Earth Observation · isp-uv-es
CloudSEN12Plus
A benchmark dataset of over 49,000 Sentinel-2 patches with expert annotations for cloud and shadow segmentation.
What CloudSEN12Plus is
CloudSEN12Plus is an Earth observation benchmark developed by the Image and Signal Processing (ISP) group at the University of Valencia, curating multispectral imagery collected by the European Space Agency's Sentinel-2 mission. It serves as an expanded and refined edition of the original CloudSEN12 collection, doubling the volume of expert-reviewed annotations.
Accurate identification of atmospheric contamination is essential for optical remote sensing workflows. Clouds, semi-transparent cirrus, and cast shadows obscure surface reflectance, which can distort downstream biogeophysical analyses if not masked reliably. CloudSEN12Plus provides standardized semantic segmentation targets across diverse global biomes, seasons, and atmospheric conditions.
The dataset supports both Top-Of-Atmosphere (Level-1C) and Bottom-Of-Atmosphere (Level-2A) surface reflectance imagery. By providing stratified splits and paired multi-band arrays alongside detailed expert masks, it enables researchers to train and validate machine learning architectures for automated pixel-level cloud screening.
How it was made
The dataset gathers Sentinel-2 MultiSpectral Instrument (MSI) image patches at original resolutions of 509x509 and 2000x2000 pixels. To make array dimensions divisible by 32 for standard convolutional neural network backbones, the images are padded with zeros along their left and bottom margins to 512x512 and 2048x2048 pixels respectively.
Each patch contains thirteen spectral bands spanning coastal aerosol to shortwave infrared wavelengths, supplemented by cloud masks. Annotations are categorized into high-quality expert segmentations, scribble-level annotations detailing cloud centres and borders, and baseline automated predictions generated by UnetMobV2 models. The data is partitioned using a stratified sampling strategy into training, validation, and test splits.
Key facts
- Dataset size
- Over 49,000 patches
- Sensor
- Sentinel-2 MultiSpectral Instrument (MSI)
- Spectral bands
- 13 bands (B01 through B12, including B8A)
- Patch dimensions
- 512x512 and 2048x2048 pixels (zero-padded from 509x509 and 2000x2000)
- Processing levels
- Level-1C (L1C) and Level-2A (L2A)
- Split strategy
- Stratified train, validation, and test splits
- Annotation types
- High-quality expert labels, scribble labels, and model predictions
- Storage formats
- Parquet, MLS, and GeoTIFF
- Primary curator
- Cesar Aybar (Image & Signal Processing group)
What you can use it for
- Cloud and shadow semantic segmentation
- Train and benchmark deep learning segmentation models to separate clear ground, thick clouds, thin clouds, and cloud shadows in Sentinel-2 imagery.
- Weakly supervised segmentation research
- Develop segmentation models that learn from scribble annotations demarcating cloud centres and borders rather than full manual segmentations.
- Cross-processing level transfer
- Evaluate whether cloud masking networks trained on Level-1C top-of-atmosphere radiance transfer reliably to surface reflectance Level-2A data.
