Hugging Face's logo

Datasets:
tacofoundation
/
cloudsen12

Modalities:
Geospatial
Languages:
English
Tags:
clouds
sentinel-2
image-segmentation
deep-learning
remote-sensing
License:
Dataset card Files Community
cloudsen12 / README.md
csaybar's picture
csaybar
Update README.md
f287a30 verified
preview code
|
raw
history blame
14.6 kB
metadata
license:
  - cc0-1.0
language:
  - en
tags:
  - clouds
  - sentinel-2
  - image-segmentation
  - deep-learning
  - remote-sensing
pretty_name: cloudsen12plus
viewer: false

Dataset Image

This dataset follows the TACO specification.

cloudsen12plus

Website: https://cloudsen12.github.io/

The largest dataset of expert-labeled pixels for cloud and cloud shadow detection in Sentinel-2

CloudSEN12+ version 1.1.0 is a significant extension of the CloudSEN12 dataset, which doubles the number of expert-reviewed labels, making it, by a large margin, the largest cloud detection dataset to date for Sentinel-2. All labels from the previous version have been curated and refined, enhancing the dataset's truestworthiness. This new release is licensed under CC0, which puts it in the public domain and allows anyone to use, modify, and distribute it without permission or attribution.

The images are padded from 509x509 to 512x512 and 2000x2000 to 2048x2048 to ensure that the patches are divisible by 32. The padding is filled with zeros in the left and bottom sides of the image. For those who prefer traditional storage formats, GeoTIFF files are available in our ScienceDataBank repository.

CloudSEN12+ v.1.1.0 offers three distinct modes, tailored for diverse research and application needs:

  • cloudsen12-l1c: Patches derived from Sentinel-2 Level-1C imagery, including high-quality labels, scribble annotations, and unlabeled data.

  • cloudsen12-l2a: Similar to cloudsen12-l1c but based on Sentinel-2 Level-2A data as processed by Google Earth Engine.

  • cloudsen12-extra: A supplementary collection of metadata to enhance contextual understanding of landscapes. Cloud masks from multiple sources have been normalized to align with the CloudSEN12 class schema. This mode includes:

    • s1_vv: Normalized Sentinel-1 Global Backscatter Model Land Surface (VV polarization).
    • s1_vh: Normalized Sentinel-1 Global Backscatter Model Land Surface (VH polarization).
    • elevation: Elevation data (meters) sourced from the MERIT Hydro dataset.
    • LC10: ESA WorldCover 10m v100 land cover product.
    • cloudmask_qa: Cloud mask from Sentinel-2 Level-1C.
    • cloudmask_sen2cor: Cloud mask from Sentinel-2 Level-2A.
    • cloudmask_cloudscore: Cloud detection results from Google Cloud Masking.
    • cloudmask_s2cloudless: Cloud mask generated by Sentinel Hub Cloud Detector.
    • cloudmask_unetmobv2: A cloud mask produced by a UnetMobV2 model trained on the CloudSEN12 dataset.

  • Changelog: Version 1.1.0:

    • We save all GeoTIFF files with discard_lsb=2 to improve the compression ratio.
    • Fixed 2000x2000 rotated patches. The datapoints are now correctly oriented. Check the patches:
      • ROI_2526__20200709T105031_20200709T105719_T31UDQ
      • ROI_0070__20190708T130251_20190708T130252_T24MUA
      • ROI_4565__20200530T100029_20200530T100502_T32TQP
    • Improved the quality of the following patches:
      • ROI_1098__20200515T190909_20200515T191310_T11WPN
      • ROI_1735__20190814T163849_20190814T164716_T15SXS
      • ROI_0760__20190516T022551_20190516T022553_T56WMD
      • ROI_3696__20200419T075611_20200419T080344_T35MRN
      • ROI_2864__20170529T105621_20170529T110523_T31TCN
    • We removed the following patches due to poor quality:
      • ROI_3980__20190228T005641_20190228T005640_T58WDB
      • ROI_1489__20210228T070831_20210228T070834_T40TDP

