On 30 June, Chenying Liu successfully defended her doctoral dissertation, "Representation Learning with Weak Labels in Remote Sensing," at the Technical University of Munich (TUM). Her research was conducted within the Munich Center for Machine Learning (MCML) and focused on developing representation learning methods that effectively utilize weak supervision for Earth Observation.

Recent advances in artificial intelligence have significantly improved the analysis of Earth Observation data. However, large-scale, high-quality annotations remain expensive and difficult to obtain. At the same time, weak supervision from sources such as OpenStreetMap, existing land cover products, and other geospatial datasets has become increasingly available. Effectively exploiting these imperfect yet abundant semantic resources is therefore an important step toward scalable and transferable AI for remote sensing.

Representation Learning with Weak Labels in Remote Sensing

Dr. Liu's dissertation investigates how different forms of weak supervision can be incorporated into representation learning to improve remote sensing models under limited annotation availability.

The dissertation makes several key contributions:

Learning from incomplete annotations.
Adaptive learning strategies are developed to recover missing semantic information, enabling more effective semantic segmentation and multi-label classification from incomplete annotations.

Weak-label pretraining for representation learning.
The dissertation demonstrates that large-scale noisy semantic labels can effectively support representation learning through cross-modal pretraining, improving the scalability of remote sensing foundation models.

Foundation models for Earth Observation.
The proposed LandSegmenter framework integrates heterogeneous semantic information from multiple weak supervision sources to learn more transferable representations for land use and land cover mapping.

Uncertainty-aware weak supervision.
The work extends weakly supervised learning by explicitly modeling label uncertainty, providing a data-centric perspective that improves the reliability and robustness of learned representations.

Together, these contributions demonstrate that weak labels are not merely imperfect annotations but valuable sources of semantic information for scalable representation learning. The dissertation establishes a systematic framework for leveraging heterogeneous weak supervision to develop more transferable and reliable AI models for Earth Observation.

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. Chenying Liu on this important academic milestone and looks forward to seeing how her research continues to advance representation learning, weak supervision, and geospatial AI in the years ahead.