Intelligent Earth Seminar: Anna Jungbluth (European Space Agency)

Advancing Earth Observation and Climate Research through Self-Supervised Learning: From Land Monitoring to Cloud Modelling

5 March 14:00
Earth Sciences: Lecture Theatre

 

Anna Jungbluth is currently working as a Research Fellow in the Climate and Long-Term Action Division of the European Space Agency (ESA) developing geospatial models for Earth Observation and Climate research. Her research focuses on self-supervised learning, foundation models, and instrument-to-instrument translation to create long-term homogeneous data records of satellite observations. Before joining ESA, she earned her PhD in Physics from the University of Oxford, researching renewable energy and photovoltaics. 

Anna developed an interest in ML for science through an internship with the NASA and ESA-funded Frontier Development Lab research programme, where she worked in an interdisciplinary team of researchers to develop machine learning models for studying the Sun. She returned to the programme in the following years, leading various Heliophysics and Earth Observation projects. Anna is passionate about increasing diversity in STEM and helping to empower the next generation of interdisciplinary scientists.
 

Advancing Earth Observation and Climate Research through Self-Supervised Learning: From Land Monitoring to Cloud Modelling 

Earth Observation (EO) satellites generate petabytes of data annually to monitor environmental changes, from land surface dynamics to atmospheric processes. While raw satellite data is increasingly accessible, expert annotations needed to derive insights are much more sparse, limiting the potential of supervised machine learning approaches. Self-supervised learning (SSL) addresses this challenge by first learning meaningful features from unlabeled data, after which models can be adapted to specific tasks using only a small number of labeled examples.
Our research demonstrates SSL's versatility in two critical domains: land monitoring and cloud modeling. For land monitoring, we develop models that are pre-trained on Sentinel-1 Synthetic Aperture Radar (SAR) or Sentinel-2 optical imagery, and later fine-tuned for tasks like vegetation prediction, biomass estimation, and land cover mapping. We compare different state-of-the-art SSL techniques based on spatial masking and feature matching, and show that SSL strongly reduces the requirements for labeled examples.
For our second application, we extend this approach to 3D cloud modeling - crucial for reducing uncertainties in climate models. We first learn cloud features through SSL on geostationary satellite imagery, and fine-tune our models on limited radar measurements of atmospheric profiles. By explicitly encoding space and time, we create geospatially-aware models that can generate large-scale 3D cloud maps with low regional biases. With this, we demonstrate the potential of SSL to advance Earth System understanding, and maximally leverage the wealth of information in EO data.