Landslide Delineation from Sentinel-1/2 Pre/Post-Event Imagery

Building an open multi-inventory Sentinel-1/2 dataset for landslide delineation and benchmarking pre/post-event change detection models, with optional InSAR-based forecasting.

Requirements

  • M.Sc. in Data Science, Computer Science, Artificial Intelligence, Mathematics, or similar
  • Knowledge of Python and deep learning frameworks (PyTorch)
  • Familiarity with geospatial data processing (QGIS, Rasterio) or willingness to learn
  • Basic concepts of image processing and remote sensing
  • Interest in dataset engineering and reproducible data pipelines

Description

Landslides are among the most frequent and damaging natural hazards, and rapidly identifying their extent after a triggering event (heavy rainfall, earthquakes) is critical for emergency response and risk mapping. Deep learning models that compare pre- and post-event satellite imagery can automate this delineation, but progress in the field is limited by the scarcity of large, harmonized, multi-region training data: existing landslide inventories are scattered across different sources, formats and quality levels, and are rarely paired with ready-to-use satellite time series.

We already collected a first, rough landslide inventory dataset and want to extend and consolidate it into a solid benchmark built on the European Copernicus Sentinel-1 (SAR) and Sentinel-2 (multispectral) missions, which offer free, global, frequent-revisit coverage well suited to before/after event comparisons.

The core of the thesis is a data engineering effort: reviewing candidate landslide inventories to assess whether they are usable (geometry quality, event dates, spatial coverage), acquiring the corresponding Sentinel-1/2 pre- and post-event imagery for each inventory, and packaging everything into a single, well-documented dataset released on the Hugging Face Hub. Once the dataset is ready, the second part of the thesis focuses on training and benchmarking deep learning models that operate on pre/post-event image pairs to delineate landslide extent, establishing baselines for future work.

Optionally, the work can be extended with Sentinel-1 InSAR products, either as an additional input for delineation or, in a more research-oriented direction, to explore ground-displacement time series for early landslide forecasting.

Main Activities

  • Inventory Review: Analyzing candidate landslide inventories in QGIS to check geometry quality, temporal metadata, and suitability for inclusion in the final dataset.
  • Data Acquisition: Writing Python pipelines to download Sentinel-1 and Sentinel-2 pre/post-event imagery for each selected inventory (e.g. via Copernicus Data Space / Planetary Computer APIs).
  • Dataset Packaging: Designing a data structure and format for pre/post-event S1/S2 image pairs with aligned labels, and packaging the resulting dataset.
  • Publication: Uploading and documenting the dataset on the Hugging Face Hub.
  • Modeling: Training and evaluating change-detection/segmentation models for landslide delineation on pre/post-event image pairs, establishing baseline results.
  • BONUS: Downloading and integrating Sentinel-1 InSAR products; using them as an additional delineation input or as time series for landslide forecasting.

Contacts