Master Thesis - Enhancing Regression Performance in Geo-Foundation Models for Earth Observation Applications with Multi-Temporal Learning

Forschungszentrum Jülich, Germany

Updated: 8 days ago

In this thesis you will work in the European Space Agency (ESA)-funded project Fast-EO (Fostering Advances in Foundation Models via Unsupervised and Self-Supervised Learning for Downstream Tasks in Earth Observation).

You will investigate novel approaches to improve the capabilities of Geo-Foundation Models (GeoFMs) for Earth Observation (EO) applications, with a particular focus on downstream regression tasks such as biomass estimation and crop yield prediction. GeoFMs are typically pretrained on large-scale EO datasets using self-supervised learning, enabling them to learn general-purpose representations that transfer across a wide range of applications. However, despite their strong transfer capabilities, GeoFMs sometimes fail to match the performance of task-specific supervised models on downstream regression tasks.

A key challenge lies in effectively modeling the temporal nature of EO data. Since downstream applications require varying temporal resolutions and sequence lengths, from daily observations to monthly or yearly time series, current temporal modeling strategies are often suboptimal and struggle to generalize across diverse tasks. In addition, the limitations of existing pretraining objectives and architectures for regression problems are still not well defined.

In this thesis, you will investigate the factors limiting the performance of current GeoFMs and develop novel methods to improve their temporal representation learning and downstream regression performance. The goal is to design more flexible temporal modeling strategies and improved pretraining approaches that enable GeoFMs to generalize more effectively across a broad range of real-world EO applications.


Similar Positions