Hugging Face's logo

Datasets:
paulhoehn
/
Sen12Landslides

DOI:
License:
Dataset card Files
xet
Community
1
The Dataset Viewer has been disabled on this dataset.

Sen12Landslides: A Spatio-Temporal Dataset for Satellite-Based Landslide and Anomaly Detection

This repository hosts Sen12Landslides, a large-scale, multi-sensor benchmark for spatio-temporal landslide and anomaly detection. It comprises 39,556 NetCDF patches (128×128 px with 10m resolution) spanning 15 time steps, derived from both Sentinel-1 (VV, VH; ascending/descending) and Sentinel-2 (10 spectral bands B02–B12). Every patch includes additionally:

  • A binary landslide mask
  • A high-resolution digital elevation model (DEM)
  • Rich metadata (event dates, trigger type, geolocation etc.)

We also provide two task-specific subsets which can be found in the GitHub Repo:

  • S12LS-LD for supervised landslide detection
  • S12LS-AD for anomaly detection

This release features over 75 000 annotated landslide polygons and ~12 000 multi-temporal image patches across SAR, optical, and elevation modalities, enabling rapid development and benchmarking of geohazard detection models:

Full Dataset

Modality Samples Annotated Ann. Rate
S1-asc 13,306 6,492 48.8%
S1-dsc 12,622 6,347 50.3%
S2 13,628 6,737 49.4%
Aligned 11,719 6,026 51.4%

Task Splits

Modality S12LS-LD S12LS-AD
S1-asc 4,793 (100%) 13,306 (48.8%)
S1-dsc 4,666 (100%) 12,622 (50.3%)
S2 4,988 (100%) 13,628 (49.4%)
Aligned 4,392 (100%) 11,719 (51.4%)

Paper link: https://www.nature.com/articles/s41597-025-06167-2

For any UPDATES, baselines, preprocessing scripts, data loaders or model-training pipelines please visit: GitHub repository

Note:

  • Sentinel-1 data in this dataset is provided in its original linear format and has not been converted to decibels (dB).
  • Sentinel-2 data has not been corrected for the radiometric offset of 1000 introduced by ESA for products processed with Baseline 04.00 or later (i.e., from January 25, 2022 onward).

If you wish to perform these corrections:

  • For Sentinel-1, convert values to dB using 10 * log10(x) (except bands DEM and MASK)
  • For Sentinel-2, add 1000 to each pixel value of the reflectance bands (B02-B12)

Functions to apply both conversions are available in the utils.py file of the GitHub repository.

Some of the data used in this dataset was downloaded using terragon, a tool for streamlined Earth observation data acquisition using multiple APIs like Microsoft Planetary Computer, GEE, ASF or CDSE.

Data Structure

The dataset is organized into three main folders based on the satellite data source:

Sen12Landslides/
├── s1asc/
│   ├── s1asc_part01.tar.gz
│   ├── ...
│   └── s1asc_part13.tar.gz
├── s1dsc/
│   ├── s1dsc_part01.tar.gz
│   ├── ...
│   └── s1dsc_part12.tar.gz
├── s2/
│   ├── s2_part01.tar.gz
│   ├── ...
│   └── s2_part28.tar.gz
└── inventories.shp.zip

inventories.shp.zip contains the official ground-truth landslide polygons used for annotation, aligned with the image patches in the dataset.

Each .tar.gz file in s1asc/, s1dsc/, or s2/ holds multiple .nc (NetCDF) files. Each file is a 128×128 pixel patch covering a specific location and includes 15 time steps of Sentinel-1 or Sentinel-2 data.

Filenames follow this pattern:

<region>_<sensor>_<id>.nc
e.g. italy_s2_6982.nc

Each file includes multi-temporal imagery, a binary MASK for landslide areas, and metadata like event date and pre/post indices. Sentinel-2 files may also contain DEM and cloud masks (SCL).

