Skip to content

Uncertainty-Aware Earth Observation Foundation Models

Type of project

M.Sc.

Contact

Spyros Kondylatos (spko@dhigroup.com)

Topic

Remote sensing provides unprecedented opportunities for observing Earth from space. Traditionally, Earth Observation (EO) workflows use satellite data, such as Sentinel-1 and Sentinel-2, to extract actionable insights. Recently, EO Foundation Models have emerged as a powerful alternative to conventional workflows. These models are pre-trained on extensive archives of EO data, encoding substantial domain knowledge within their learned representations [1].

However, the critical nature of EO applications demands trustworthy and reliable deep learning models. Uncertainty estimation can significantly enhance model reliability by providing confidence measures for predictions [2]. Current geospatial Foundation Models primarily focus on delivering large-scale deterministic representations while overlooking their uncertainty. Recent research has introduced the concept of representation uncertainty in EO, incorporating an uncertainty layer into model representations [3]. Nevertheless, this work has been limited to single-modality training with supervised datasets.

This thesis will extend this foundational work to multi-modal EO Foundation Models. The student will leverage state-of-the-art pre-trained multimodal EO Foundation Models and investigate integrating an uncertainty layer into their representations. The research will examine how different combinations of input modalities influence estimated uncertainty and explore how zero-shot uncertainty behaves across diverse downstream tasks.

Relevance to DHI

The DHI Earth Observation Center of Excellence applies machine learning and deep learning methodologies to EO applications. Furthermore, DHI specializes in water-related applications, including bathymetry, coastal monitoring, ocean observations, and related domains. These applications typically represent out-of-distribution scenarios for most EO models, as most training datasets focus on terrestrial environments.

Developing uncertainty-aware Foundation Models can substantially enhance the trustworthiness of EO applications in water-related contexts. By providing quantitative insights into model confidence and prediction reliability, uncertainty estimates enable practitioners to identify when model outputs warrant trust versus when additional validation or expert review is required—a critical capability for high-stakes environmental monitoring and decision-making.

What is the student's contribution

  • Investigate methodologies for integrating uncertainty estimation capabilities into standard EO Foundation Model architectures
  • Develop and evaluate training strategies for uncertainty-aware multi-modal Foundation Models
  • Evaluate downstream task performance across multiple applications, including out-of-distribution evaluation on water-related applications (e.g., analyzing whether uncertainty estimates reflect increased uncertainty in unfamiliar environmental domains)

References

[1] Zhu, X.X., Xiong, Z., Wang, Y. et al. On the foundations of Earth foundation models. Commun Earth Environ 7, 103 (2026). https://doi.org/10.1038/s43247-025-03127-x

[2] Gawlikowski, J., Tassi, C.R.N., Ali, M. et al. A survey of uncertainty in deep neural networks. Artif Intell Rev 56 (Suppl 1), 1513–1589 (2023). https://doi.org/10.1007/s10462-023-10562-9

[3] Spyros Kondylatos, Nikolaos Ioannis Bountos, Dimitrios Michail, Xiao Xiang Zhu, Gustau Camps-Valls, Ioannis Papoutsis; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2025, pp. 6552-6562