Romain Thoreau - PhD in Machine Learning for Environmental Sciences

About me

I am an Assistant Professor (maitre de conférences in French) in computer science at AgroParisTech in the MIA Paris-Saclay lab. I am interested in machine learning for environmental sciences. Currently, I focus on multimodal and multisource representation learning, with applications to the monitoring of industrial pollution and wildfire simulation.

Previously, I was a postdoctoral research fellow at CNES in collaboration with the SPE lab from CNRS - Corsica University. I completed my PhD (manuscript) at ONERA, Magellium and ISAE-SUPAERO in close collaboration with the Toulouse Mathematics Institute on the semantic segmentation of airborne hyperspectral images for mapping impermeable surfaces in large urban areas.

I try to conduct my research following the principles of Slow Science and Open Science. I also attempt to limit the carbon footprint of my research, and I therefore do not travel by plane to attend conferences.

Open positions

Research internship in machine learning for multimodal data analysis co-supervised with Vincent Guigue.

News

June 2025: Our papers introducing DEFLECT and p3VAE have been accepted to ICCV 2025 and Springer Machine Learning, respectively!

June 2025: For the tenth MILO seminar, we were glad to have a presentation of Maud Biquard (ISAE-SUPAERO / CNES) about image restoration with compressive variational autoencoders!

May 2025: For the ninth MILO seminar, we were delighted to welcome Benoit Dufumier (CEA) and Javiera Castillo-Navarro (CNAM) for a presentation about multimodal contrastive learning!

April 2025: For the eighth MILO seminar, we were glad to welcome Luca Savant from Turin University for a great talk about Gaussian Splattings for Earth observation!

Selected publications

	Can multimodal representation learning by alignment preserve modality-specific information? R. Thoreau, J. Levillain and D. Derksen Accepted to MACLEAN workshop @ ECML-PKDD 2025. [arXiv] [code] [slides]
	Parameter-Efficient Adaptation of Geospatial Foundation Models through Embedding Deflection R. Thoreau, V. Marsocci and D. Derksen (Equal contribution) ICCV 2025*. [paper] [arXiv] [code] [poster] [video]
	Physics-informed Variational Autoencoders for Improved Robustness to Environmental Factors of Variation R. Thoreau, L. Risser, V. Achard, B. Berthelot and X. Briottet Springer Machine Learning. [paper] [arXiv] [code] [slides]
	Toulouse Hyperspectral Data Set: a benchmark data set to assess semi-supervised spectral representation learning and pixel-wise classification techniques R. Thoreau, L. Risser, V. Achard, B. Berthelot and X. Briottet ISPRS Journal of Photogrammetry and Remote Sensing. [paper] [data set] [code] [arXiv]
	Active Learning for Hyperspectral Image Classification: a Comparative Review R. Thoreau, V. Achard, L. Risser, B. Berthelot and X. Briottet IEEE Geoscience and Remote Sensing Magazine. [paper] [code]

Teaching

Current courses

I teach computer science (programming with Python / SQL, machine learning, image analysis...) to undergraduate and graduate students at AgroParisTech. I am also involved, as a non-expert, in classes of ethics, history and philosophy of sciences.

Past courses (2020 - 2025)

Machine learning (lectures) at Paul Sabatier University, M1 MAPI3,
Introduction to machine learning (lectures) at Paul Sabatier University, 3rd year BS in Mathematics / Computer Science,
Introduction to statistics (lectures) at Paul Sabatier University, 2nd year BS in Biochemistry,
Integration and probability theory (tutorials) at ENSEEIHT engineering school,
Statistical foundations of Machine Learning (practical works) at ISAE-SUPAERO, Data & Decision Sciences MS,
Combinatorial optimization (practical works) at Paul Sabatier University, Computer Science MS,
Algorithmic (practical works) at Paul Sabatier University, Computer Science BS,
Introduction to python (practical works) at Paul Sabatier University, Civil Engineering BS.

Student supervision

April to October 2025, Hugo André, final-year student at ISAE-SUPAERO on deep generative modeling for the inversion of industrial stack plumes from satellite images, in collaboration with Pierre-Yves Foucher (ONERA), Camille Desjardins, Laura Le Barbier and Solange Lemai-Chenevier (CNES).
March to August 2023, Arthur Cassou, final-year student at ENSTA Bretagne on the integration of physical priors in generative models for hyperspectral data, in collaboration with Laurent Risser, Véronique Achard and Xavier Briottet.
May to July 2022, Gaspard Loupit, second-year student at ISAE-SUPAERO on the comparison of features extracted by spatial-spectral convolutions to morphological profiles through LIME explanations for hyperspectral images, in collaboration with Véronique Achard.
April to July 2022, Clément Lefranc, master student in computer science at Paul Sabatier University on the development of a QGIS plug-in for active learning on Earth observation data, namely AL4EO.

MILO scientific seminars

MILO seminars are scientific seminars about mathematics and computing, e.g. machine learning, applied or applicable to Earth observation. I've been organizing the seminars since May 2024 within the COMET Signal and Image Processing community, a community of experts initiated by CNES, the French spatial agency, in order to gather academic researchers and engineers from the space industry in France. If you would like to present your work, please reach out!

[Past and coming seminars]