- Annotation hardness and trustworthiness estimation
- Use provided model predictions and expert labels to analyse classification ambiguity across varying terrain types and cloud optical depths.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from isp-uv-es/CloudSEN12Plus at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 47,362 rows of default/train
begin
int- min
- 0
- median
- 16.2B
- mean
- 17.3B
- max
- 40.8B
clear_percentage
int- min
- -1
- median
- 58
- mean
- 55.80
- max
- 100
cloud_shadow_percentage
int- min
- 0
- median
- 5
- mean
- 8.04
- max
- 87
datapoint_id
string_text- min
- 48
- median
- 48
- mean
- 48.16
- max
- 49
equi_id
string_text- min
- 17
- median
- 17
- mean
- 17
- max
- 17
equi_zone
string_label- AS29%
- AF23%
- NA16%
fixed
int- min
- 0
- median
- 0
- mean
- 7.6e-3
- max
- 1
label_type
string_label- nolabel62%
- high19%
- scribble19%
length
int- min
- 1.3M
- median
- 2.8M
- mean
- 3.3M
- max
- 54M
metadata_length
int- min
- 52,604
- median
- 1.1M
- mean
- 932k
- max
- 1.1M
proj_centroid
string_text- min
- 20
- median
- 25
- mean
- 25.64
- max
- 46
proj_epsg
int- min
- 32,603
- median
- 32,642
- mean
- 32658.05
- max
- 32,760
proj_geometry
string_text- min
- 77
- median
- 90
- mean
- 89.63
- max
- 90
proj_shape
int- min
- 509
- median
- 509
- mean
- 530.63
- max
- 2,000
proj_transform
string_text- min
- 27
- median
- 30
- mean
- 29.93
- max
- 30
roi_id
string_text- min
- 8
- median
- 8
- mean
- 8.16
- max
- 9
s2_date
string_text- min
- 20
- median
- 20
- mean
- 20
- max
- 20
s2_id
string_text- min
- 60
- median
- 60
- mean
- 60
- max
- 60
s2_id_gee
string_text- min
- 38
- median
- 38
- mean
- 38
- max
- 38
s2_mean_solar_azimuth_angle
float- min
- 27.68
- median
- 147.38
- mean
- 126.79
- max
- 308.22
s2_mean_solar_zenith_angle
float- min
- 5.73
- median
- 54.67
- mean
- 52.05
- max
- 81.87
split
string_label- train100%
thick_percentage
int- min
- 0
- median
- 17
- mean
- 27.50
- max
- 100
thin_percentage
int- min
- 0
- median
- 0
- mean
- 8.65
- max
- 100
url
string_label- https://huggingface.co/datasets/isp-uv-es/CloudSEN12Plus/resolve/main/train/train_509_nolabel_chunk1.mlstac31%
- https://huggingface.co/datasets/isp-uv-es/CloudSEN12Plus/resolve/main/train/train_509_nolabel_chunk2.mlstac31%
- https://huggingface.co/datasets/isp-uv-es/CloudSEN12Plus/resolve/main/train/train_509_scribble.mlstac19%
Strengths and limits
Good at
- Provides expert-quality manual masks covering both thick cloud and challenging thin semi-transparent cloud categories.
- Doubles the number of expert-reviewed annotations compared to the earlier CloudSEN12 release.
- Includes all thirteen Sentinel-2 MSI spectral bands, allowing algorithms to exploit cirrus and shortwave infrared channels.
- Offers cloud-optimized data formats that permit inspecting and querying samples without complete local archive downloads.
Watch out for
- Images contain zero-padding along their left and bottom borders to satisfy power-of-two spatial dimensions, which requires careful boundary handling during training.
- Scribble annotations only cover centres and outlines rather than complete dense pixel-level boundaries.
- The dataset is restricted to Sentinel-2 MSI imagery and does not include imagery from other satellite constellations.
Questions
Can I look inside CloudSEN12Plus without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is CloudSEN12Plus under?
Its authors declare cc0-1.0. Read the terms on the original page before reusing the data.
What classes are included in the high-quality label masks?
The dense high-quality masks provide four distinct integer classes: clear ground (0), thick cloud (1), thin cloud (2), and cloud shadow (3).
How are scribble labels categorized?
Scribble masks differentiate seven classes: clear ground (0), thick-cloud border (1), thick-cloud centre (2), thin-cloud border (3), thin-cloud centre (4), cloud-shadow border (5), and cloud-shadow centre (6).
Which image sizes are present in the dataset?
Patches originate as 509x509 and 2000x2000 pixel tiles, which have been zero-padded to 512x512 and 2048x2048 pixels respectively so their dimensions are multiples of 32.
Does the dataset include both raw top-of-atmosphere and surface reflectance data?
Yes, CloudSEN12Plus supports both Sentinel-2 Level-1C (top-of-atmosphere) and Level-2A (atmospherically corrected surface reflectance) products.
How to cite it
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, 'CloudSEN12+: The largest dataset of expert-labeled pixels for cloud and cloud shadow detection in Sentinel-2', Data in Brief, pages 110852, 2024, DOI: 10.1016/j.dib.2024.110852
Written from CloudSEN12Plus's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.