dcodrut/dl4gam_alps

A Multi-modal Dataset for Glacier Mapping (Segmentation) in the European Alps

Contents

• Introduction
• Dataset components
• Other specifications
• Patch sampling
• Geographic cross-validation
• How to cite

Introduction

The dataset consists of Sentinel-2 images from 2015 (mainly), 2016 and 2017, and binary segmentation masks for glaciers, based on an inventory built by glaciology experts (Paul et al. 2020). Secondly, given that glacier ice is not always visible in the images, due to seasonal snow, shadow/cloud cover, and debris cover (which is particularly significant), the dataset also includes additional features (see additional components) that can help in the segmentation task.

The glacier inventory contains 4395 glaciers covering 1805.9 km². We consider only the glaciers larger than 0.1 km² (n = 1646). After a manual QC, we additionally removed 53 glaciers for which the images were of poor quality (e.g. too many clouds) and for which alternative dates within the same year could not be found. We end up with 1593 glaciers but covering 93.3% (1684.7 km²) of the glacier inventory area. The dataset was constructed by collecting images for each glacier individually, with a 1.28 km buffer around each glacier (e.g. see Example for Aletsch Glacier).

This repository includes both patchified and glacier-wide raster versions. An example of one patch with all the features: Additionally, we provide glacier-wide rasters for 2023, which can be used for glacier area change detection (see https://dcodrut.github.io/dl4gam_alps/). preds contains the predictions of an U-Net ensemble (see our paper).

Dataset components

Main components

• Sentinel-2 images (all bands, including cloud and shadow masks which can be used for loss masking)
• glacier mask containing the IDs of the glaciers present in the image or -1 (i.e. missing) for non-glacier pixels; this can be turned into a binary mask for segmentation

Additional components

• six glacier masks: the current glacier (on which the patch/image is focused) plus five additional buffers (-20 m, -10 m, 10 m, 20 m and 50 m) that can be used for customized loss functions/metrics, e.g. the False Positive Rate on the 20 to 50 m buffer
• debris mask (0: no debris, 1: debris) based on a mix of three publications (Scherler et al. 2018, Herreid & Pellicciotti 2020, Linsbauer et al. 2021). This debris mask could be used for e.g. measure the performance of the models on debris-covered glaciers or for a multi-class segmentation task. However, except for Switzerland, the masks should be used with caution (low quality due to how the masks were derived and temporal mismatches). The three sources were combined as follows:
  • we used Linsbauer et al. 2021 for the glaciers in Switzerland;
  • Herreid & Pellicciotti 2020 for those outside Switzerland and larger than 1 km²;
  • Scherler et al. 2018 for the rest.
• DEM (Copernicus GLO-30) + five derived features (using xDEM): slope, aspect, terrain ruggedness index, planform and profile curvatures
• dh/dt (surface elevation change) map over 2010-2015 (Hugonnet et al. 2021); this can provide information on the glacier parts that are not visible in the images e.g. due to debris but which are still melting (=> losing elevation)
• v (surface velocity) map over 2015 (ITS_LIVE)

Other specifications:

• temporal coverage: one acquisition per glacier, from either 2015 (mainly), 2016, or 2017; and one acquisition per glacier for 2023 (glacier-wide rasters only, without ground truth glacier masks)
• spatial coverage: all the glaciers in the European Alps larger than 0.1 km², spreading over four countries: Austria, Italy, France, Switzerland.
• 2251 patches sampled with overlap from the 1593 glaciers; or 11440 for the large version, obtained with an increased sampling overlap; see Patch sampling
• the dataset download size is 5.8 GB (11 GB when unarchived); or 29.5 GB (52 GB when unarchived) for the large version
• the dataset is provided at 10m GSD (after bilinearly resampling some of the Sentinel-2 bands and the additional features which come at a lower resolution)
• the dataset provides fixed training, validation, and test geographical splits (70-10-20, by glacier area)
• five different splits are provided, according to a five-fold cross-validation scheme (see Geographic cross-validation)
• all the features/masks are stacked and provided as NetCDF files (one or more per glacier), structured as data/{glacier_id}/{glacier_id}_{patch_number}_{center_x}_{center_y}.nc
• data is projected and geocoded in local UTM zones

Patch sampling

Given each glacier's raster, we sample one or more patches, of size 256x256 (i.e. 2.56 km x 2.56 km) as follows:

• version small (2251 patches):
  • we sample one patch with the center on the glacier's centroid
  • we build a grid of patch centers with a step size of 128 pixels and keep all the patches which have the center on the (current) glacier
• version large (11440 patches, used in our paper):
  • additional to the centroid, we sample four patches from the "edges"
  • we use again the grid sampling but with a 64 pixels step

We show below an example of the two versions of sampled patches (together with their centers) for one glacier (Glacier de la Pilatte):

Geographic cross-validation

Given that we are interested in glacier area change estimation and we want estimates for all the glaciers, we use a five-fold cross-validation scheme such that we avoid using the training results. Then, a further split is performed in each iteration to extract a small validation fold. This acts as a spatial gap between training and testing folds (otherwise it's difficult to have a perfect separation given the irregularity of the glacier shapes). The IDs of the glaciers from each fold and cross-validation iteration are provided in map_all_splits_all_folds.csv. The scheme is depicted below, followed by the map of the first iteration:

How to cite

If you use this dataset in your research, please cite the following paper:

• https://doi.org/10.22541/essoar.173557607.70204641/v1