Fusing InSAR Coherence into Multi-Modal Flood Mapping

Adding Sentinel-1/NISAR InSAR coherence to the GEOID-Flood benchmark and studying how to fuse it with existing SAR, optical and DEM modalities to improve flood segmentation.

Requirements

  • M.Sc. in Data Science, Computer Science, Artificial Intelligence, Mathematics, or similar
  • Knowledge of Python and deep learning frameworks (PyTorch)
  • Basic understanding of SAR imaging and interferometry (InSAR)
  • Familiarity with geospatial raster data (Rasterio) or willingness to learn
  • Interest in multi-modal data fusion for semantic segmentation

Description

GEOID-Flood is a LINKS Foundation benchmark for flood segmentation, built from 219 flood events across 65 countries and combining pre/post-event Sentinel-1 (GRD and RTC), pre-event Sentinel-2, and the Copernicus GLO-30 DEM, with manually validated background/permanent-water/flooded-water labels. Current models process these modalities to detect flooded areas, but none of them exploit InSAR coherence: the loss of phase correlation between two SAR acquisitions is a well-known, physically grounded signal for surface change, and drops sharply over areas disturbed by flooding (submerged vegetation, moving water, damaged structures), potentially detecting inundation missed by amplitude-based SAR or optical channels alone.

The goal of this thesis is to add InSAR coherence derived from Sentinel-1 (and, where available, NISAR) to the GEOID-Flood dataset and study whether it improves flood segmentation performance. We already have a pipeline to compute pre/post-event coherence pairs, so the first step of the work is downloading and integrating coherence for the GEOID-Flood events into the existing dataset structure. The core research question is about fusion: coherence is a very different signal from SAR backscatter or optical reflectance, with its own noise characteristics and failure modes (e.g. decorrelation from vegetation or long temporal baselines unrelated to flooding), so simply stacking it as another input channel may not be the best approach. The thesis will explore and compare different strategies to fuse InSAR coherence with the existing modalities, from early fusion to modality-specific encoders and learned/attention-based fusion mechanisms, building on the TerraTorch-based training pipeline already used in GEOID-Flood.

Main Activities

  • Literature Review: Surveying InSAR coherence for change/flood detection and multi-modal fusion strategies for segmentation.
  • Coherence Data Acquisition: Downloading Sentinel-1 (and optionally NISAR) InSAR coherence pairs for the GEOID-Flood events using the existing coherence pipeline.
  • Dataset Integration: Extending the GEOID-Flood dataset structure and loaders to include the new coherence modality.
  • Fusion Strategy Design: Designing and implementing candidate fusion approaches (early, late, learned/attention-based) to combine coherence with SAR, optical and DEM inputs.
  • Benchmarking: Training and evaluating the fused models against the existing GEOID-Flood baselines, and analyzing where coherence helps or hurts.

Contacts