A lack of on-the-ground information at the start of a humanitarian crisis is a major obstacle to a quick and effective response. It is critical to know the location and the best way/s to access affected populations. This is particularly difficult in locations where infrastructure and communications networks may be damaged or disrupted.
Project overview
SKAI (formerly RUDA) uses Artificial Intelligence (AI) and satellites to help WFP reduce the amount of time needed to understand the impact of disasters.
By using Artificial Intelligence (AI) to analyze images provided by satellites, WFP can respond to emergencies without the delays and logistics obstacles associated with a traditional response. SKAI dramatically speeds up the process of determining what has happened, what to expect on the ground, and helps define the kind of support that is required.
When a tropical cyclone hits a low-lying coastal region — a quick assessment of everything from structural damage to transportation networks, in addition to basic information on the lay of the land, is paramount when preparing for action.
Gathering this information by traditional means, such as sending in a human team, might take weeks. But it doesn’t have to be this way— and SKAI is leading the change.
With the appropriate use of satellite imagery and image analysis powered by AI—humanitarian disaster response has the potential to be both faster and better aligned with the needs of affected populations.
Satellites are already widely used to provide broad, regional coverage of phenomena such as climate-related crop failures, post-storm housing damage, and the movement of people during political conflicts. Drones, whether remotely controlled or fully-autonomous, are increasingly being used to fill in the gaps at a more local level, tracking changes that might happen over the course of hours or days.
The big hurdle is that all this visual data needs to be interpreted. Today, for instance, imagery collected from a 20-minute drone flight takes a highly trained, remote-sensing analyst half a day’s work to analyze. The monotony of the workload can introduce inconsistency, fatigue-based and other human errors, and limited opportunities for internal audits and improved accountability.
Results - 2022
Multiple WFP teams implemented SKAI during the assessment of several major disasters:
> In April, severe flooding and landslides in South Africa caused the death of 448 people and displaced over 40,000 people.
> In June, floods in Pakistan affected at least 33 million people. Across both locations, SKAI assessed 1,100,449 buildings for flooding damage.
> Following Hurricane Ian in late 2022, Google Research AI, worked with GiveDirectly in the USA to provide USD 2 million in cash assistance to more than 2,900 low-income households in heavily affected areas identified by the SKAI algorithm.
SKAI improved the speed of WFP's damage assessment by 92 percent, while achieving 77 percent in cost savings, surpassing expectations.
Working with a robust training and evaluation set is the foundation of a machine learning model able to recognize such things as damaged buildings and infrastructure. Such an extensible platform may eventually integrate additional assessments, such as movement of people, crop failures, and other variables.
The ambition for SKAI is to build an AI-powered platform able to function across a variety of geographic locations, lighting conditions, disasters, and damage types. SKAI is unique in having been trained with 27 past-onset disasters that were previously tagged by manual analysts.
A working group formed by UN agencies and private collaborators have been working to turn SKAI into a software product for the broader humanitarian sector, with multiple use cases. Its machine learning source code has also been made open-source on the Github repository.
