The research

Work packages

The chair is organised as a pipeline: new geometric features to represent multivariate time series (WP1), new geometric learning models to classify and fuse them (WP2), and new geometric generative models to denoise, interpolate and predict (WP3). Everything is evaluated on real, public remote-sensing datasets, and released as an open toolbox. Frugality is measured, not assumed — energy consumption is tracked with tools such as CodeCarbon.

WP 1 Geometric invariant features for InSAR & PolInSAR N. Le Bihan · A. Mian · P.-O. Amblard · G. Ginolhac Geometric phases as low-dimensional, gauge-invariant signatures of ground displacement — glaciers, volcanoes, subsidence. Read more → WP 2 New geometric machine learning models F. Chatelain · G. Ginolhac · N. Le Bihan · A. Mian Extending SPDNet to transformer architectures for the temporal dimension, and to federated learning for multimodal data. Read more → WP 3 Flow matching on manifolds G. Ginolhac · Y. Mhiri · N. Le Bihan · F. Chatelain Riemannian generative models for denoising SAR imagery, and interpolating or predicting images between acquisition dates. Read more →

Datasets Sentinel-1/2, TerraSAR-X, UAVSAR — InSAR series over Mexico City and volcanoes, via existing collaborations incl. ONERA.

Toolbox All algorithms released as an open-source toolbox for the community.

Beyond remote sensing Seismic and sonar data with industrial partners Géolithe and Exail; EEG with the DataIA institute.