Novel machine-learning-based approach for automated snow avalanche detection from SAR images for the Austrian alps

Kathrin Lisa Kapper*, Stefan Muckenhuber, Thomas Gölles, Andreas Trügler, Jakob Abermann, Eirik Malnes, Muhamed Kuric, Jakob Grahn, Alexander Prokop, Birgit Schlager, Wolfgang Schöner

*Corresponding author for this work

Research output: Contribution to conferenceAbstractpeer-review


Knowledge of the snow avalanche coverage and activity of a larger region is essential for a variety of applications, e.g. avalanche warnings and hazard management. Recent advances in automated machine-learning-based algorithms set the scene for fast and comprehensive detection of avalanches from satellite images. In the framework of this project, we develop an automated state-of-the-art avalanche-detection system for the Austrian Alps, including a best-practice data-processing pipeline and a learning-based approach applied to synthetic aperture radar (SAR) satellite images. For this purpose, we make use of the openly available Copernicus Sentinel-1 SAR images that have successfully been used for avalanche detection, together with a variety of published training data sets. In a first step, the labelled training data, which comprise around 26 000 manually detected avalanche outlines from Switzerland and Greenland, were downloaded and preprocessed. The SAR images were selected to correspond to the regions and time slots of the training data and were preprocessed to yield optimum detection results. In addition, SAR images from the Austrian Alps from an avalanche-rich winter season will be used to evaluate how well our detection algorithm generalizes to this independent data that is potentially differently distributed from the training data. Furthermore, selected ground-truth data from Switzerland, Greenland, and Austria will allow us to validate the accuracy of the detection approach. As a novel approach to improve detection performance, we propose to include encoded weather data into the avalanche detection pipeline. The weather data include several meteorological parameters, such as precipitation and wind speed, over a certain time range that are downscaled to fit the corresponding pixels of the SAR image. In this way, simple averages of meteorological parameters but also full snowpack models can be aggregated over time.
Original languageEnglish
Publication statusPublished - 30 Sep 2022
EventInternational Symposium on Snow - Davos, Switzerland
Duration: 25 Sep 202230 Sep 2022


ConferenceInternational Symposium on Snow
Internet address


Dive into the research topics of 'Novel machine-learning-based approach for automated snow avalanche detection from SAR images for the Austrian alps'. Together they form a unique fingerprint.

Cite this