Lisbon, Portugal · open to work remotly or relocate to any place in Europe
Geospatial engineer & Earth Observation
I work with remote sensing and deep learning , with geospatial foundation models and satellite image time series, asking whether embeddings learned in one place and one season still transfer to other domains.
Before research came two years of commercial LiDAR, photogrammetry and survey engineering, and before that a degree in art conservation.
01 — About
My current focus is geospatial foundation models, machine learning and deep learning applied to Earth observation, with satellite image time series — SAR, optical and multispectral — as the central object of study. The guiding question studied on my master thesis is how learned embeddings transfer across regions, seasons and sensing conditions.
That research sits on top of a genuinely different first career. Years spent producing survey deliverables that had to be checked, defended and signed off built the habit of asking what a dataset's error budget actually is, where the uncertainty comes from, and whether a number can be trusted. Classical geomatics is not a detour from the present work — it is the foundation for it, and it shapes how I read machine learning results: with attention to what the data physically is, how it was acquired, and what a headline accuracy figure conceals.
Before geomatics there was a third field entirely: art history, conservation and restoration of fine arts, including an Erasmus exchange in Rome. That training taught close observation, material analysis, documentation discipline and the patience of slow, reversible work on objects that cannot be replaced. It also produced someone comfortable being a beginner in a new discipline — which turned out to be the transferable skill that mattered most.
The result: I have surveyed in the field, flown drones, processed point clouds under commercial deadlines, and then trained transformers on Sentinel time series and interrogated their embeddings for hidden geographic bias. I can speak credibly to engineers, remote sensing scientists and machine learning researchers, because the work has genuinely been done in all three registers.
Fine arts and conservation, then geospatial engineering, then AI for Earth observation — each a full change of vocabulary, method and community, carried through to a degree or a working role.
Field acquisition and drone piloting, proprietary survey software, scientific Python, deep learning, cloud-native formats, distributed compute — end to end rather than one layer of it.
Commercial survey rewards throughput, tolerances and deliverables that pass inspection. Research rewards ablations, baselines and claims that survive scrutiny. I have worked under both.
Spark, Zarr, PyTorch and RIEGL's proprietary chain share almost nothing with each other, and were picked up across three institutions, one company and one engineering internship in a few years.
A cross-disciplinary programme in Romania put geomorphologists, engineers, climate scientists, psychologists, sociologists and anthropologists around one problem, with a joint output required.
Frozen embeddings, lightweight classifiers on top of them, and — more distinctively — probing what those representations actually encode: domain shift quantification, adversarial adaptation, label-efficiency ablations.
Building, harmonising and modelling multi-date Sentinel-1 and Sentinel-2 series with irregular acquisition, applied to crop-type classification and agricultural monitoring.
Vision transformers, U-Net and CNNs; semantic segmentation, object detection, classification of imagery and point clouds — with train/test protocols that respect spatial and temporal autocorrelation.
SAR (including interferometry), optical and multispectral, hyperspectral, VHR imagery and LiDAR — and what each modality can and cannot physically resolve.
Aerial LiDAR, mobile mapping, terrestrial laser scanning, UAV photogrammetry, terrain modelling, structure monitoring, volumetrics, GNSS, total station and bathymetry — performed commercially, to client tolerance.
A deliberate specialism, not an incidental duty: fair baselines, evaluation splits that do not leak, retained accuracy rather than headline numbers, and unexplained improvements treated as bugs until proven otherwise.
COG, Zarr, STAC catalogues and EO data cubes; Spark and Dask for distributed computation; Parquet for large point-cloud and imagery workloads — including contributing to the modernisation of Copernicus Sentinel data formats.
Hazard dynamics in fragile landscapes, multi-hazard cascades, exposure and social vulnerability, and impact chain modelling — the step that connects an EO measurement to why it is being made.
02 — Current research
MSc thesis · 03/2026 – 09/2026 IRISA — Obelix team, Vannes Université Bretagne Sud + PLUS Salzburg
Supervisors: Charlotte Pelletier and Karima Hadj-Rabah.