  • Consideration:

    • The field roi_id field serves as a unique identifier for the geographical location of each patch. In other words, it is used to link S2 images with a specific geographic location. However, the roi_id between the 509x509 and 2000x2000 patches are not the same. For example, the roid_id: ROI_0008 in the 509x509 patches is not the same as the ROI_0008 in the 2000x2000 patches. In this version, we fixed this issue by summing the max value of the 509x509 patches to the 2000x2000 patches. In this way, the roi_id between the 509x509 and 2000x2000 patches are unique. If users of 2000x2000 patches need to match the original roi_id published in the previous version, they can use the following formula:
      • old_roi_id_2000 = old_roi_id_2000 - 12101 where 12101 is the max value of the 509 patches. We also reported the previous roi as old_roi_id.
drawing
*CloudSEN12+ spatial coverage. The terms p509 and p2000 denote the patch size 509 × 509 and 2000 × 2000, respectively. `high`, `scribble`, and `nolabel` refer to the types of expert-labeled annotations*

🔄 Reproducible Example

Open In Colab

Load this dataset using the tacoreader library.

import tacoreader
import rasterio as rio

print(tacoreader.__version__) # 0.5.2

# Remotely load the Cloud-Optimized Dataset 
dataset = tacoreader.load("tacofoundation:cloudsen12-l1c")
#dataset = tacoreader.load("tacofoundation:cloudsen12-l2a")

# Read a sample
sample_idx = 2422
s2_l1c = dataset.read(sample_idx).read(0)
s2_label = dataset.read(sample_idx).read(1)

# Retrieve the data
with rio.open(s2_l1c) as src, rio.open(s2_label) as dst:    
    s2_l1c_data = src.read([4, 3, 2], window=rio.windows.Window(0, 0, 512, 512))
    s2_label_data = dst.read(window=rio.windows.Window(0, 0, 512, 512))

# Display
fig, ax = plt.subplots(1, 2, figsize=(10, 5))
ax[0].imshow(s2_l1c_data.transpose(1, 2, 0) / 3000)
ax[0].set_title("Sentinel-2 L1C")
ax[1].imshow(s2_label_data[0])
ax[1].set_title("Human Label")
plt.tight_layout()
plt.savefig("taco_check.png")
plt.close(fig)
drawing

🛰️ Sensor Information

The sensor related to the dataset: sentinel2msi

🎯 Task

The task associated with this dataset: semantic-segmentation

📂 Original Data Repository

Source location of the raw data:https://huggingface.co/datasets/isp-uv-es/CloudSEN12Plus

💬 Discussion

Insights or clarifications about the dataset: https://huggingface.co/datasets/tacofoundation/cloudsen12/discussions

🔀 Split Strategy

How the dataset is divided for training, validation, and testing: stratified

📚 Scientific Publications

Publications that reference or describe the dataset.

Publication 01 

@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 others},
  journal={Scientific Data},
  volume={9},
  number={1},
  pages={782},
  year={2022},
  publisher={Nature Publishing Group UK London}
}

Publication 02 

@inproceedings{aybar2023lessons,
  title={Lessons Learned From Cloudsen12 Dataset: Identifying Incorrect Annotations in Cloud Semantic Segmentation Datasets},
  author={Aybar, Cesar and Montero, David and Mateo-Garc{'\i}a, Gonzalo and G{'o}mez-Chova, Luis},
  booktitle={IGARSS 2023-2023 IEEE International Geoscience and Remote Sensing Symposium},
  pages={892--895},
  year={2023},
  organization={IEEE}
}

Publication 03

  • DOI: 10.1016/j.dib.2024.110852
  • Summary: Extended version of CloudSEN12. We include 2000 x 2000 patches to the dataset.
  • BibTeX Citation:
@article{aybar2024cloudsen12+,
  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 others},
  journal={Data in Brief},
  volume={56},
  pages={110852},
  year={2024},
  publisher={Elsevier}
}

🤝 Data Providers

Organizations or individuals responsible for the dataset.

