Transformers for building damage assessment

DAHiTrA is a novel deep-learning model employing hierarchical transformers to classify building damage. Related to approaches used in LLMs, this approach encodes spatial features at multiple resolutions, comparing changes from imagery collected before and after a disaster. Trained on a new high-resolution satellite imagery dataset collected around Hurricane Ida in 2021, the model has proven to be easily adapted to new regions with limited additional fine-tuning. Traditional post-disaster damage assessments can take months, but DAHiTrA, with satellite images available within 24 hours, provides damage reports the next day. This speed and accuracy improve resource allocation for communities and governments. DAHiTrA is used by government agencies, emergency management, and international organizations for effective disaster management.

As new AI approaches are integrated into existing workflows for disaster response they will enable faster and more accurate damage assessment and resource allocation and distribution, reducing impacts on affected regions and populations.