Geospatial foundation models are promoted as a way to cut the annotation cost of crop-type mapping, on the assumption that their embeddings capture crop information that stays useful across regions and growing seasons. This thesis takes that assumption seriously enough to test it — asking whether the representations actually separate crop identity from the geographic and environmental context in which the observations were made, or whether the two stay entangled so that a model trained in one country quietly inherits that country's conditions.
Cross-country transfer retained by the embeddings, against 59% for the raw Sentinel time series.
Cross-year transfer retained, against 88% for the raw baseline.
of the labels needed for embedding models to match or beat the raw-data baselines.
from domain-adversarial adaptation — real, but most of the transfer loss remained.
The finding: TESSERA embeddings transfer well across growing seasons but less effectively across countries. Pre-training reduces spatial domain shift without removing it — crop information stays entangled with geographic and environmental factors.
Geospatial foundation models are proposed as a way to reduce the annotation burden of crop-type mapping, assuming that their embeddings capture crop information that remains useful across regions and growing seasons. However, it is still unclear whether these representations separate crop identity from the geographic and environmental context in which they were observed. This thesis evaluates this assumption using TESSERA embeddings, a time series geospatial foundation model, and 13 harmonised EuroCrops classes across 10 European countries for 2018 and across Austria for 2016–2023. Lightweight classifiers were trained on one country or year and evaluated on others, using the raw Sentinel-1 and Sentinel-2 time series as a baseline. Probing experiments tested whether country, coordinates, and acquisition year could be recovered from the embeddings, while domain-adversarial adaptation was used to reduce the recoverable domain information. Cross-country transfer retained 66% of in-distribution accuracy for the embeddings, compared with 59% for the raw series, while cross-year transfer retained 92–94%, compared with 88%. Embedding-based models also matched or exceeded the raw-data baselines using only one quarter of the labels. Adversarial adaptation improved mean target accuracy by 2.8 percentage points, but most of the transfer loss remained. These results show that TESSERA embeddings transfer well across growing seasons but less effectively across countries. Pre-training reduces spatial domain shift but does not remove it, indicating that crop information remains entangled with geographic and environmental factors.
03 — Experience
Research position inside the Obelix team at IRISA (UMR CNRS 6074), a remote sensing and machine learning group whose work centres on satellite image time series, domain adaptation and land cover mapping — the exact intersection the thesis occupies.
Carried out as work in an active laboratory rather than an isolated student project: a full experimental programme on the transferability of geospatial foundation model embeddings, designed, implemented and evaluated over seven months with regular supervision from a specialist in satellite image time series classification and domain adaptation.
An engineering internship in cloud-native Earth observation data infrastructure, working alongside an international open-source team. The work centred on data formats for large EO archives: Cloud-Optimized GeoTIFF and, principally, Zarr — exploring what the new Zarr specification changes about storing and serving satellite data at scale.
This fed into a contribution to ESA's EOPF Toolkit, whose purpose is to modernise and harmonise the data products of the Copernicus Sentinel missions, and closed with a study on flood mapping in Valencia.
xarray.
A full commercial survey engineering role covering acquisition through to delivered product.
Introduction to professional topographic survey: the fundamentals of the survey process, extracting the required features from acquired data, and the comparative characteristics, use cases and limitations of the three principal acquisition families — terrestrial (TLS), mobile (MLS) and airborne (ALS) laser scanning. The internship converted directly into the engineer role.
Held alongside full-time study for the Geospatial Engineering degree — the degree was self-supported.
04 — Projects
Research, engineering and coursework projects. Filter by the kind of work you are looking for.
MSc thesis · Université Bretagne Sud
A full experimental programme testing whether TESSERA embeddings separate crop identity from geography — across ten European countries and eight Austrian years, with raw-time-series baselines, probing, domain-adversarial adaptation and label-budget ablations.
Development Seed · ESA EOPF Toolkit
Work on cloud-native storage of Copernicus Sentinel data, exploring the new Zarr specification's potential against established COG-based access patterns, in service of modernising and harmonising Sentinel data products.
