LSD4WSD : An Open Dataset for Wet Snow Detection with SAR Data and Physical Labelling

LSD4WSD V2.0

Learning SAR Dataset for Wet Snow Detection - Full Analysis Version. 

The aim of this dataset is to provide a basis for automatic learning to detect wet snow. It is based on Sentinel-1 SAR GRD satellite images acquired between August 2020 and August 2021 over the French Alps. The new version of this dataset is no longer simply restricted to a classification task, and provides a set of metadata for each sample.

Modification and improvements of the version 2.0.0 :

• Number of massif: add 7 new massif to cover the all Sentinel-1 images (cf `info.pdf`).
• Acquisition: add images of the descending pass in addition to those originally used in the ascending pass.
• Sample: reduction in the size of the samples considered to 15 by 15 to facilitate evaluation at the central pixel.
• Sample: increased density of extracted windows, with a distance of approximately 500 meters between the centers of the windows.
• Sample: removal of the pre-processing involving the use of logarithms.
• Sample: removal of the pre-processing involving the normalisation.
• Labels: new structure for the labels part: dictionary with keys: `topography`, `metadata` and `physics`.
• Labels: `physics`: addition of direct information from the CROCUS model for 3 simulations: Liquid Water Content, snow height and minimum snowpack temperature.
• Labels: `topography`: information on the slope, altitude and average orientation of the sample.
• Labels: `metadata` : information on the date of the sample, the mountain massif and the run (ascending or descending).
• Dataset: removal of the train/test split*

*We leave it up to the user to use the Group Kfold method to validate the models using the alpine massif information.

Finally, it consists of 2467516 samples of size 15 by 15 by 9. For each sample, the 9 metadata are provided, using in particular the Crocus physical model:

• topography:
  • elevation (meters) (average),
  • orientation (degrees) (average),
  • slope (degrees) (average),
• metadata:
  • name of the alpine massif,
  • date of acquisition,
  • type of acquisition (ascending/descending),
• physics
  • Liquid Water Content (km/m2),
  • snow height (m),
  • minimum snowpack temperature (Celsius degree).

The 9 channels are in the following order:

• Sentinel-1 polarimetric channels: VV, VH and the combination C: VV/VH in linear,
• Topographical features: altitude, orientation, slope
• Polarimetric ratio with a reference summer image: VV/VVref, VH/VHref, C/Cref**

** The reference image selected is that of August 9th 2020, as a reference image without snow (cf. Nagler&al)

An overview of the distribution and a summary of the sample statistics can be found in the file info.pdf.

The data is stored in .hdf5 format with gzip compression. We provide a python script to read and request the data. The script is dataset_load.py. It is based on the h5py, numpy and pandas libraries. It allows to select a part or the whole dataset using requests on the metadata. The script is documented and can be used as described in the README.md file

The processing chain is available at the following Github address.

The authors would like to acknowledge the support from the National Centre for Space Studies (CNES) in providing computing facilities and access to SAR images via the PEPS platform.

The authors would like to deeply thank Mathieu Fructus for running the Crocus simulations.

Erratum :

In the dataloader file, the name of the "aquisition" column must be added twice, see the correction below.:

dtst_ld = Dataset_loader(path_dataset,shuffle=False,descrp=["date","massif","aquisition","aquisition","elevation","slope","orientation","tmin","hsnow","tel",],) 

If you have any comments, questions or suggestions, please contact the authors: 

• matthieu.gallet@univ-smb.fr
• fatima.karbou@meteo.fr
• abdourrahmane.atto@univ-smb.fr
• emmanuel.trouve@univ-smb.fr