Prithvi-EO-2.0 Transformers Models

Hugging Face–compatible checkpoints converted from the official Prithvi-EO-2.0 foundation models (IBM, NASA, Jülich Supercomputing Centre). Each subfolder is a standalone model repo layout (config.json, model.safetensors, preprocessor, and remote code) for spatiotemporal feature extraction on HLS multispectral imagery.

Model Description

Prithvi-EO-2.0 is a 3D ViT encoder pretrained with masked autoencoding on NASA HLS V2 (30 m, six bands: B02–B07). TL variants add learned temporal (year + Julian day) and location (lat/lon) embeddings.

This collection bundles 6 converted checkpoints:

All folders ship self-contained remote code (modeling_prithvi.py, processor, pipeline) and load with trust_remote_code=True.

Developed by: IBM / NASA / JSC
Packaged for Hugging Face by: BiliSakura
License (weights): Apache-2.0
Original paper: Prithvi-EO-2.0: A Versatile Multi-Temporal Foundation Model for Earth Observation Applications

Legacy .pt filename mapping is in conversion_manifest.json.

Usage

Processors default to do_resize: false. Pass HLS reflectance values at native 224×224 (or another size divisible by the patch size); HLS mean/std normalization is applied by default.

Custom pipeline (recommended)

from transformers import pipeline
import numpy as np

REPO = "/home/czy/local/models/BiliSakura/Prithvi-EO-2.0-transformers"
SUBFOLDER = "prithvi-eo-v2-300m-tl"

pipe = pipeline(
    task="prithvi-eo-feature-extraction",
    model=REPO,
    trust_remote_code=True,
    model_kwargs={"subfolder": SUBFOLDER},
)

# Four HLS frames (T, H, W, C) in reflectance units — list of temporal frames
frames = [np.random.uniform(500, 3000, (224, 224, 6)).astype("float32") for _ in range(4)]
features = pipe(
    frames,
    pool=True,
    return_tensors=True,
    temporal_coords=[[2018, 26], [2018, 106], [2018, 201], [2018, 266]],
    location_coords=[19.5, -99.1],
)
print(features.shape)  # torch.Size([1, 1024])

# Dense token map (CLS + spatiotemporal patches)
tokens = pipe(frames, pool=False, return_tensors=True)
print(tokens.shape)    # torch.Size([1, 785, 1024]) for 224×224, patch 16, T=4

You can also pass a single array shaped (C, T, H, W) or (T, H, W, C).

Direct model loading

from transformers import AutoModel, AutoImageProcessor
import torch

model_dir = f"{REPO}/{SUBFOLDER}"
model = AutoModel.from_pretrained(model_dir, trust_remote_code=True)
processor = AutoImageProcessor.from_pretrained(model_dir, trust_remote_code=True)

batch = processor(frames, return_tensors="pt")
with torch.no_grad():
    outputs = model(**batch, temporal_coords=batch["temporal_coords"], location_coords=batch["location_coords"])
print(outputs.pooler_output.shape)

Standard image-feature-extraction

pipe = pipeline(
    task="image-feature-extraction",
    model=f"{REPO}/{SUBFOLDER}",
    trust_remote_code=True,
)

Normalization

The bundled image processor applies HLS pretraining normalization per band:

• Mean: [1087, 1342, 1433, 2734, 1958, 1363]
• Std: [2248, 2179, 2178, 1850, 1242, 1049]
• Nodata: -9999 → 0.0001 before normalization

Band order: B02, B03, B04, B05, B06, B07 (Blue, Green, Red, Narrow NIR, SWIR1, SWIR2).

Re-converting checkpoints

conda activate rsgen
python /home/czy/local/models/BiliSakura/Prithvi-EO-2.0-transformers/convert_all_checkpoints.py
python /home/czy/local/models/BiliSakura/Prithvi-EO-2.0-transformers/test_prithvi.py --all

Citation

@article{Prithvi-EO-V2-preprint,
    author  = {Szwarcman, Daniela and Roy, Sujit and others},
    title   = {{Prithvi-EO-2.0: A Versatile Multi-Temporal Foundation Model for Earth Observation Applications}},
    journal = {arXiv preprint arXiv:2412.02732},
    year    = {2024}
}