Post-doctoral researcher

October 2023 - ongoing Currently with the ECEO lab of EPFL. Working with Devis Tuia and Marc Rußwurm on ADOPT: AI for Detecting Ocean Plastic Pollution with Tracking, a project founded by the Swiss Data Science Center.

Post-doctoral researcher

March 2022 - September 2023 I did my post-doc within the ASTRAL project (Apprentissage StaTistique pour l’imageRie sAr muLtidimensionnelle), financed by the joint ASTRID program of ANR and AID. During my postdoc I worked with Florence Tupin, Loïc Denis, Clément Rambour and Nicolas Thome. Among the objectives of the project are the exploration of network architectures suitable for multi-dimensional complex SAR data (inteferometric, polarimetric and tomographic data) and the estimation of uncertainties associated to the network prediction.

Ph.D candidate

January 2019 - March 2022 The subject of my Ph.D was: “Deep Learning for SAR imagery: from denoising to scene understanding”. Throughout my Ph.D thesis, I have explored supervised, semi-supervised and unsupervised deep learning approaches for speckle reduction from Synthetic Aperture Radar images. The fluctuations caused by speckle seriously limit the exploitation of SAR data. The speckle phenomenon has a deterministic yet unpredictable phenomenon: it is often modeled as a random variable and treated as a multiplicative noise. At first, we proposed to create a dataset of groundtruth images to train a deep model for speckle reduction, namely SAR-CNN. To account for specific acquisition modality, introducing speckle spatial correlation, we then introduce a semi-supervised approach to train neural networks directly on SAR data:SAR2SAR. The algorithm relies on the noise2noise framework and on SAR time-series. However, one does not always have access to more images acquired over a same area. With MERLIN we show that, somehow counterintuitively, one can exploit the phase of a single-look complex SAR image to produce two independent subimages, namely the real and the imaginary part, creating ideal conditions to perform unsupervised learning on single SAR images.