OmniCloudMask model weights
Weights for OmniCloudMask, a cloud and cloud shadow segmentation library for satellite imagery. The models take Red, Green and NIR bands at 10–50 m resolution, or 5–50 m for V3 and V4, and work across sensors, including Sentinel-2, Landsat 8, PlanetScope and Maxar. Running at 5 m widens the range of imagery the model accepts but is no more accurate than 10 m; see choosing a resolution.
You don't need to download these files by hand. The omnicloudmask Python package fetches the right ones on first use:
pip install omnicloudmask
from omnicloudmask import predict_from_array
mask = predict_from_array(rgn_array) # (3, H, W) array of Red, Green, NIR
See the documentation for usage, and the paper for method and validation.
Output classes
The class scheme follows CloudSEN12:
| Value | Class |
|---|---|
| 0 | Clear |
| 1 | Thick cloud |
| 2 | Thin cloud |
| 3 | Cloud shadow |
License
Copyright (c) 2026 The Government of Western Australia, managed by the Department of Primary Industries and Regional Development.
The licence depends on the model version. V2, V3 and V4 are MIT. V1 is CC BY-NC 4.0, licensed for non-commercial use only. The full licence terms, attribution and third-party notices are in LICENSES.md.
| File | Version | Licence |
|---|---|---|
PM_model_OCM_7.97_R_G_NIR_3_smp_regnety_004.pycls_in1k_PT_state.safetensors |
V4 | MIT |
PM_model_OCM_7.97_R_G_NIR_3_smp_edgenext_small.usi_in1k_PT_state.safetensors |
V4 | MIT |
PM_model_OCM_7.43_R_G_NIR_regnety_004.pycls_in1k_PT_state.safetensors |
V3 | MIT |
PM_model_OCM_7.43_R_G_NIR_edgenext_small.usi_in1k_PT_state.safetensors |
V3 | MIT |
PM_model_OCM_6.43_RG_NIR_regnety_004.pycls_in1k_PT_state.safetensors |
V2 | MIT |
PM_model_OCM_6.43_RG_NIR_edgenext_small.usi_in1k_PT_state.safetensors |
V2 | MIT |
PM_model_2.2.10_RG_NIR_509_regnety_004.pycls_in1k_PT_state.safetensors |
V1 | CC BY-NC 4.0 |
PM_model_2.2.10_RG_NIR_509_convnextv2_nano.fcmae_ft_in1k_PT_state.safetensors |
V1 | CC BY-NC 4.0 |
The same terms apply to copies of these weights hosted elsewhere or in other formats, such as the Google Drive mirror used by the omnicloudmask package, which hosts the V1 weights as .pth files.
The omnicloudmask package uses the latest version by default, so you only get V1 weights if you request them with model_version=1.
Why V1 is different
Two of V1's inputs are licensed for non-commercial use only:
- Training data. V1 was trained on the original CloudSEN12 release from Science Data Bank, which the Science Data Bank record and the CloudSEN12 download page list as CC BY-NC 4.0.
- Backbone. The V1 ConvNeXt model was fine-tuned from the timm ConvNeXt V2 ImageNet weights (
convnextv2_nano.fcmae_ft_in1k), which are licensed CC BY-NC 4.0.
Whether the licence of a training dataset or a pretrained model legally carries over to weights trained from it is not settled. Releasing V1 under CC BY-NC 4.0 is a cautious choice that respects the terms of both inputs.
V2 and later start from MIT-licensed backbones and were trained on the CloudSEN12+ re-release (CC0) and other data whose terms allow commercial use, listed under Training data and attribution. If you need a model for commercial use, use V2 or later.
The OmniCloudMask library code is MIT licensed. See the GitHub repository.
Model versions
| Version | Architectures | Training data added |
|---|---|---|
| V4 | regnety_004, edgenext_small (segmentation-models-pytorch + timm) |
CloudSEN12+ Scribble, CloudSEN12+ 2k, OCM Scribble |
| V3 | regnety_004, edgenext_small (fastai + timm) |
CloudSEN12+ High from Planetary Computer, CloudSEN12+ High super-resolved, KappaSet, OCM Hard Negatives |
| V2 | regnety_004, edgenext_small (fastai + timm) |
None (CloudSEN12+ High, re-sourced from the CC0 release) |
| V1 | regnety_004, convnextv2_nano (fastai + timm) |
Original CloudSEN12 High |
Each version is also trained on the datasets of the versions before it, except that V2 onwards use the CloudSEN12+ copy of CloudSEN12 High rather than the original. Each version is an ensemble of its two models. See the model changelog for details.
Pretrained backbones
All models were initialised from ImageNet-1k weights distributed through timm:
| Backbone | Used in | Source | Licence |
|---|---|---|---|
regnety_004.pycls_in1k |
V1–V4 | RegNetY | MIT |
edgenext_small.usi_in1k |
V2–V4 | EdgeNeXt | MIT |
convnextv2_nano.fcmae_ft_in1k |
V1 | ConvNeXt V2 | CC BY-NC 4.0 |
These backbones were pretrained on ImageNet, whose images are provided under non-commercial research terms. The backbones themselves are released under the licences listed above regardless, and the project's position is that the ImageNet image terms do not apply to these weights either.