Development Seed
Closing study of the internship: mapping the Valencia floods from a Sentinel-1 backscatter time series served directly out of cloud-native storage, rather than downloaded scene by scene.
Université Bretagne Sud
Classifying the BigEarthNet Sentinel-2 benchmark at scale with a parallel big-data approach, rather than on a single machine.
Université Bretagne Sud
Parallel classification of point-cloud datasets too large for single-machine processing, stored columnar in Parquet and processed with Spark — combining point-cloud domain knowledge from industry with distributed computing.
Université Bretagne Sud
A semantic segmentation pipeline detecting swimming pools from aerial and satellite imagery — a small-object, high-imbalance segmentation problem.
Université Bretagne Sud
Machine learning applied to environmental time series to anticipate drought conditions, involving feature construction from temporal signals and model selection.
PLUS Salzburg
A tool reproducing the processing steps and operators of ESA SNAP programmatically — converting an interactive, click-driven desktop workflow into scripted, reproducible, repeatable code.
PLUS Salzburg · Advanced Remote Sensing
SAR processing through to interferogram generation, covering the coregistration and phase handling the technique requires.
PLUS Salzburg
Semantic classification of LiDAR point clouds in Python notebooks — applying machine learning to data of the same kind previously handled commercially with proprietary survey software.
PLUS Salzburg · Advanced Remote Sensing
Supervised classification and segmentation of EO imagery, including class design, training and accuracy assessment.
PLUS Salzburg · Digital Earth
A Big Earth Data study of burn severity following the 2017 Pedrógão Grande wildfires in Portugal, built on EO data cube concepts rather than single scenes.
PLUS Salzburg · Spatial Simulations
Systems thinking in spatial representations: simulating spatial and temporal dynamics of grazing to study how a complex system responds before the ecosystem is disrupted.
PLUS Salzburg · Methods in Spatial Analysis
Terrain-based spatial analysis — viewshed and visibility modelling over digital elevation data, and the methodological choices that determine whether the result means anything.
PLUS Salzburg · Spatial Databases
Designing and querying spatial databases: schema design, spatial indexing and SQL over geometry, as the storage layer beneath geospatial analysis.
BIP · University of Bucharest
A team-built causal model linking an earthquake hazard in a Romanian village to exposure, structural and social vulnerability, and downstream impacts — developed from field observation and presented to programme lecturers.
No projects in this category.
05 — Community & field
08–10 April 2026 Université Bretagne Sud, Campus Tohannic, Vannes
An international three-day school on artificial intelligence for Earth observation, co-organised by the Obelix team at IRISA with the Erasmus Mundus Copernicus Master in Digital Earth, the SequoIA cluster (PANORAMIX chair) and ESA Φ-lab, supported by Université Bretagne Sud and the European Space Agency. It followed the AI4EO 2025 symposium held in Rennes the previous autumn.
The programme covered foundation models for Earth observation, MLOps, responsible AI and generative models, delivered as lectures by invited scientists alongside hands-on coding workshops and a collaborative data-driven project run with ESA Φ-lab. Attendance was free but competitive, with applications reviewed and places capped; participants could obtain 3 ECTS.
Invited speakers included Pedram Ghamisi and Weikang Yu (HZDR), Adam Stewart (TUM), Stéphane May (CNES), Iris de Gélis (Observatoire de Paris, PSL), Nicolas Audebert (IGN) and Pierre Adorni (IRISA).
Contributing to an event of this kind — international speakers, several partner institutions including a space agency, a selected cohort of postgraduate participants — means working on the delivery side of scientific exchange rather than only the receiving side.
03/2025 – 04/2025 University of Bucharest, Romania Certificate awarded
Five weeks of virtual sessions followed by five days of fieldwork. The virtual component was deliberately cross-disciplinary, with lecturers from climate science, geomorphology, engineering, psychology, sociology and anthropology — covering vulnerability in the Anthropocene, geohazard dynamics in fragile landscapes, sea level change and adaptation, landscape and community resilience, and social vulnerability in coastal environments.
