On 29 June 2026, Jakob Gawlikowski successfully defended his doctoral dissertation at the Technical University of Munich (TUM). His research focused on improving the trustworthiness of deep learning for Earth Observation through uncertainty quantification, distribution shift detection, and multimodal data fusion.

Deep learning has become a powerful tool for analyzing rapidly growing volumes of Earth Observation data. As these models are increasingly deployed in operational applications, ensuring that their predictions remain reliable under changing environmental conditions has become a critical challenge. Developing AI systems that can recognize uncertainty, detect unfamiliar data, and provide interpretable predictions is therefore essential for trustworthy Earth Observation.

Trustworthiness of Deep Learning for Earth Observation: Uncertainty Quantification and Distribution Shifts

Dr. Gawlikowski's dissertation develops methodological foundations for trustworthy deep learning in remote sensing by addressing uncertainty estimation, out-of-distribution detection, multimodal learning, and model interpretability.

The dissertation makes several key contributions:

Uncertainty quantification for deep learning.
The dissertation provides a comprehensive analysis of uncertainty in deep learning, reviewing uncertainty sources, modeling approaches, and their role in developing reliable AI systems for Earth Observation.

Structured uncertainty modeling for semantic segmentation.
Structured stochastic segmentation networks are introduced to better capture uncertainty and spatial correlations in satellite image segmentation, leading to more informative confidence estimates.

Out-of-distribution detection for remote sensing.
Novel methods are developed to identify distribution shifts in satellite imagery, including a detailed analysis of how cloud contamination can result in overconfident model predictions.

Trustworthy multimodal SAR and optical data fusion.
The dissertation proposes source-wise out-of-distribution detection and adaptive fusion strategies that improve the robustness of multimodal remote sensing models under challenging observation conditions.

Explainable multi-source learning.
An efficient relevance propagation framework is introduced to quantify the contribution of different input data sources, providing improved interpretability of multimodal deep learning models.

Together, these contributions establish a comprehensive framework for improving the reliability, robustness, and interpretability of deep learning models for Earth Observation, supporting their deployment in real-world applications.

Examination Committee

The examination committee included:

• Prof. Xiaoxiang Zhu (Technical University of Munich, PhD supervisor)
• Prof. Katharina Anders (Technical University of Munich), chair of the committee
• Prof. Rudolph Triebel (German Aerospace Center (DLR) and Karlsruhe Institute of Technology (KIT)), examiner
• Prof. Gustau Camps-Valls (University of Valencia), external examiner

The group warmly congratulates Dr. Jakob Gawlikowski on this important academic milestone and looks forward to seeing how his research continues to advance trustworthy artificial intelligence for Earth Observation in the years ahead.