Stage M2 Exploring Pretext Tasks for Building Statistically Constrained Latent Spaces of SAR Time Series

This internship, hosted at the ICube laboratory in Strasbourg (Illkirch), focuses on self-supervised learning for Sentinel-1 SAR satellite image time series. SAR data enables day-and-night Earth monitoring regardless of cloud cover, but deep networks need costly labeled data. Self-supervised learning addresses this by training an encoder, via pretext tasks, to build a generalizable latent space reusable with only a few labels.
Existing pretext tasks, such as masked autoencoders and denoising, were designed for optical or Gaussian-like data and suit SAR poorly. SAR signals are affected by multiplicative, heavy-tailed speckle and by acquisition geometry (layover, shadow, foreshortening). Standard augmentations and Euclidean ℓ2 losses therefore yield latent spaces that are neither physically meaningful nor robust.
The goal is to build latent spaces consistent with SAR statistics at two levels: (1) physics-aware pretext tasks, with speckle and temporal resampling, geometry-aware masking, and divergence-based losses (Kullback–Leibler, Rényi, soft-DTW); and (2) latent representations tied to robust statistical models, such as compound-Gaussian parameters or covariance matrices on Riemannian manifolds.
The intern will first implement standard pretext tasks and help define a few-label evaluation protocol (linear probing). They will then design SAR-consistent augmentations and divergence-based losses, studying their differentiability and numerical stability. Results may be published (GRETSI, IGARSS, ICASSP).

Contacts: Antoine Bralet : abralet@unistra.fr, Matthieu Gallet (matthieu.gallet@univ-smb.fr), Yassine Mhiri (yassine.mhiri@univ-smb.fr)

Link to full offer: https://fichiers.numerique.gouv.fr/explorer/items/files/c5e2b561-ce8a-417a-beb4-d7b112200be0