SKAI

After an earthquake, hurricane or wildfire, responders need to know which buildings were damaged or destroyed. That count gives a first estimate of how many people have lost shelter or may be trapped, and of what rebuilding will cost.¹ SKAI compares an image of each building from before the disaster with one from after and scores how likely it is to be damaged.^1,3 In many past disasters its results agreed with expert assessments on between 85 and 98 per cent of buildings.¹

Google Research built it with the UN World Food Programme's Innovation Accelerator and the AI for Social Good team at InstaDeep.¹ The method was described in a paper for the NeurIPS 2020 workshop on humanitarian assistance and disaster response, posted in November 2020.⁴ It uses semi-supervised learning: the model learns from a small set of buildings that people have labelled plus a much larger set that nobody has.⁴ The paper tested it on the 2010 Haiti earthquake, the 2017 Santa Rosa wildfire and the 2016 conflict in Syria.⁴ The developers ask disaster response groups who want to use it to write to skai-developers@googlegroups.com.¹

SKAI runs on Google Cloud.² Dataflow cuts the images into one small picture per building, Vertex AI handles labelling and training, and a Cloud Storage bucket holds everything.^2,3 Building locations come from Google's Open Buildings, from OpenStreetMap, or from your own CSV of coordinates; the two open sources differ a lot in coverage from region to region.³ The images must be Cloud Optimized GeoTIFFs.³ People then label a sample of buildings (2,000 in the example command) in a tool called Eagle Eye as undamaged, minor, major or destroyed, and those grades are folded into damaged or not.³ The trained model scores every building in the area and writes the results to a CSV.³

For a model of a place, the output is a point layer of buildings with a damage score after a specific event. It can feed a count of homes lost in each neighbourhood. The code is under the Apache 2.0 licence.¹ Microsoft's building damage assessment toolkit does similar work by fine-tuning an image model on damage drawn by hand in the affected area.⁵