On 30 June, Wei Huang successfully defended his doctoral dissertation at the Technical University of Munich (TUM). His research focused on developing label-efficient learning methods for remote sensing under limited annotations and distribution shifts.

Artificial intelligence has become a key technology for extracting information from the rapidly growing volume of Earth Observation data. However, obtaining high-quality annotated satellite imagery remains one of the major bottlenecks for developing robust and transferable models. As Earth Observation datasets continue to expand in scale and diversity, learning effectively from limited labeled data has become a fundamental challenge for remote sensing and geospatial AI.

Label-efficient Learning for Earth Observation

Dr. Huang's dissertation develops a systematic framework for label-efficient learning in remote sensing, addressing semi-supervised learning, domain generalization, and semi-supervised domain adaptation.

The dissertation makes several key contributions:

Adaptive and debiased semi-supervised learning.
Novel learning strategies are proposed to improve pseudo-label quality while mitigating class imbalance and confirmation bias, enabling more reliable learning from limited annotated data.

Label-efficient building footprint extraction.
The dissertation develops methods that combine adaptive pseudo-labeling, debiasing strategies, and building height information to substantially improve building extraction under extremely limited annotation budgets.

Robust semantic segmentation.
Decoupled and class-balanced learning frameworks are introduced to address noisy pseudo-labels and long-tailed class distributions, improving segmentation performance across diverse remote sensing datasets.

Representation learning for domain generalization.
The work proposes representation enhancement and stabilization strategies that improve the generalization of deep learning models across different geographic regions and imaging conditions.

Semi-supervised domain adaptation.
A bidirectional alignment framework is developed to effectively transfer knowledge across domains using only a small number of labeled target samples.

Together, these contributions establish a comprehensive framework for label-efficient learning in remote sensing, advancing the development of robust and transferable AI models for Earth Observation under limited annotation availability and distribution shifts.

Examination Committee

The examination committee included:

• Prof. Xiaoxiang Zhu (Technical University of Munich, PhD supervisor)
• Prof. Benjamin Busam (Technical University of Munich), chair of the committee
• Prof. Lorenzo Bruzzone (Università di Trento), examiner
• Prof. Richard Bamler (Technical University of Munich), examiner

The group warmly congratulates Dr. Wei Huang on this important academic milestone and looks forward to seeing how his research continues to advance label-efficient learning, remote sensing, and geospatial AI in the years ahead.