SKAI is a machine learning based tool for performing automatic building damage assessments on aerial imagery of disaster sites.

If you are working in the disaster response space and/or are interested in using SKAI, please reach out to the developers at [email protected].


A humanitarian disaster such as an earthquake, hurricane, or wildfire is a highly disruptive event that affects a region and people in complex ways. Disaster assessment is the process of understanding, quantifying, and locating these harmful effects, in order to provide crisis responders with situation awareness and help them plan rescue and recovery activities.

One predominant type of disaster assessment is identifying all buildings that were damaged or destroyed in a disaster. This helps estimate how much of the population is unsheltered or possibly trapped in rubble, or how much it will cost to rebuild a neighborhood.

SKAI uses machine learning and aerial imagery to automatically identify the locations of damaged buildings in a disaster region. This significantly speeds up damage assessment turn-around times and lowers labor costs. Model-generated assessments match expert generated assessments on between 85% and 98% of buildings assessed in many past disasters ranging from hurricanes to wildfires.

For more information, please refer to our NeurIPS workshop paper and our blog post.


SKAI requires a Linux workstation with a recent version of Python (>=3.7). SKAI was designed to run on Google Cloud, so you need to have a Google Cloud project set up as well.

Please see detailed setup instructions here.

Using SKAI

Please see detailed instructions here.


This software was developed in collaboration with the United Nations World Food Programme (WFP) Innovation Accelerator. The WFP Innovation Accelerator identifies, supports and scales high potential solutions to hunger worldwide. The Innovation Accelerator supports WFP innovators and external start-ups and companies through financial support, access to a network of experts and a global field reach.


View Github