Name Role URL
Image & Signal Processing host https://isp.uv.es/
European Space Agency (ESA) producer https://www.esa.int/

🧑‍🔬 Curators

Responsible for structuring the dataset in the TACO format.

Name Organization URL
Cesar Aybar Image & Signal Processing https://csaybar.github.io/

🏷️ Labels

The dataset contains four classes: clear, thick cloud, thin cloud, and cloud shadow.

Name Category Description
clear 0 Pixels without cloud and cloud shadow contamination. They are primarily identified using bands B4- B3-B2, B1-B12-B13, and the cirrus band.
thick cloud 1 Opaque clouds that block all the reflected light from the Earth's surface. We identify them by assuming clouds exhibit distinctive shapes and maintain higher reflectance values in bands B4-B3-B2, B1-B12-B13, and the cirrus band.
thin cloud 2 Semitransparent clouds that alter the surface spectral signal but still allow to recognize the background. This is the hardest class to identify. We utilize CloudApp [1] to better understand the background, both with and without cloud cover.
cloud shadow 3 Dark pixels where light is occluded by thick or thin clouds. Cloud shadows depend on clouds presence and, by considering the solar position, we can identify and map these shadows through a reasoned projection of the cloud shape.

🌈 Optical Bands

Spectral bands related to the sensor: L1C

Name Common Name Description Center Wavelength Full Width Half Max Index
B01 coastal aerosol Band 1 - Coastal aerosol - 60m 443.5 17.0 0
B02 blue Band 2 - Blue - 10m 496.5 53.0 1
B03 green Band 3 - Green - 10m 560.0 34.0 2
B04 red Band 4 - Red - 10m 664.5 29.0 3
B05 red edge 1 Band 5 - Vegetation red edge 1 - 20m 704.5 13.0 4
B06 red edge 2 Band 6 - Vegetation red edge 2 - 20m 740.5 13.0 5
B07 red edge 3 Band 7 - Vegetation red edge 3 - 20m 783.0 18.0 6
B08 NIR Band 8 - Near infrared - 10m 840.0 114.0 7
B8A red edge 4 Band 8A - Vegetation red edge 4 - 20m 864.5 19.0 8
B09 water vapor Band 9 - Water vapor - 60m 945.0 18.0 9
B10 cirrus Band 10 - Cirrus - 60m 1375.5 31.0 10
B11 SWIR 1 Band 11 - Shortwave infrared 1 - 20m 1613.5 89.0 11
B12 SWIR 2 Band 12 - Shortwave infrared 2 - 20m 2199.5 173.0 12

Spectral bands related to the sensor: L2A

Band Name Description Center Wavelength (nm) Bandwidth (nm) Index
B01 Coastal aerosol Band 1 - Coastal aerosol - 60m 443.5 17.0 0
B02 Blue Band 2 - Blue - 10m 496.5 53.0 1
B03 Green Band 3 - Green - 10m 560.0 34.0 2
B04 Red Band 4 - Red - 10m 664.5 29.0 3
B05 Red edge 1 Band 5 - Vegetation red edge 1 - 20m 704.5 13.0 4
B06 Red edge 2 Band 6 - Vegetation red edge 2 - 20m 740.5 13.0 5
B07 Red edge 3 Band 7 - Vegetation red edge 3 - 20m 783.0 18.0 6
B08 NIR Band 8 - Near infrared - 10m 840.0 114.0 7
B8A Red edge 4 Band 8A - Vegetation red edge 4 - 20m 864.5 19.0 8
B09 Water vapor Band 9 - Water vapor - 60m 945.0 18.0 9
B11 SWIR 1 Band 11 - Shortwave infrared 1 - 20m 1613.5 89.0 10
B12 SWIR 2 Band 12 - Shortwave infrared 2 - 20m 2199.5 173.0 11
AOT - Aerosol Optical Thickness - - 12
WVP - Water Vapor Pressure. The height the water would occupy if the vapor were condensed into liquid and spread evenly across the column - - 13