On March 20 2026, Dr. Christoph Schweden successfully defended his doctoral dissertation at the Technical University of Munich (TUM). His research was conducted in collaboration with the German Aerospace Center (DLR), focusing on uncertainty quantification in deep learning for Earth Observation.

With the rapid development of deep learning, remote sensing has undergone a fundamental transformation, enabling large-scale and high-resolution analysis of the Earth system. However, as models become increasingly complex and are applied across diverse geographic regions and data conditions, assessing the reliability of their predictions has become a central challenge. In this context, uncertainty quantification (UQ) plays a key role in moving beyond predictive accuracy towards a more comprehensive understanding of model behavior.

Advancing Uncertainty-Aware Learning in Earth Observation

Dr. Schweden’s dissertation systematically investigates uncertainty in deep learning models for remote sensing, with a particular focus on land use and land cover classification. His work addresses how uncertainty arising from data, models, and human annotation can be explicitly modeled and utilized to improve reliability and efficiency.

The dissertation makes several key contributions:

• Incorporating human label uncertainty.
  By leveraging distributional labels derived from expert annotations, the work demonstrates that annotation variability can be used to improve both model generalization and calibration.
• Label embeddings based on Bayesian reasoning.
  A flexible framework for modeling label uncertainty is introduced, enabling the integration of probabilistic label representations into deep learning models and significantly improving calibration and robustness.
• Distance-aware neural network architectures for out-of-distribution detection.
  Architectural approaches such as spectral-normalized neural Gaussian processes and deterministic uncertainty estimation are adapted for remote sensing, leading to substantial improvements in detecting previously unseen data conditions.
• Uncertainty estimation with conformal prediction.
  The dissertation provides a systematic evaluation of conformal prediction for remote sensing applications, demonstrating its ability to produce statistically valid uncertainty estimates and to support efficient active learning strategies.

Through these contributions, the work advances the state of the art in uncertainty-aware learning for Earth Observation and contributes to the development of more reliable and interpretable geospatial AI systems.

Examination Committee

The examination committee included:

• Prof. Xiaoxiang Zhu (Technical University of Munich, PhD supervisor)
• Prof. Urs Hugentobler (Technical University of Munich), chair of the committee
• Prof. Göran Kauermann (Ludwig-Maximilians-Universität München), co-supervisor
• Prof. Konrad Schindler (ETH Zürich), external examiner

The group warmly congratulates Dr. Christoph Schweden on this important academic milestone and wish him every success in his future research career.