Training data and attribution
| Dataset | Versions | Licence |
|---|---|---|
| CloudSEN12, original release (Science Data Bank) | V1 | CC BY-NC 4.0 |
CloudSEN12+ (tacofoundation/cloudsen12) |
V2–V4 | CC0 1.0 |
| KappaSet | V3–V4 | CC BY 4.0 |
| Sentinel-2 imagery from Microsoft Planetary Computer (CloudSEN12+ High re-source, OCM Hard Negatives and OCM Scribble) | V3–V4 | Copernicus Sentinel data terms |
| OCM Hard Negatives labels | V3–V4 | CC0 1.0. Labels created by the Department of Primary Industries and Regional Development, Government of Western Australia |
| OCM Scribble labels | V4 | CC0 1.0. Labels created by the Department of Primary Industries and Regional Development, Government of Western Australia |
The V4 training data, including the OCM datasets, is published as the OmniCloudMask Combined Training Dataset.
The super-resolved training tiles used in V3 and V4 were generated with 2xNomosUni_esrgan_multijpg (CC BY 4.0) by Philip Hofmann.
Contains modified Copernicus Sentinel data.
We thank the CloudSEN12 team (Aybar et al.) and the KappaSet authors (Shtym et al., KappaZeta) for making their labelled data available. The spatial distribution of training samples is shown on the training data map.
Citation
If you use these weights, please cite:
@article{wright2025omnicloudmask,
title = {Training sensor-agnostic deep learning models for remote sensing: Achieving state-of-the-art cloud and cloud shadow identification with OmniCloudMask},
author = {Wright, Nicholas and Duncan, John M. A. and Callow, J. Nik and Thompson, Sally E. and George, Richard J.},
journal = {Remote Sensing of Environment},
volume = {322},
pages = {114694},
year = {2025},
doi = {10.1016/j.rse.2025.114694}
}
Please also cite the training data and pretrained backbones for the version you use:
V3 and V4
@article{aybar2024cloudsen12plus,
title = {CloudSEN12+: The largest dataset of expert-labeled pixels for cloud and cloud shadow detection in Sentinel-2},
author = {Aybar, Cesar and Bautista, Lesly and Montero, David and Contreras, Julio and Ayala, Daryl and Prudencio, Fernando and Loja, Jhomira and Ysuhuaylas, Luis and Herrera, Fernando and Gonzales, Karen and Valladares, Jeanett and Flores, Lucy A. and Mamani, Evelin and Qui{\~n}onez, Maria and Fajardo, Rai and Espinoza, Wendy and Limas, Antonio and Yali, Roy and Alc{\'a}ntara, Alejandro and Leyva, Martin and Loayza-Muro, Ra{\'u}l and Willems, Bram and Mateo-Garc{\'\i}a, Gonzalo and G{\'o}mez-Chova, Luis},
journal = {Data in Brief},
volume = {56},
pages = {110852},
year = {2024},
doi = {10.1016/j.dib.2024.110852}
}
@misc{kappaset,
title = {KappaSet: Sentinel-2 KappaZeta Cloud and Cloud Shadow Masks},
author = {Shtym, Tetiana and Wold, Olga and Domnich, Marharyta and Voormansik, Kaupo and Rohtsalu, Mariana and Truup{\~o}ld, Joosep and Murin, Nikita and Toqeer, Abdullah and Odera, Catherine Akinyi and Harun, Fariha and S{\"u}nter, Indrek and Trofimov, Heido and Kostiukhin, Anton and J{\"a}rveoja, Mihkel},
howpublished = {Zenodo},
version = {1},
year = {2022},
doi = {10.5281/zenodo.7100327},
url = {https://doi.org/10.5281/zenodo.7100327}
}
@inproceedings{radosavovic2020regnet,
title = {Designing Network Design Spaces},
author = {Radosavovic, Ilija and Kosaraju, Raj Prateek and Girshick, Ross and He, Kaiming and Doll{\'a}r, Piotr},
booktitle = {2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
pages = {10425--10433},
year = {2020},
doi = {10.1109/CVPR42600.2020.01044}
}
@inproceedings{maaz2022edgenext,
title = {EdgeNeXt: Efficiently Amalgamated CNN-Transformer Architecture for Mobile Vision Applications},
author = {Maaz, Muhammad and Shaker, Abdelrahman and Cholakkal, Hisham and Khan, Salman and Zamir, Syed Waqas and Anwer, Rao Muhammad and Shahbaz Khan, Fahad},
booktitle = {Computer Vision -- ECCV 2022 Workshops: Tel Aviv, Israel, October 23--27, 2022, Proceedings, Part VII},
series = {Lecture Notes in Computer Science},
