This software automatically delineates calving front positions from Landsat-8 and Landsat-9 Level-1 data archives. It processes multispectral imagery using a specialized artificial neural network (ANN). In particular, it uses a convolutional neural network of the U-Net architecture to semantically segment images into a glacier/land and water class. The glacier calving front, which is described by the boundary between these two classes, is then extracted by vectorizing and masking the model prediction. The figure below gives a broad overview of the processing workflow.
The model is trained using manually delineated calving front position from 19 Greenland glaciers. We evaluated the performance of this model using three independent test datasets, including glaciers in Svalbard, at the Antarctic Peninsula and in Patagonia. The quality of the automatically extracted calving fronts is comparable to that of manually delineated calving fronts. The processing system and its validation are described in detail in this article.
We offer three options for using this software. We recommend using the containerized implementation via Docker (option 1) or Singularity (option 2). It is also possible to download the ANN weights and use the routines in this repository (option 3).
The docker container contains the code, model weights and all requirements to run the processing system (tested for Docker version 23.0). Download the container by running:
docker pull eloebel/glacier-front-extraction:latest
Place the Landsat archives to be processed in an input folder. Run the container and define the path of this input folder and set the glacier name as an environment variable.
docker run --volume=/home/INPUT_IMAGES:/input --env glacier=daugaard_jensen eloebel/glacier-front-extraction:latest
Glaciers outside the reference dataset can also be processed. To do this, set the glacier name to custom and define lon and lat coordinates (decimal degrees) as environment variables. The processing window is approximately 15 km by 15 km, centered on these defined coordinates. Example for a calving front extraction for 28.57°W and 71.91°N:
docker run --volume=/home/INPUT_IMAGES:/input --env glacier=custom --env lon=-28.57 --env lat=71.91 eloebel/glacier-front-extraction:latest
The ANN calving front prediction is stored in a separate folder within the defined input directory. The prediction contains the coastline, the calving front (LineString Shapefile), the prediction mask (geoTiff) and an overview figure.
Docker contains can also be run with Singularity (tested for Singularity version 3.8). Download the container and convert into a Singularity image by running:
singularity pull docker://eloebel/glacier-front-extraction:latest
Place the Landsat archives you want to process in an input folder. Run the container and bind the path to this input folder. In addition, you have to bind a data folder which is needed for storing temporary data. The glacier name is defined using an environment variable.
singularity run --bind /home/INPUT_IMAGES:/input --bind /home/DATA:/data --env glacier=daugaard_jesen glacier-front-extraction_latest.sif
Glaciers outside the reference dataset can also be processed. To do this, set the glacier name to custom and define lon and lat coordinates (decimal degrees) as environment variables. The processing window is approximately 15 km by 15 km, centered on these defined coordinates. Example for a calving front extraction for 28.57°W and 71.91°N:
singularity run --bind /home/INPUT_IMAGES:/input --bind /home/DATA:/data --env glacier=custom --env lon=-28.57 --env lat=71.91 glacier-front-extraction_latest.sif
The ANN calving front prediction is stored in a separate folder within the defined input directory. The prediction contains the coastline, the calving front (LineString Shapefile), the prediction mask (geoTiff) and an overview figure. The data folder containing temporary data (especially the pre-processed satellite images) can be deleted. Otherwise, the ANN is applied to the data in subsequent container runs.
Clone the repository and install all required python packages. You can use the requirements.txt with conda:
conda create --name <env_name> --file requirements.txt
Download the ANN model weights from the latest release and unpack them into the main repository folder. Place the Landsat archives to be processed in the input folder (see placeholder).
python GLACIER_FRONT_EXTRACTION.py daugaard_jensen
Glaciers outside the reference dataset can also be processed. To do this, set the glacier name to custom and define lon and lat coordinates (decimal degrees) as environment variables. The processing window is approximately 15 km by 15 km, centered on these defined coordinates. Example for a calving front extraction for 28.57°W and 71.91°N:
python GLACIER_FRONT_EXTRACTION.py custom -28.57 71.91
The ANN calving front prediction is stored in a separate folder within the defined input directory. The prediction contains the coastline, the calving front (LineString Shapefile), the prediction mask (geoTiff) and an overview figure.
This is an example of our software applied to Landsat images of the Daugaard Jensen Glacier in East Greenland. To quantify the change in the position of the calving front, you can use our implementation of the rectilinear box method.
The following Greenland glaciers can be selected directly from the environment variable. For these glaciers, the processing system extracts not only the entire coastline but also the masked glacier front. Glaciers outside the reference dataset can also be processed. To do this, set the glacier name to custom and define the lon and lat coordinates (decimal degrees).
| Glacier | environment variable |
|---|---|
| Kangiata Nunaata Sermia | kangiata_nunaata_sermia |
| Helheim Glacier | helheim |
| Kangerdlussuaq Glacier | kangerdlugssuaq |
| Jakobshavn Isbræ | jakobshavn_isbrae |
| Eqip Sermia | eqip_sermia_region |
| Store Glacier | store |
| Rink Isbræ | rinks_isbrae |
| Daugaard Jensen Glacier | daugaard_jensen |
| Ingia Isbræ | ingia_isbrae |
| Upernavik Isstrøm | upernavik_isstrom_north |
| Waltershausen Glacier | waltershausen_glacier |
| Hayes Glacier | hayes |
| Sverdrup Glacier | sverdrup |
| Kong Oscar Glacier | kong_oscar |
| Døcker Smith Glacier | docker_smith |
| Harald Molke Bræ | harald_moltke_brae |
| Tracy Glacier | tracy_gletscher |
| Humboldt Glacier | humboldt |
| Zachariae Isstrøm | zachariae_isstrom |
| Nioghalvfjerdsbræ | nioghalvfjerdsbrae |
| Hagen Bræ | hagen_brae |
| Academy Glacier | academy |
| Ryder Glacier | ryder_glacier |
If you find our software helpful and use it in your research, please use the following BibTeX entry.
@Article{loebel2023,
AUTHOR = {Loebel, E. and Scheinert, M. and Horwath, M. and Humbert, A. and Sohn, J. and Heidler, K. and Liebezeit, C. and Zhu, X. X.},
TITLE = {Calving front monitoring at sub-seasonal resolution: a deep learning application to Greenland glaciers},
JOURNAL = {The Cryosphere Discussions},
VOLUME = {2023},
YEAR = {2023},
PAGES = {1--21},
URL = {https://tc.copernicus.org/preprints/tc-2023-52/},
DOI = {10.5194/tc-2023-52}
}
Erik Loebel
TU Dresden | Geodetic Earth System Research
tu-dresden.de/geo/ipg/gef
erik.loebel@tu-dresden.de
@ErikLoebel