturalist and exports them into a flat, machine-friendly Parquet format.

Each record includes:

* **Taxonomy**: species, genus, family names + GBIF IDs
* **Geolocation**: latitude, longitude
* **Event metadata**: event date, dataset source
* **Image metadata**: direct image URL (iNat CDN), license info
* **Local identifiers**: GBIF occurrence IDs, dataset keys

No images are stored in this dataset. Only URLs and metadata are provided.

---

## **Use Cases**

* Large-scale plant species classification
* Vision transformer pretraining (ViT, ConvNeXt, etc.)
* Weakly-supervised learning using URLs
* Region-aware or habitat-aware plant ID models
* Training LoRA adapters for specific plant subsets
* Ecological modelling using species + geolocation
* Dataset bootstrapping for downstream fine-tuning tasks

---

## **Dataset Structure**

Total rows: **96,100,000+**
Format: **Parquet**
Split: `train` (single split)

---

## **Features**

| Column            | Type   | Description                                       |
| ----------------- | ------ | ------------------------------------------------- |
| `gbif_id`         | string | GBIF occurrence ID                                |
| `species_id`      | string | GBIF species ID                                   |
| `genus_id`        | string | GBIF genus ID                                     |
| `family_id`       | string | GBIF family ID                                    |
| `species_name`    | string | Scientific species name                           |
| `genus_name`      | string | Scientific genus name                             |
| `family_name`     | string | Scientific family name                            |
| `lat`             | string | Latitude                                          |
| `lon`             | string | Longitude                                         |
| `event_date`      | string | Observation event timestamp                       |
| `dataset_key`     | string | GBIF dataset key                                  |
| `dataset_name`    | string | Dataset name (usually iNaturalist research-grade) |
| `basis_of_record` | string | Observation type                                  |
| `image_url`       | string | Direct image URL (iNaturalist Open Data)          |
| `license_raw`     | string | License URL for the media                         |
| `rights_holder`   | string | Name of the copyright holder                      |

---

## **Licensing**

All media follows the original **iNaturalist Open Data** licensing provided by contributors.
Each row includes `license_raw` and `rights_holder`.

You must comply with the specific Creative Commons license associated with each image URL.
This dataset itself (metadata only) is released under **CC0**, but **images are NOT included and are NOT CC0**.

---

## **How to Load**

### **Python (HF Datasets)**

```python
from datasets import load_dataset

ds = load_dataset("juppy44/gbif-plants-raw", split="train", streaming=True)

for row in ds.take(5):
    print(row["species_name"], row["image_url"])
```

### **Polars**

```python
import polars as pl

df = pl.read_parquet("shard_0000.parquet")
print(df.head())
```

---

## **Notes**

* Some observations may contain outdated or deprecated GBIF taxonomy entries.
* Image URLs point to iNaturalist CDN mirrors and may expire in rare cases.
* No filtering by image quality, angle, or duplicates has been applied.
* Future versions may include cleaned subsets or image-validated variants.

---

## **Citation**

If you use this dataset, please cite:

```
iNaturalist Open Data. https://www.inaturalist.org/pages/open-data
GBIF: The Global Biodiversity Information Facility. https://www.gbif.org/
juppy44. gbif-plants-raw (2025). https://huggingface.co/datasets/juppy44/gbif-plants-raw
```

---

## **Contact / Contributions**

Open to collaboration on:

* cleaned or taxon-specific subsets
* embedding-based deduplication
* full image dataset export
* pretraining versions for ViT / ConvNeXt
* LoRA adapter training pipelines

Feel free to open issues or discussions on the Hugging Face repo.