A typhoon strikes an island in the Pacific Ocean, downing power lines and cell towers. An earthquake hits a remote mountainous region, destroying structures and leaving no communication infrastructure behind. Communication gaps created in the wake of a natural disaster or disease outbreak can create a seemingly impenetrable fog between first responders and affected citizens, leading to confusion, mismanagement of resources, and further loss of human life.
Thanks to the “eyes in the sky” and the latest advancements in artificial intelligence, emergency responders and health workers will be able to assess the scale and impact of natural disasters and humanitarian support needed anywhere in the world, no matter how remote.
On the opposite side of the globe, computer scientists like Dalton Lunga at the Department of Energy’s Oak Ridge National Laboratory can process large volumes of satellite imagery of an impacted area using advanced machine learning techniques to extract structural information of any region and generate maps that are critical for understanding population distribution and natural disaster assessment.
Traditionally, the impact of natural disasters was heavily dependent on first responders scouting affected areas and working with community members to determine where humanitarian support was needed the most.
“Although it makes an impact, at a large scale, that approach is not sustainable and often not all key affected areas receive resources efficiently,” Lunga said. “What if some affected areas are not even accessible? How can we, as computer scientists working with domain experts, build systems that are less human-dependent for humanitarian support missions?”
Such systems would be key in delivering satellite imagery-based data products in the form of maps to support local humanitarian and emergency response agencies, all before the first responders even arrive on the scene. The result is a more streamlined, data-driven recovery effort, with the greatest aid going to those with the greatest need.
Lunga, a lead scientist on multiple projects at ORNL, conducts research on new machine learning driven geospatial image analytics in the lab’s Geographic Data Science Group and is a member of the ORNL AI Initiative. His work utilizes massive geospatial datasets and the world’s most powerful supercomputers to solve problems not only in disaster response, but also in national security, disease eradication and environmental monitoring.
Lunga’s work at the lab allows him to apply lessons from foundational AI methods to solve problems of greater impact to society. This combination of learning and problem solving was instilled in him from an early age and carried through his education in electrical and mechanical engineering in South Africa.
“I’ve always been a hands-on person, someone who was always driven by wanting to know why things work the way they do and what I could change for something to work differently,” Lunga said. “Growing up, I was always fascinated by many things, like flying machines and how it was possible to ‘put’ a human voice inside a radio.”
Lunga pursued his doctoral degree in computer and electrical engineering at Purdue University and gained a deeper comprehension of machine learning and signal processing while working on a project to understand remote sensing images for supporting agricultural mapping applications. Lunga was drawn into remote sensing because of the capability to exploit a large amount of information with the potential to make an impact at a global scale.
“The vision of one day being able to develop large scale data products based solely on imagery and other sources of information became an idea that stayed with me through the end of my Ph.D. program,” Lunga said.
After Purdue, Lunga moved back to South Africa, where he led efforts to develop machine learning techniques for natural language processing. However, he always wished he could do more applied research connected to his newly developed capabilities in machine learning and signal processing in an environment that offered more challenges and computing support. Through his connections at Purdue, an open post-doctoral appointment at ORNL landed in his inbox, and he was encouraged to apply at the recommendation of his advisor.
“There was little hesitance on my part to consider the opportunity, as I had to weigh on leaving a senior research scientist role for a postdoc,” he said. “I thought about it and saw it as a bridge to get re-engaged with the things that mattered more, where I could play a bigger role as part of a larger community.”
ORNL has three resources that make it stand out from other institutions, Lunga said. One is the support from management and networks with a larger community of stakeholders. The second resource is the world-class computing and expertise that enable researchers to tackle the big questions.
“Whereas other scientists limited by computing resources, I am limited by the imagination of the solutions I can try on a problem,” he said.
The third is more intangible, Lunga said, and comes from the community of scientists both in and out of his own field that are open to collaboration and supporting one another.
“For example, as a member of the ORNL AI initiative, I am able to engage with others in an interdisciplinary environment and share ideas or dissect a problem. This is a very unique position that I associate with being here at ORNL,” he said. “By collaborating with other scientists, we can work toward building AI tools that we can leverage for various problems.”
Lunga envisions both short-term and long-term impacts of his work. In the immediate future, he said, there will be a growing adaptation of AI into humanitarian workflows to enable decision making, with personal computing devices enabling those impacted by natural disasters to play an active role and interact with these systems.
This society-wide impact can be seen most prominently in the work he has done with the Bill and Melinda Gates Foundation to support their disease eradication efforts in sub-Saharan Africa. He has worked on a project to find and map human settlements that may not have been accounted for during local census surveys. In Nigeria, he said, census information was over a decade old, so the growth and movement of people across the country had not been properly documented, and another census could take multiple years to complete. Using recent satellite imagery and infrastructure mapping from AI, though, Lunga and his colleagues were able to map the entire country and provide a population distribution estimate in a few weeks compared to the years it would have taken the national government.
“We often refer to these unaccounted people, these lives, as ‘the missing pixels,'” he said. “We are putting those missing pixels back on the map while at the same time supporting post-disaster recovery efforts, opening up economic development and stretching social services into areas that may not have been possible without AI and global imagery.
“The bigger picture is for every person in affected areas to be connected to humanitarian support systems and to be accounted for as part of one society.”
In the longer timeframe, Lunga envisions GeoAI systems that will process planet-scale datasets with high precision and embedded intelligence capabilities to monitor and focus on areas as they are triggered by certain events. The insights generated can then be transmitted in near-time to stakeholders for actionable further steps, he said. As such, his work would evolve more towards integrated systems, where AI will be just one component of a bigger GeoAI system that can process quadrillions of pixels of data using supercomputers and other computing devices.
These future AI systems will be less experience-based and modeled more on how humans reason and make wise decisions in new situations with limited experience. With these new systems, humanitarian agencies will be able to access the information they require without the need for a world-class supercomputer.
“What if you are in some part of the planet, like in India or Africa, where the only access you have to technology is your mobile phone?” Lunga said. “You could not even think of running our current AI systems on those devices, so we must come up with less complicated methods that can be trained efficiently from less experience yet are still highly performing for all.”