publisher = {Springer},
pages = {3--20},
year = {2022},
doi = {10.1007/978-3-031-25082-8_1}
}
V2
@article{aybar2024cloudsen12plus,
title = {CloudSEN12+: The largest dataset of expert-labeled pixels for cloud and cloud shadow detection in Sentinel-2},
author = {Aybar, Cesar and Bautista, Lesly and Montero, David and Contreras, Julio and Ayala, Daryl and Prudencio, Fernando and Loja, Jhomira and Ysuhuaylas, Luis and Herrera, Fernando and Gonzales, Karen and Valladares, Jeanett and Flores, Lucy A. and Mamani, Evelin and Qui{\~n}onez, Maria and Fajardo, Rai and Espinoza, Wendy and Limas, Antonio and Yali, Roy and Alc{\'a}ntara, Alejandro and Leyva, Martin and Loayza-Muro, Ra{\'u}l and Willems, Bram and Mateo-Garc{\'\i}a, Gonzalo and G{\'o}mez-Chova, Luis},
journal = {Data in Brief},
volume = {56},
pages = {110852},
year = {2024},
doi = {10.1016/j.dib.2024.110852}
}
@inproceedings{radosavovic2020regnet,
title = {Designing Network Design Spaces},
author = {Radosavovic, Ilija and Kosaraju, Raj Prateek and Girshick, Ross and He, Kaiming and Doll{\'a}r, Piotr},
booktitle = {2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
pages = {10425--10433},
year = {2020},
doi = {10.1109/CVPR42600.2020.01044}
}
@inproceedings{maaz2022edgenext,
title = {EdgeNeXt: Efficiently Amalgamated CNN-Transformer Architecture for Mobile Vision Applications},
author = {Maaz, Muhammad and Shaker, Abdelrahman and Cholakkal, Hisham and Khan, Salman and Zamir, Syed Waqas and Anwer, Rao Muhammad and Shahbaz Khan, Fahad},
booktitle = {Computer Vision -- ECCV 2022 Workshops: Tel Aviv, Israel, October 23--27, 2022, Proceedings, Part VII},
series = {Lecture Notes in Computer Science},
publisher = {Springer},
pages = {3--20},
year = {2022},
doi = {10.1007/978-3-031-25082-8_1}
}
V1
@misc{aybar2022cloudsen12data,
title = {CloudSEN12 - a global dataset for semantic understanding of cloud and cloud shadow in Sentinel-2},
author = {Aybar, Cesar and Ysuhuaylas, Luis and Loja, Jhomira and Gonzales, Karen and Herrera, Fernando and Bautista, Lesly and Yali, Roy and Flores, Angie and Diaz, Lissette and Cuenca, Nicole and Espinoza, Wendy and Prudencio, Fernando and Inga, Joselyn and Llactayo, Valeria and Montero, David and Sudmanns, Martin and Tiede, Dirk and Mateo-Garc{\'\i}a, Gonzalo and G{\'o}mez-Chova, Luis},
howpublished = {Science Data Bank},
version = {V1},
year = {2022},
doi = {10.57760/sciencedb.06669},
url = {https://cstr.cn/31253.11.sciencedb.06669},
note = {CSTR:31253.11.sciencedb.06669}
}
@article{aybar2022cloudsen12,
title = {CloudSEN12, a global dataset for semantic understanding of cloud and cloud shadow in Sentinel-2},
author = {Aybar, Cesar and Ysuhuaylas, Luis and Loja, Jhomira and Gonzales, Karen and Herrera, Fernando and Bautista, Lesly and Yali, Roy and Flores, Angie and Diaz, Lissette and Cuenca, Nicole and Espinoza, Wendy and Prudencio, Fernando and Llactayo, Valeria and Montero, David and Sudmanns, Martin and Tiede, Dirk and Mateo-Garc{\'\i}a, Gonzalo and G{\'o}mez-Chova, Luis},
journal = {Scientific Data},
volume = {9},
number = {1},
pages = {782},
year = {2022},
doi = {10.1038/s41597-022-01878-2}
}
@inproceedings{radosavovic2020regnet,
title = {Designing Network Design Spaces},
author = {Radosavovic, Ilija and Kosaraju, Raj Prateek and Girshick, Ross and He, Kaiming and Doll{\'a}r, Piotr},
booktitle = {2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
pages = {10425--10433},
year = {2020},
doi = {10.1109/CVPR42600.2020.01044}
}
@inproceedings{woo2023convnextv2,
title = {ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders},
author = {Woo, Sanghyun and Debnath, Shoubhik and Hu, Ronghang and Chen, Xinlei and Liu, Zhuang and Kweon, In So and Xie, Saining},
booktitle = {2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
pages = {16133--16142},
year = {2023},
doi = {10.1109/CVPR52729.2023.01548}
}
Model tree for NickWright/OmniCloudMask
Base model
timm/convnextv2_nano.fcmae_ft_in1k