Next month, I will be joining INRIA in Montpellier, within the EVERGREEN research team. I am very much excited!
EVERGREEN is basically sitting at the intersection of three domains:
- Earth observation
- Machine learning
- Environmental & agricultural applications
Their name stands for Earth obserVation and machine lEarning foR aGRo-Environmental challENges.
The project
I will have the chance to join them as a research engineer on a project named GEO-ReSeT (Generalized Earth Observation with Remote Sensing and Text). This project is targeting the training of a geospatial foundation-model presenting capabilities across text, remote sensing imagery and features in maps. All these modalities will be linked thanks to their location on the globe. The goal here is to have a model capable of zero shot or few shot solutions on land cover, land-use mapping or visual Q&A in domains like environmental monitoring, urban planning, agriculture and so on.
What I have been doing so far
I stepped into research in may 2024 while interning at Miami University under the supervision of Dr Philippe Giabbanelli. That’s the first time I got to tinker with LLMs, back when GPT-4o was just getting released. I used LLMs to study how we could create empathetic stories from agent-based models for decision makers but also assessed their abilities to learn conceptual modeling on the fly. I had a lot of fun in doing so and I learnt a lot. And that’s most probably the first time I had this feeling of « just one more improvement and then I’ll stop » kind of feeling. The one that keeps you so engaged in the project that you never feel like working.
Then for my last internship before graduation, I decided to join Airbus Defence and Space. I wanted to get an opinion of both the research and the industry while still being a student. Again, it was a very pleasant experience. I discovered remote sensing for the first time and the amazing applications that come with it. I got to meet amazing people and there also I learnt a lot, especially on vision-language models and fine-tuning methods. What I especially liked there was to see all the different kind of problems being tackled in agriculture, environmental monitoring etc. In that sense it’s satisfying to see the practical application of your work.
And after graduation?
Right after this internship, I graduated from my engineering school and took some time off in India. I made the most of this time to focus on some personal projects to learn things I never really got the time to do, mingled with coding agents and agentic engineering, and lately I have been involved on other research projects with Dr Philippe Giabbanelli on topics like agentic engineering workflows, large language models to explain agent-based simulations and to connect causal maps. I also got the chance to travel a bit in North India and Thailand.
Thar Desert, India
Bangkok, Thailand
I have been wondering (a lot) what to do and where to go after graduation. I have tasted a small portion of research and had a little bite in the industry, which both have their assets and drawbacks. Given the current state of the job market as a junior it has not been an easy time to find something appealing or to find something at all. But this « time off » in India gave me some time to think about it and eventually I have decided to stick to research after graduation. The first reason is obviously because I enjoy doing research a lot. It seems to me that it is a lot more flexible than corporate life but most of all it is a lot more rewarding. What I mean is that you are more likely to be in control of your work, explore some areas you wouldn’t otherwise and get the opportunity to share your work with the world. Moreover I believe it is a great way to learn and gain expertise, which we don’t have after graduation, especially in generalist engineering schools like mine. Ok we have done a few projects classifying cats and dogs here and there, but even if internships were nice experiences, it fundamentally leave you after graduation with so much left to learn and so much to explore. While I was trying to figure out my first job, someone gave me the following advice: « choose the place where you will learn the most ». It is dead simple as an advice, yet this is the best advice I ever had.
More than simply learning things and enjoying research, I find it exciting to work on the forefront of human knowledge, open source the fruits of your work and trying to tackle challenges that matter. What I like about the EVERGREEN team is that they mix everything I have been doing so far (AI and remote sensing) towards problems that I believe are very important to face (deforestation monitoring being one of them).
What I will focus on
In order to train the foundation-model of GEO-ReSeT, we will be needing text. Specifically we will be needing unstructured text, linked to geographic named entities and their geographic footprint. Collecting such data will most probably need to involve LLMs at some point to disambiguate entities based on the context provided within the text descriptions.
I already tried to collect some modest datasets in the past, but this time it will be a serious opportunity to do so for a real project, grounded in literature and supervised by experts in the field. I couldn’t be more excited!
Special thanks to everyone I have met along the way who helped me in one way or another towards this new chapter :)