PhD defense Zoé Berenger: Deep learning and SAR tomography for monitoring forest structures
Thursday 11 December, 2025, at 14:00 (Paris time) at Télécom Paris
Télécom Paris, 19 place Marguerite Perey F-91120 Palaiseau [getting there], amphi 2 and in videoconferencing
Jury
-
Yajing Yan, Maı̂tre de conférences, Université Savoie Mont-Blanc (Rapporteuse)
-
Laetitia Thirion-Lefèvre, Professeure, CentraleSupélec, Université Paris Saclay (Rapporteuse)
-
Christian Germain, Professeur, Université de Bordeaux (Examinateur)
-
Andreas Reigber, Professor, DLR (Examinateur)
-
Florence Tupin, Professeure, Télécom Paris, Institut Polytechnique de Paris (Directrise de thèse)
-
Loı̈c Denis, Professeur, Université Jean Monnet Saint-Etienne (Co-encadrant de thèse)
-
Laurent Ferro-Famil, Professeur, ISAE-SUPAERO/CESBIO, Université de Toulouse (Co-encadrant de thèse)
Abstract
Forests regulate the Earth’s climate by absorbing and storing carbon, but monitoring their height and structure at global scale is difficult: field measurements are limited, and existing radar-based 3D reconstruction methods either lack precision or are too computationally slow. Synthetic Aperture Radar (SAR) tomography offers canopy-penetrating and all-weather imaging, yet current techniques face trade-offs that limit their suitability for upcoming missions such as ESA’s BIOMASS satellite.
Learn more
This thesis investigates whether deep learning can improve SAR tomography reconstruction over forested areas. A first contribution consists in training an encoder-decoder network on simulated data in a supervised manner to refine coarse tomographic reconstructions. This method produces sharper vertical forest profiles than traditional approaches while remaining computationally efficient, proving that deep learning is a suitable tool for this problem. A second, self-supervised method removes the need for simulated references by enforcing consistency with SAR measurements and constraining solutions through a new regularized learning strategy inspired by Equivariant Imaging. Tested on diverse forest datasets, it achieves results comparable to the supervised approach and superior to conventional techniques, with fast inference suitable for large-scale monitoring.