Multi-Sensor Wildfire Smoke Delineation across Geostationary and Polar-Orbiting Satellites

Exploring deep learning smoke segmentation across heterogeneous satellite sensors, with a focus on adapting GOES-based models to European EO platforms.

Requirements

  • M.Sc. in Data Science, Computer Science, Artificial Intelligence, Mathematics, or similar
  • Strong knowledge of Python and deep learning frameworks (PyTorch, Lightning)
  • Basic background in computer vision, preferably semantic segmentation or domain adaptation
  • Basic knowledge of remote sensing and multispectral satellite imagery

Description

Accurate delineation of wildfire smoke plumes is critical for near-real-time air quality assessment, fire spread monitoring, and emergency response. Smoke is spectrally ambiguous with clouds, geometrically diffuse, and highly variable across fire regimes — making it a challenging target for automated segmentation.

SmokeViz (NeurIPS 2025) provides a strong starting point: over 160,000 NOAA expert smoke annotations on GOES ABI imagery, refined via pseudo-label dimension reduction (PLDR) to align coarse temporal labels with the most representative satellite frame. This dataset opens the door to data-driven smoke segmentation, but is limited to North American wildfires and a single sensor family.

This thesis explores whether and how such models can be extended to European operational sensors — primarily the recently launched MTG-FCI (geostationary, 10-minute full-disk, spectral channels similar to GOES ABI) — and complementary polar-orbiting instruments: VIIRS, Sentinel-3 OLCI, and Sentinel-2 MSI. The GOES→MTG transition is a natural first step given the spectral proximity of the two sensors, though differences in calibration, viewing geometry, and European fire regimes may introduce non-trivial distribution shifts worth characterizing.

The research is deliberately exploratory: the specific methodology (zero-shot transfer, domain adaptation, pseudo-label refinement, multi-sensor fusion, or a combination) will be guided by early experimental results. A key open question is how well GOES-trained models transfer out of the box, and how much labeled European data is actually needed to close any remaining gap.

Main Activities

  • Literature review on satellite smoke segmentation, domain adaptation for remote sensing, and multi-sensor fusion
  • Characterization of spectral band correspondences and distribution shifts between GOES ABI and MTG-FCI
  • Data collection and preprocessing for European fire events across target sensors (MTG, VIIRS, OLCI, Sentinel-2), cross-referenced with EFFIS and Copernicus EMS records
  • Adaptation and fine-tuning experiments starting from SmokeViz-pretrained models
  • Evaluation of smoke vs. cloud discrimination and geographic generalization across fire regimes

Contacts