KISSIMMEE, Fla.—Continuously improving artificial intelligence around computer vision, analytic workflows, data collection, and enterprise infrastructure, all wrapped into responsible uses of the technology, make up the high-level goals for the National Geospatial-Intelligence Agency (NGA) around its application of AI, agency head Vice Adm. Frank Whitworth said last week.
NGA manages various AI programs and efforts, most notably its Maven program based on computer vision AI used to automatically detect and identify potential targets of interest in imagery and video. Maven is a program within NGA AI, which is supported by Whitworth’s five goals around AI.
Computer vision (CV) must keep improving around positive identification, geolocation, and speed, which necessitates improvements around “existing overhead imagery broad area search, targeting, and full motion video lines of effort, Whitworth said of his first goal during a keynote address at the GEOINT 2024 Symposium.
“We need to deliver target detections from new modalities, deliver CV models to military service partners on autonomous systems, and automate geolocation accuracy of detections,” he said in his May 6 address.
Better computer vision models are “useful for analysts and for operators,” Mark Munsell, director of the Data and Innovation Directorate at NGA, told Defense Daily
on May 10 during a follow-up interview to GEOINT.
Whitworth said the next goal is more integration of “AI into the analytical workforce,” to include taking advantage of generative AI for searching and reasoning, improving models through analyst feedback, and using synthetic learning and labeling.
The third goal is to make AI part of the routine “collection orchestration” for tasking satellites, and accounts for priorities, limitations, current and crisis needs, all occurring “with very little fanfare,” he said.
Implementing an enterprise AI infrastructure that lowers costs and increases turnaround times for imagery services, services, data storage, and data access is the fourth goal.
“It’s going to need a common platform with data management for labels, models, and detections, and then to provide labels, models, and detections as a service to our customers,” Whitworth said.
NGA also wants to lead the with the proper use of AI and is creating a new training program for the responsible use of the technology by developers and users, he said of the fifth goal.
Achieving these goals is necessary “to produce GEOINT data and AI models of increasing accuracy, that generate high quality and useful detections and reports to be applied to defense and intelligence missions,” Whitworth said.
In addition to NGA Maven, the agency manages another AI program of record called ASPEN. Maven is focused on the combatant commands and helping to find and identify targets whereas ASPEN aids in analysis, Munsell said in the interview.
“NGA ASPEN is focused on helping analysts do their jobs to include warning, and writing reports, and reporting,” Munsell said.
NGA also manages other AI efforts that help with production and analysis, and that essentially get after Whitworth’s goals, he said.
Some of these efforts include natural language processing the agency is using for its maritime and aviation safety navigation mission, machine learning to build gravity models of the Earth, and this summer NGA will use large language models to help analysts with their work, he said.
Munsell has said previously that AI technology continues to make remarkable advances. Coming out of the GEOINT conference, a key takeaway is that the large models developed and trained for language and vision applications have advanced to where in “the next big phase” the geospatial-intelligence community needs to apply and “fine-tune them in our data…and in our production work,” he said. “I think that’s going to be a big improvement over the next couple of years.”
Munsell mentioned OpenAI’s GPT-4V, which stands for generative pre-trained transformer for vision, which has been trained on a vast array of sources to detect things in photos and videos.
Computer vision keeps getting better but the large models “are the next big thing,” he said.