EO-VAE: Towards A Multi-sensor Tokenizer for Earth Observation

EO-VAE is a multi-sensor variational autoencoder designed to serve as a foundational tokenizer for the Earth Observation (EO) domain. Unlike traditional approaches that require separate models for different sensors, EO-VAE utilizes a single model to encode and reconstruct flexible channel combinations through dynamic hypernetworks.

Model Summary

• Model Name: EO-VAE
• Paper: EO-VAE: Towards A Multi-sensor Tokenizer for Earth Observation Data
• License: Apache-2.0
• Task: Image-to-Image / Tokenization (Remote Sensing)

Usage

import torch
from eo_vae.models.new_autoencoder import EOFluxVAE

model = EOFluxVAE.from_pretrained(
    repo_id="nilsleh/eo-vae", 
    ckpt_filename="eo-vae.ckpt",
    config_filename="model_config.yaml",
    device="cpu",
)

# Run reconstruction / latent extraction
x = torch.randn(1, 3, 256, 256)
# Example wavelengths for Sentinel-2 RGB
wvs = torch.tensor([0.665, 0.56, 0.49], dtype=torch.float32) 

with torch.no_grad():
    recon = model.reconstruct(x, wvs)                # [B, 3, 256, 256]
    z = model.encode_spatial_normalized(x, wvs)      # [B, 32, 32, 32] for 256x256 input

These are the wavelengths used across modalities:

WAVELENGTHS = {
    'S2RGB': [0.665, 0.56, 0.49],
    'S1RTC': [5.4, 5.6],
    'S2L2A': [
        0.443, 0.490, 0.560, 0.665, 0.705, 0.740,
        0.783, 0.842, 0.865, 1.610, 2.190, 0.945,
    ],
    'S2L1C': [
        0.443, 0.490, 0.560, 0.665, 0.705, 0.740,
        0.783, 0.842, 0.865, 0.945, 1.375, 1.610, 2.190,
    ],
}

If you use this model in your work, please cite:

EO-VAE: Towards A Multi-sensor Tokenizer for Earth Observation Data

@article{eo-vae,
  title={EO-VAE: Towards A Multi-sensor Tokenizer for Earth Observation Data},
  author={Lehmann, Nils and Wang, Yi and Xiong, Zhitong and Zhu, Xiaoxiang},
  journal={arXiv preprint arXiv:2602.12177},
  year={2026}
}