Post-Doctoral Research Visit F/M Learning Fine-Grained, Temporally-Aware Geolocation

Contexte et atouts du poste

This post-doctoral position is part of the EARTH-FM project (Enhancing Archival Remote sensing data for Training Holistic Foundation Models), a joint INRIA-CNES challenge aimed at advancing the state of the art in multi-modal remote sensing Foundation Models. The project exploits large multi-temporal and multi-scale remote sensing image archives collected by CNES, such as Spot World Heritage and Pléiades World Heritage, while also exploring methodologies for data acquired through upcoming Earth Observation missions (e.g. CO3D). The work addresses challenges specific to remote sensing data, including high diversity in acquisition conditions and the integration of new modalities.

The recruited fellow will join the

Mission confiée

The objective of this position is to combine implicit neural representations and self-supervised multi-sensor learning paradigms to develop spatio-temporal Earth embeddings for agricultural and environmental applications. Building on recent advances in geographic Implicit Neural Representations (INRs), the model will jointly encode geographic location and temporal dynamics into compact vector representations, addressing a key limitation of current geospatial AI approaches, which either ignore fine-grained geographic context or fail to capture temporal dynamics. Self-supervised learning will allow the model to exploit large volumes of unlabeled Satellite Image Time Series (SITS), extracting robust, task-agnostic representations while reducing dependence on costly manual annotation.

The resulting embedding framework will be pre-trained on unlabeled SITS data, including datasets collected by the TETIS lab and/or the EVERGREEN team-project, covering French agricultural areas among others, and successively fine-tuned on downstream tasks such as land cover mapping, crop type classification, and monitoring of soil surface dynamics.

Principales activités

The fellow will design and develop a self-supervised training framework that leverages multi-modal Satellite Image Time Series (SITS), combining optical (Sentinel-2) and radar (Sentinel-1) data, jointly encoded with geographic location and temporal dynamics. A central activity will be to enhance the quality and generalization capability of the resulting spatio-temporal Earth embeddings by incorporating geographic Implicit Neural Representations (INRs) and spectrum-aware multi-sensor encoders. The work will primarily leverage publicly available crop mapping datasets, in addition to data collected by the TETIS lab and/or the EVERGREEN team-project, covering French agricultural areas.

Concrete activities include implementing and evaluating self-supervised and multi-modal co-learning architectures, integrating spatio-temporal encoding strategies for jointly modeling location and time, and benchmarking the developed embedding framework on downstream agricultural tasks including land cover mapping, crop type classification, and monitoring of soil surface dynamics. Results will be disseminated through publications in international journals and conferences, with active collaboration with the partner teams involved in the project.

Compétences

Technical skills and level required: PhD in machine learning, computer vision, remote sensing, or a closely related field. Solid background in deep learning, with hands-on experience in self-supervised learning and/or multi-modal representation learning. Familiarity with modern architectures (transformers, CNNs) and with training models at scale on GPU/HPC infrastructure. Experience working with remote sensing or geospatial imagery is a strong asset, as is knowledge of continual learning and multi-temporal data analysis.

Languages: Proficiency in Python and common deep learning frameworks (e.g. PyTorch). Working knowledge of English (written and spoken) for publishing and international collaboration. French is not required but can be an asset for interaction with CNES teams.

Relational skills: Ability to work collaboratively within a distributed, multi-team environment (INRIA project-teams and CNES experts). Good communication skills to present results clearly, listen, and exchange ideas across disciplines. Autonomy, rigor, and the capacity to organize one's own research while contributing to shared objectives.

Other valued / appreciated: Track record of publications in international machine learning, computer vision, or remote sensing venues. Curiosity about Earth Observation applications and an interest in transferring research toward operational use.

Avantages

• Subsidized meals
• Partial reimbursement of public transport costs
• Leave: 7 weeks of annual leave + 10 extra days off due to RTT (statutory reduction in working hours) + possibility of exceptional leave (sick children, moving home, etc.)
• Possibility of teleworking and flexible organization of working hours
• Professional equipment available (videoconferencing, loan of computer equipment, etc.)
• Social, cultural and sports events and activities
• Access to vocational training
• Contribution to mutual insurance (subject to conditions)

Rémunération

Gross Salary: 2788 € per month