Forecasting Avalanche Hazard by Combining Deep Learning Snow Regression with Weather Forecasts

Coupling a Sentinel-1/3 snow depth regression pipeline with weather forecast data to anticipate avalanche hazard conditions.

Requirements

  • M.Sc. in Data Science, Computer Science, Artificial Intelligence, Mathematics, or similar
  • Knowledge of Python and deep learning frameworks (PyTorch, Lightning)
  • Familiarity with geospatial raster data (Rasterio) or willingness to learn
  • Basic concepts of time series and signal processing
  • Interest in SAR/optical remote sensing and meteorological data

Description

Monitoring the snowpack is essential for water resource management, hydropower planning, and, in mountain regions, for anticipating hazards such as avalanches and rapid snowmelt. At LINKS Foundation, we are working on a deep learning pipeline that regresses snow depth over the Aosta Valley by combining Sentinel-1 SAR backscatter, Sentinel-3 LST, and a static elevation model (FABDEM), trained and validated against ground station measurements. On its own, such a pipeline only produces snapshot estimates of the current snow state from satellite observations: it cannot say how the snowpack, and the hazard it represents, will evolve over the next hours or days.

The goal of this thesis is to build an avalanche hazard forecasting system by coupling snow depth regression with numerical weather forecast data (temperature, precipitation, wind). The candidate will first work on the snow depth regression itself, consolidating and improving the Sentinel-1/Sentinel-3 pipeline. The resulting snow depth/SWE maps then become the initial condition of a forecasting component that ingests weather forecasts to project how the snowpack will change, and to derive avalanche hazard indicators (e.g. rapid accumulation, rain-on-snow, melt rate) that are recognized proxies for avalanche risk.

The thesis is therefore structured in two parts: first the snow depth regression, then coupling it with weather forecast products and defining and evaluating avalanche hazard indicators derived from the combined output.

Main Activities

  • Snow Depth Regression: Working on the Sentinel-1/Sentinel-3 snow depth regression pipeline and validating it against station data.
  • Weather Data Integration: Identifying and acquiring suitable weather forecast products (e.g. ECMWF/ARPA forecasts) and aligning them spatially and temporally with the snow regression outputs.
  • Forecasting Model Design: Developing a model that combines the current snow state with weather forecasts to predict short-term snowpack evolution.
  • Hazard Indicator Definition: Formalizing one or more avalanche hazard indicators (e.g. rapid accumulation, rain-on-snow, rapid melt) derived from the forecasted snow state.
  • Evaluation: Validating forecasts against station data and, where available, historical avalanche records.

Contacts