Congratulations to our EAGLE student Christina Krause, who successfully defended her MSc thesis titled “Active Learning and Parameter-Efficient Feature Adaptation for Cross-City Seismic Building Type Transfer Using DINOv2 and Street-Level Imagery” on November 12, 2026.

Christina’s thesis explored whether street-level imagery offers a scalable source for characterizing buildings in the thematic context of seismic risk assessment. Models trained in one city lose much of their accuracy when applied to another, and labelling data for every new city is costly. Christina investigated in her master thesis how a vision foundation model can close this gap. Using Santiago de Chile as source and Lima as target city, which differ in building stock as well as in imaging platform, two levers for cross-city transfer, parameter-efficient adaptation with low-rank adaptation (LoRA) on the source city and active learning (AL) of target-city labels, starting from frozen DINOv2 features, were systematically compared. The results revealed that the most effective lever depends on the attribute. For building height, source-side adaptation achieves a performance gain that frozen-feature annotation matches only after approximately 1,750–2,500 target labels, depending on the acquisition strategy. For construction material, by contrast, source-side adaptation alone is insufficient, and only target labels improve performance substantially. Without target labels, the LoRA-adapted model reaches a weighted F1 of 0.51 for the SBST on the classes shared by both cities, compared with 0.42 for the Reference CNN. Combining source-side adaptation with target-side annotation yields the highest performance for all three attributes and reaches 0.71 on the shared classes for the SBST attribute. The choice of acquisition function, in contrast, proves secondary to the learning regime in this setup. The thesis further indicates that building classes absent from the source label space resist detection because they are new in the label space but hardly distinct in the feature space. For transferring exposure models to data-scarce cities, the findings support adapting a pretrained foundation model on the labelled source city and continuing to train it on a few hundred actively selected target images, while building types absent from the source city remain the main gap.

The work was supervised by Prof. Hannes Taubenböck, Prof. Christian Geiß and Dr Patrick Aravena-Pelizari (DLR).  

Congratulations, Christina, on the successful defense, on the impressive presentation of your great piece of research work!