His research focuses on physics-consistent machine learning for natural hazard monitoring and process-based forecasting, addressing the need for reliable and scalable approaches in a changing climate.

With the increasing frequency and intensity of natural hazards, Earth Observation combined with machine learning has become a key tool for monitoring and prediction. However, purely data-driven models often struggle with generalization and physical consistency, while traditional physics-based approaches can be computationally expensive and difficult to scale. In this context, integrating physical knowledge into machine learning models offers a promising direction toward more robust and trustworthy forecasting systems.

Advancing Physics-Consistent Learning for Natural Hazard Analysis

Dr. Xu’s dissertation systematically investigates how physical principles can be incorporated into machine learning frameworks for Earth Observation, with a particular focus on hazard detection and flood forecasting. His work addresses challenges related to data scarcity, transferability, and model reliability across diverse environmental conditions.

The dissertation makes several key contributions:

Unsupervised and transferable hazard monitoring.
The work develops methods for detecting natural hazards without relying on labeled data, enabling robust performance across unseen regions and scenarios.

Physics-informed flood forecasting.
Hydrodynamic principles are integrated into learning-based models to enable accurate and scalable flood prediction using multi-source satellite observations.

Benchmark datasets for evaluation.
Large-scale, process-aware datasets are established to support systematic benchmarking and comparison of machine learning approaches for flood forecasting.

Neural solvers for spatiotemporal dynamics.
New modeling approaches are proposed for efficiently capturing complex flood dynamics, enabling scalable and resolution-independent predictions.

Towards reliable and interpretable machine learning.
The dissertation explores physics-aware and uncertainty-aware learning strategies to improve model consistency, generalization, and usability in real-world applications.

Beyond methodological contributions, the work demonstrates strong practical relevance, with developed approaches attracting interest for real-world deployment and integration into applied platforms.

Examination Committee

• Prof. Xiaoxiang Zhu (Technical University of Munich, PhD supervisor)
• Prof. Niklas Boers (Technical University of Munich), chair of the committee
• Prof. Jonathan Bamber (University of Bristol; TUM Institute for Advanced Study)
• Prof. Jian Peng (Helmholtz Centre for Environmental Research)

The group warmly congratulates Dr. Qingsong Xu on this important academic milestone and wishes him every success in his future research career.