Generalization to unseen disasters and regions

Establish reliable generalization of building damage assessment methods to disasters and geographic regions not represented in the training data.

Background

The chapter reviews building damage assessment from the perspectives of data, evaluation, and methods. It emphasizes that existing datasets and benchmarks often have limited geographic diversity or evaluate training and test samples from the same region, which can produce overly optimistic performance estimates.

The authors identify transfer to unseen events and regions as a persistent unresolved difficulty, particularly for methods that must operate under changes in building typology, urban layout, sensor characteristics, and disaster conditions.