mespinosami/copgen-base

COP-GEN: Latent Diffusion Transformer for Copernicus Earth Observation Data

COP-GEN is a generative foundation model for Copernicus Earth observation data. It learns a joint distribution over all major Copernicus modalities — Sentinel-1 SAR, Sentinel-2 multispectral (L1C and L2A), DEM, and LULC — enabling both unconditional generation and cross-modal conditional synthesis (e.g. generate S2 RGB from S1 SAR, or generate all modalities jointly).

Model Details

• Developed by: Miguel Espinosa, Eva Gmelich Meijling, Valerio Marsocci, Elliot J. Crowley, Mikolaj Czerkawski
• Model type: Latent Diffusion Transformer (multimodal, multi-resolution)
• Modalities: S1RTC (VV, VH), S2L1C (all bands + cloud mask), S2L2A (all bands), DEM, LULC, timestamps, lat-lon
• License: CC-BY-4.0
• Paper: arXiv:2603.03239
• Repository: github.com/miquel-espinosa/COP-GEN

Architecture

COP-GEN operates in a shared latent space produced by a set of modality-specific KL-regularised VAEs. The diffusion backbone is a transformer trained jointly over all modalities, supporting arbitrary conditioning at inference time — any subset of modalities can be held as conditions while the rest are generated.

Uses

Direct Use

Generate synthetic Copernicus EO scenes, either unconditionally or conditioned on one or more input modalities. Useful for data augmentation, gap-filling missing modalities, and studying cross-sensor relationships.

Downstream Use

The latent representations and generated samples can serve as inputs to downstream EO tasks: land cover classification, change detection, cloud removal, SAR-to-optical translation, and more.

How to Get Started

from libs.copgen import CopgenModel

model = CopgenModel(
    model_path="path/to/model_checkpoint.pth",
    config_path="path/to/model_config.py"
)

# Conditional generation: provide one or more modalities as conditions
samples = model.generate(
    modalities=["S2L2A_B02_B03_B04_B08", "S1RTC_vh_vv"],
    conditions={"S1RTC_vh_vv": s1_tensor},
    n_samples=4,
)

# Unconditional generation
samples = model.generate(
    modalities=["S2L2A_B02_B03_B04_B08", "S1RTC_vh_vv"],
    n_samples=4,
)

See examples/conditional_generation.py and examples/unconditional_generation.py for full worked examples.

Training Details

Training Data

Trained on Major-TOM global Copernicus data, covering Sentinel-1 RTC, Sentinel-2 L1C and L2A, DEM, and LULC. A pre-compiled Edinburgh subset is available at mespinosami/copgen-edinburgh-subset for local development and reproduction.

Training Procedure

1. Modality-specific KL-VAEs are trained separately per modality and resolution.
2. All modalities are encoded into a shared latent space.
3. A diffusion transformer backbone is trained jointly over the merged latents, with random masking of modalities to enable conditional generation at inference.

See the GitHub README for full training instructions.

Evaluation

Evaluated on the COP-GEN-Benchmark test set (495 held-out global scenes). Distribution-level metrics (FID and related) are reported in Table 1 of the paper. To reproduce:

pip install -r benchmark/stochastic/requirements.txt
python -m benchmark.stochastic.run --output metrics.csv

Citation

@article{copgen2026,
    title   = {COP-GEN: Latent Diffusion Transformer for Copernicus Earth
               Observation Data},
    author  = {Espinosa, Miguel and Gmelich Meijling, Eva and Marsocci,
               Valerio and Crowley, Elliot J. and Czerkawski, Mikolaj},
    year    = {2026},
    journal = {arXiv preprint arXiv:2603.03239},
    url     = {https://arxiv.org/abs/2603.03239},
}