when Hurricane Ian swept through Florida in late September, leaving a devastating trail of strong winds and flooding. But a week after the storm passed, some in the three hardest-hit counties saw an unexpected beacon of hope.
Nearly 3,500 residents in Collier, Charlotte and Lee counties received a push notification on their smartphones offering $700 in cash assistance with no issues. A Google algorithm deployed in partnership with the nonprofit GiveDirectly estimated based on satellite imagery that these people lived in severely damaged neighborhoods and needed some help.
GiveDirectly is testing this new approach to emergency aid in partnership with Google.org, the search and advertising company’s charity. The person providing the money is a user of a benefits app called Providers, which manages food stamp payments. Locating messages with the help of Google’s AI software, GiveDirectly only provides assistance to people who live in areas devastated by Ian, faster than manually sorting through the app’s roster.
This is the first time GiveDirectly has used the technology in the United States, but it has tested a similar idea in Togo in the months after the pandemic battered the world economy. There, UC Berkeley researchers provide assistance to families based on signs of poverty detected by image algorithms and clues in cell phone bills.
The Florida project is powered by a mapping tool called Delphi, developed by four Google machine learning experts who worked with GiveDirectly for more than six months starting in late 2019. The software highlights communities in need after disasters such as hurricanes by overlaying real-time maps of storm damage with poverty data from sources such as the U.S. Centers for Disease Control and Prevention. The storm damage data is provided by another Google tool called Skai, which uses machine learning to analyze satellite imagery before and after a disaster and estimate the severity of damage to buildings.
“You now have a map of where is socioeconomically vulnerable and where is broken,” said Alex Diaz, who leads the AI for Social Good team at Google.org. This can help support on-the-ground and expedited assistance. “
The algorithms that underpin Skai’s damage assessment were trained by manually labeling satellite images of hundreds of buildings known to have been damaged in the affected area. The software can then quickly detect damaged buildings throughout the affected area. A research paper on the underlying technology presented at the 2020 Symposium on Artificial Intelligence for Disaster Response claims that automatically generated damage assessments match those of human experts with between 85% and 98% accuracy.
This month in Florida, GiveDirectly sent a $700 push notification to any user of the Providers app with a registered address near Collier, Charlotte and Lee counties where Google’s AI system believes more than 50 percent of buildings have been damaged . So far, 900 people have accepted the offer, and half of them have been paid. If each recipient accepts GiveDirectly’s offer, the organization will pay $2.4 million in direct financial assistance.