The field component ran in Bucharest and the Vrancea seismic region: a walking survey of structural retrofitting on buildings still carrying the vulnerability exposed by the 1977 earthquake, then a shallow landslide in Subcarpathian molasse, a reservoir site where construction induced slope instability, an earthquake-triggered landslide as a hazard cascade, terrain reshaped by mudflows, a village surveyed on foot to compare structural vulnerability across housing types, and active mud volcanoes — which framed the problem of disaster risk reduction inside protected areas.
The assessed output was an impact chain model, built in a small mixed team: a structured causal representation of an earthquake scenario in Valea Lupului, tracing the hazard through exposure, physical and social vulnerability, to the resulting impacts. That is the analytical step remote sensing work often stops short of.
06 — Skills
Accumulated across three institutions, one survey company and one engineering internship — several of these ecosystems share almost nothing with each other.
Methods: vision transformers · U-Net · CNNs · foundation models and frozen-embedding workflows · semantic segmentation · object detection · time series classification · domain adaptation and adversarial training · dimensionality reduction · probing
Study and work conducted in Portuguese, English and partly in Italian during the Rome exchange, with two academic years spent in French- and German-speaking countries.
07 — Education
Erasmus Mundus Joint Master Degree · full scholarship holder
A two-year European joint programme delivered across two universities, competitively selected and funded by an Erasmus Mundus scholarship, combining Earth observation science with data science and requiring relocation between countries between the two years.
Remote sensing across the full range of modalities — hyperspectral, SAR, optical, multispectral, VHR and LiDAR — alongside EO data cubes, the structure and mission architecture of Copernicus and the Sentinel constellation, big data concepts, spatial databases and scientific programming for EO workflows.
Courses: Advanced Remote Sensing · Methods in Spatial Analysis · Systems Thinking in Spatial Representations · Spatial Databases · Digital Earth: Big Earth Data Concepts
Practical work: supervised classification and segmentation of satellite imagery · LiDAR point cloud classification in Python · SAR processing through to interferogram generation · and, most substantially, a Python API reproducing the processing steps and operators of ESA's SNAP toolbox programmatically.
A machine learning and data engineering specialisation applied to geospatial problems: machine learning, deep learning and computer vision; foundation models and large geospatial AI models; foundations of remote sensing analysis; efficient remote sensing image processing; big data and parallel computing; cloud-based computing.
Project work: drought prediction from environmental time series · parallel classification of large point clouds in Parquet with Spark · distributed BigEarthNet classification · semantic segmentation for swimming pool detection.
Faculdade de Ciências, Universidade de Lisboa
A full engineering degree covering remote sensing, geomatics, programming, satellite systems, GNSS and navigation, GIS and mapping. The practical training spanned both ends of the methodological range: classical instrument survey with total stations and levelling, and modern acquisition including bathymetric survey and UAV-based mapping. Completed while working part-time.
Accademia di Belle Arti di Roma
A semester specialising in conservation and restoration of cultural property, in Italian, within a different pedagogical tradition.
Faculdade de Belas Artes, Universidade de Lisboa
Art history, and the theory and practice of conservation and restoration of fine arts. Training in close visual analysis, material and condition assessment, documentation, and intervention ethics.
Copernicus Master in Digital Earth — competitive European funding awarded to selected candidates for a two-university joint master.
International spring school on AI for Earth observation, Vannes, April 2026 — IRISA/Obelix, Copernicus Master in Digital Earth, Cluster SequoIA and ESA Φ-lab.
European drone licence, held and used professionally for UAV survey missions.
View certificateQuantifying Vulnerability to Natural Hazards in Changing Climate Patterns — University of Bucharest, 2025.
C2 listening and reading, C1 writing and speaking.
View certificateEuropean space sector training badge.
View badgeSite safety certification for industrial and survey environments.
Held and used for field survey mobilisation.
08 — Contact
I am open to work on engineering roles in Earth observation, geospatial machine learning and cloud-native geodata — in Europe or remote.