MOSAIC checkpoints
Trained weights for MOSAIC: Data-Efficient Cross-Sensor Alignment for Earth Observation Representation Learning (DSAA 2026).
Code: github.com/ymittal23/MOSAIC
These files are the paper best.pt checkpoints (full training state, ~2.1 GB each). They are not stored in the GitHub repository.
| File | Paper variant | Pretraining data |
|---|---|---|
mosaic-ms12-best.pt | MOSAIC-MS12 | Sen12MS-CR all seasons, 98K pairs, 12-band optical |
mosaic-rgb-98k-best.pt | MOSAIC-RGB-98K | Sen12MS-CR all seasons, 98K pairs, RGB |
mosaic-rgb-spring-best.pt | MOSAIC-RGB | Sen12MS-CR spring, 23K pairs, RGB |
Download
pip install huggingface_hub
huggingface-cli download ymittal23/MOSAIC mosaic-ms12-best.pt --local-dir checkpoints
Evaluate
From the code repo:
python src/evaluate_all.py \
--checkpoint checkpoints/mosaic-ms12-best.pt \
--eurosat-dir data/eurosat \
--eurosat-sar-dir data/eurosat_sar/EuroSAT-SAR \
--output-dir experiments/eval/eurosat \
--optical-channels 12 \
--mode all \
--seed 42
Use --optical-channels 3 with the RGB checkpoints.
Citation
@inproceedings{mittal2026mosaic,
title={MOSAIC: Data-Efficient Cross-Sensor Alignment for Earth Observation Representation Learning},
author={Mittal, Yash and Aggarwal, Deeksha and Kumar, Uttam},
booktitle={IEEE International Conference on Data Science and Advanced Analytics (DSAA)},
year={2026}
}
License
MIT License. Third-party datasets retain their own licenses.