metadata

license: mit
task_categories:
  - image-classification
  - feature-extraction
  - zero-shot-classification
language:
  - en
tags:
  - biodiversity
  - biology
  - computer-vision
  - multimodal
  - self-supervised-learning
  - florida
  - plants
pretty_name: DeepEarth Central Florida Native Plants
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: observations.parquet
      - split: test
        path: observations.parquet

DeepEarth Central Florida Native Plants Dataset v0.2.0

🌿 Dataset Summary

A comprehensive multimodal dataset featuring 33,665 observations of 232 native plant species from Central Florida. This dataset combines citizen science observations with state-of-the-art vision and language embeddings for advancing multimodal self-supervised ecological intelligence research.

Key Features

🌍 Spatiotemporal Coverage: Complete GPS coordinates and timestamps for all observations
🖼️ Multimodal: 31,136 observations with images, 7,113 with vision embeddings
🧬 Language Embeddings: DeepSeek-V3 embeddings for all 232 species
👁️ Vision Embeddings: V-JEPA-2 self-supervised features (6.5M dimensions)
📊 Rigorous Splits: Spatiotemporal train/test splits for robust evaluation

📦 Dataset Structure

observations.parquet         # Main dataset (500MB)
vision_index.parquet        # Vision embeddings index
vision_embeddings/          # Vision features (50GB total)
├── embeddings_000000.parquet
├── embeddings_000001.parquet
└── ... (159 files)

🚀 Quick Start

from datasets import load_dataset
import pandas as pd

# Load main dataset
dataset = load_dataset("deepearth/central-florida-plants")

# Access data
train_data = dataset['train']
print(f"Training samples: {len(train_data)}")
print(f"Features: {train_data.features}")

# Load vision embeddings (download required due to size)
vision_index = pd.read_parquet("vision_index.parquet")
vision_data = pd.read_parquet("vision_embeddings/embeddings_000000.parquet")

📊 Data Fields

Each observation contains:

Field	Type	Description
`gbif_id`	int64	Unique GBIF occurrence ID
`taxon_id`	string	GBIF taxon ID
`taxon_name`	string	Scientific species name
`latitude`	float	GPS latitude
`longitude`	float	GPS longitude
`year`	int	Observation year
`month`	int	Observation month
`day`	int	Observation day
`hour`	int	Observation hour (nullable)
`minute`	int	Observation minute (nullable)
`second`	int	Observation second (nullable)
`image_urls`	List[string]	URLs to observation images
`num_images`	int	Relative image number in GBIF occurrence
`has_vision`	bool	Vision embeddings available
`vision_file_indices`	List[int]	Indices to vision files
`language_embedding`	List[float]	7,168-dim DeepSeek-V3 embedding
`split`	string	train/spatial_test/temporal_test

🔄 Data Splits

The dataset uses rigorous spatiotemporal splits:

{ "train": 30935, "temporal_test": 2730 }

Temporal Test: All 2025 observations (future generalization)
Spatial Test: 5 non-overlapping geographic regions
Train: Remaining observations

🤖 Embeddings

Language Embeddings (DeepSeek-V3)

Dimensions: 7,168
Source: Scientific species descriptions
Coverage: All 232 species

Vision Embeddings (V-JEPA-2)

Dimensions: 6,488,064 values per embedding
Structure: 8 temporal frames × 24×24 spatial patches × 1408 features
Model: Vision Transformer Giant with self-supervised pretraining
Coverage: 7,113 images
Storage: Flattened arrays in parquet files (use provided utilities to reshape)

💡 Usage Examples

Working with V-JEPA 2 Embeddings

import numpy as np
import ast

# Load vision embedding
vision_df = pd.read_parquet("vision_embeddings/embeddings_000000.parquet")
row = vision_df.iloc[0]

# Reshape from flattened to 4D structure
embedding = row['embedding']
original_shape = ast.literal_eval(row['original_shape'])  # [4608, 1408]

# First to 2D: (4608 patches, 1408 features)
embedding_2d = embedding.reshape(original_shape)

# Then to 4D: (8 temporal, 24 height, 24 width, 1408 features)
embedding_4d = embedding_2d.reshape(8, 24, 24, 1408)

# Get specific temporal frame (0-7)
frame_0 = embedding_4d[0]  # Shape: (24, 24, 1408)

# Get mean embedding for image-level tasks
image_embedding = embedding_4d.mean(axis=(0, 1, 2))  # Shape: (1408,)

Species Distribution Modeling

# Filter observations for a specific species
species_data = dataset.filter(lambda x: x['taxon_name'] == 'Quercus virginiana')

# Use spatiotemporal data for distribution modeling
coords = [(d['latitude'], d['longitude']) for d in species_data]

Multimodal Learning

# Combine vision and language embeddings
for sample in dataset:
    if sample['has_vision']:
        lang_emb = sample['language_embedding']
        vision_idx = sample['vision_file_indices'][0]
        # Load corresponding vision embedding
        vision_emb = load_vision_embedding(vision_idx)

Zero-shot Species Classification

# Use language embeddings for zero-shot classification
species_embeddings = {
    species['taxon_name']: species['language_embedding']
    for species in dataset.unique('taxon_name')
}

📄 License

This dataset is released under the MIT License.

📚 Citation

If you use this dataset, please cite:

@dataset{deepearth_cf_plants_2024,
  title={DeepEarth Central Florida Native Plants: A Multimodal Biodiversity Dataset},
  author={DeepEarth Team},
  year={2024},
  version={0.2.0},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/deepearth/central-florida-plants}
}

🌟 Acknowledgments

We thank all citizen scientists who contributed observations through iNaturalist and GBIF. This dataset was created as part of the DeepEarth initiative for multimodal self-supervised ecological intelligence research.

🔗 Related Resources

📈 Dataset Statistics

Total Size: ~51 GB
Main Dataset: 500 MB
Vision Embeddings: 50 GB
Image URLs: 31,136 total images referenced
Temporal Range: 2019-2025
Geographic Scope: Central Florida, USA

Dataset prepared by the DeepEarth team for advancing multimodal self-supervised ecological intelligence research.