By Emma Wadland
The World Food Programme (WFP) is working with leading experts to take machine learning to new limits.
When disaster strikes, a single drone or satellite image can paint a thousand words but convey little about the individual lives turned upside down or the emergency response they need most.
Until recently, analysing the damage to buildings or infrastructure could take weeks and was carried out later in the response. Now, a machine learning application known as Digital Engine for Emergency Photo-analysis (DEEP) is accelerating the process.
“This is the best part of my job. WFP is not a technology firm. We use tech for people, not profit. This is the most important rule that drives me during this research,” says Marco Codastefano, Data Science Specialist in WFP’s Technology Division.
Even when it’s offline, the application can scour aerial imagery to assess damage to buildings within hours. It improves planning around health, food and shelter, and replaces foot or helicopter assessment, which cuts both risks and costs.
As the first organization to use this machine learning model in emergency situations, WFP deployed it in Mozambique in 2019 and then in Colombia, the Philippines and Lebanon in 2020. But, as Marco explains, this is a rapidly evolving field. “If we want to use technology to save lives and change lives, we need a strong relationship with academia,” he says.
Marco has begun forging research pathways, including with the Polytechnic University of Turin. One of their first actions was to confirm that DEEP could also be applied to satellite images (consider that a good satellite image is 50 pixels per cm compared to 4 or 6 pixels per cm for drone images). This avoids the need to wait for drone equipment and pilots to reach the disaster zone, which is particularly important given pandemic travel restrictions.
“The collaboration between our university and WFP will bring interesting outcomes, both in the field of international cooperation, and locally, in the framework of the activities carried out by Italian first responders,” says Filiberto Chiabrando, Associate Professor of Geomatics at Polytechnic.
However, teaching a model to predict new information is time-consuming. So, Marco and Polytechnic will soon publish their research on training a model like DEEP to identify five damage cases instead of two in the ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences.
“This technology is a necessary step towards better emergency response, not only in our country, but at a global level. It will allow us to gain crucial information for humanitarian frameworks,” says Alessio Calantropio, a researcher at Polytechnic working closely with Marco.
The German Space Agency (DLR) is also helping improve DEEP’s ability to detect buildings and expand it to include roads. This is a chance “to make geo-information from remote sensing data more usable for humanitarian relief missions,” says Nina Merkel, a Research Associate with DLR. “Working with organizations like WFP helps us develop and adapt our methods.”
“DEEP started with damage assessment but could extend to natural hazard predictions and conflict analysis,” says Lara Prades, Head of the Geospatial Unit in WFP’s Emergency Division. That idea inspired her team’s collaboration with the University of Washington in the United States to find new ways of understanding the link between conflict, food insecurity and climate change.
Academics don’t need to be persuaded to help fight hunger. “WFP has an enormous reputation,” Marco says. “Our collaboration starts with the first email.