The output of such a file looks like the following after calling: xr.open_dataset("Sen12Landslides/data/s2/italy_s2_6982.nc")

<xarray.Dataset> Size: 6MB
Dimensions:      (time: 15, x: 128, y: 128)
Coordinates:
  * x            (x) float64 1kB 7.552e+05 7.552e+05 ... 7.565e+05 7.565e+05
  * y            (y) float64 1kB 4.882e+06 4.882e+06 ... 4.881e+06 4.881e+06
  * time         (time) datetime64[ns] 120B 2022-10-05 2022-10-30 ... 2023-09-10
Data variables: (12/14)
    B02          (time, x, y) int16 492kB ...
    B03          (time, x, y) int16 492kB ...
    B04          (time, x, y) int16 492kB ...
    B05          (time, x, y) int16 492kB ...
    B06          (time, x, y) int16 492kB ...
    B07          (time, x, y) int16 492kB ...
    ...           ...
    B11          (time, x, y) int16 492kB ...
    B12          (time, x, y) int16 492kB ...
    SCL          (time, x, y) int16 492kB ...
    MASK         (time, x, y) uint8 246kB ...
    DEM          (time, x, y) int16 492kB ...
    spatial_ref  int64 8B ...
Attributes:
    ann_id:           41125,41124,37694,37696,41131,37693,37689,37695,37749,3...
    ann_bbox:         (755867.5791119931, 4880640.0, 755900.7341873142, 48806...
    event_date:       2023-05-16
    date_confidence:  1.0
    pre_post_dates:   {'pre': 7, 'post': 8}
    annotated:        True
    satellite:        s2
    center_lat:       4881280.0
    center_lon:       755840.0
    crs:              EPSG:32632

For the corresponding Sentinel-1 data, the overall structure remains the same, but the data variables are adapted to SAR input, containing VV and VH bands instead of optical bands. The metadata attributes are consistent across modalities, with the only change being the satellite attribute set to "s1" instead of "s2".

Download 

pip install --upgrade huggingface_hub

hf auth login  # paste your token from https://huggingface.co/settings/tokens (only first time)

hf download paulhoehn/Sen12Landslides --repo-type dataset --local-dir /path/to/your/local_folder

Baselines

Due to class imbalance (~3% landslides), we provide, additionaly to our macro-avg metrics in the paper, binary metrics on the landslide class for benchmarking against other detection methods.

Note: To compare landslide detection performance, use the binary metrics below rather than the macro-averaged metrics from the paper.

Benchmark Results (S12LS-LD)

Benchmark using paper architectures with binary metrics on Sentinel-2 + DEM:

Model Precision Recall F1-Score IoU AP
ConvGRU 0.53 0.67 0.59 0.42 0.61
U-TAE 0.41 0.86 0.55 0.38 0.69
Unet3D 0.52 0.68 0.59 0.42 0.63
U-ConvLSTM 0.55 0.73 0.63 0.46 0.67

Three training runs (seed=42,123,777) were performed for each model on the S12LS-LD split with lit_module=binary for 100 epochs (early stopping enabled). Test metrics were averaged across seeds on the held-out test set. See configs/ for full settings.

Acknowledgement

Sen12Landslides was funded by HELMHOLTZ IMAGING, a platform of the Helmholtz Information & Data Science Incubator [grant number: ZT-1-PF-4-028].

This work was enabled by the computational and data resources provided through the "terrabyte" HPDA project of the German Aerospace Center (DLR) and Leibniz Supercomputing Center (LRZ), where Sentinel-1 NRB data was preprocessed and accessed.

Sentinel-2 Level-2A data were downloaded from Microsoft Planetary Computer and are provided under Copernicus Sentinel license conditions (© European Union 2015–2025, ESA) (https://planetarycomputer.microsoft.com/dataset/sentinel-2-l2a).

The DEM data is produced using Copernicus WorldDEM-30 © DLR e.V. 2010-2014 and © Airbus Defence and Space GmbH 2014-2018 provided under COPERNICUS by the European Union and ESA; all rights reserved.

Downloads last month
1,135