On 6 March 2026, Dr. Fahong Zhang successfully defended his doctoral dissertation at the Technical University of Munich (TUM). His research was conducted within the framework of the Global Building Atlas project, focusing on large-scale semantic understanding of remote sensing imagery.

With the rapid growth of Earth Observation (EO) data, the ability to systematically extract geospatial information from satellite imagery has become increasingly important. While deep learning methods have significantly advanced remote sensing analysis, several key challenges remain, including domain shifts across geographic regions, limited and incomplete annotations, and the conversion of raster predictions into structured geospatial representations.

Dr. Zhang’s dissertation addresses these challenges by developing methods for instance-level semantic understanding in remote sensing imagery, with a particular focus on scalable mapping of buildings from satellite data.

Key Research Contributions

Dr. Zhang’s dissertation introduces several methodological innovations in geospatial AI and remote sensing:

• Domain Adaptive Semantic Segmentation
A pseudo feature-guided self-training framework was developed to improve the cross-regional generalization of segmentation models.

• Robust Landslide Detection Across Regions
An uncertainty-aware self-training strategy enhances model transferability and achieved 3rd place in the Landslide4Sense competition.

• Few-Shot Object Detection with Incomplete Labels
A self-training detection framework enables effective learning from partially annotated remote sensing datasets.

• Global Collinearity-aware Polygonizer
A transformer-based method converts building masks into geometrically consistent vector polygons, enabling accurate large-scale mapping.

• Global Polygonal Building Mapping Framework
These methodological advances are integrated into a system capable of generating building polygon maps across six continents, demonstrating strong performance in real-world large-scale applications.

The polygonization framework developed in the dissertation has been directly used to generate the GBA.Polygon attribute of the Global Building Atlas dataset, highlighting the practical impact of this research.

Examination Committee

The dissertation was evaluated by an international examination committee:

• Muhammad Shahzad (Chair, now with University of Reading)
• Xiaoxiang Zhu (PhD Supervisor)
• Friedrich Fraundorfer (external examiner, Technische Universität Graz)
• Sébastien Lefevre (external examiner, University of South Brittany)

The group warmly congratulates Dr. Fahong Zhang on this important academic milestone and wishes him continued success in his future career.