Offre en lien avec l’Action/le Réseau : – — –/– — –

Laboratoire/Entreprise : LITIS Lab / Université de Rouen Normandie
Durée : 36 months
Contact : paul.honeine@univ-rouen.fr
Date limite de publication : 2026-06-15

Contexte :

Sujet :
The foundation model (FM) paradigm is undoubtedly a major breakthrough in Machine Learning (ML) for Artificial Intelligence (AI). An FM is a large-scale neural network pre-trained with self-supervision on a vast unannotated dataset and designed to perform downstream tasks with minimal fine-tuning on small, annotated datasets. While FMs have made an outstanding leap in computer vision and large language models, they have not yet emerged in more complex fields, such as chemical analysis. The HyFoundationS project aims to devise an FM for chemical soil analysis, through the lens of hyperspectral imaging.

Unlike traditional cameras with their three primary colors (red, green, and blue), hyperspectral cameras capture detailed spectral information at every pixel, providing a detailed description of the material properties in the scene [1]. ML methods have been devised to explore hyperspectral images, mainly addressing spectral unmixing, classification, and segmentation tasks. With the FM paradigm reshaping the landscape of ML, there is growing interest in FMs for hyperspectral imaging, with several papers published very recently, focusing primarily on image segmentation in satellite images [2-6].

This PhD thesis is an integral part of the interdisciplinary project HyFoundationS (Hyperspectral Foundation Models for Chemical Soil Analysis). Led by the LITIS Lab, HyFoundationS aims to develop an FM for chemical analysis of soil pollution by hyperspectral imaging. In order to unleash the full potential of FMs in the analysis of soil pollution, HyFoundationS brings together an AI laboratory (LITIS), a chemistry laboratory (CARMeN), as well as a startup specialized in soil pollution analysis (Tellux). This consortium has been working together for more than 5 years, developing ML and chemical analysis for soil pollution assessment using hyperspectral cameras installed on a bench in lab conditions, which allows full environmental control over a wide variety of pollutants. HyFoundationS aims to provide major innovations to overcome key scientific and technical barriers for soil pollution analysis with FMs.

The PhD candidate will collaborate with post-doctoral fellows and senior researchers in AI, in chemical analysis, and in geosciences, to design and train a FM for chemical soil analysis. For this purpose, we aim to take advantage of recent advances in FMs for hyperspectral images and overcome their limitations as they are mainly restricted to airborne and satellite-based Earth observations, making them less relevant for chemical soil analysis. It is therefore necessary to properly define chemical analysis downstream tasks, while considering the complexity of the chemical compositions and their interactions, as well as the significant variability in the lithology matrix and the wide variety of pollutants. Therefore, developments will explore advances in cross-domain transfer learning and adapters [7-9]. Experiments and empirical evaluations will be carried out in close collaboration with CARMeN and Tellux.

Profil du candidat :
We are seeking a highly motivated PhD candidate with a strong interest in AI for science. The candidate must have an engineering or a Master 2 diploma in computer science, applied mathematics, AI, or a related field (including remote sensing). The candidate must have solid technical skills in deep learning, with experience in Python and the common ML libraries. Expertise in hyperspectral imaging is not necessary.
Candidates with a strong interest in interdisciplinary research and who can work in a collaborative environment are strongly encouraged to apply.

Formation et compétences requises :

Adresse d’emploi :
LITIS Lab, Rouen

Document attaché : 202605171426_HyFoundationS.pdf