Thermal Satellite Imagery for the Detection and Monitoring of Nuclear Weapons: Early Approaches for Automating Imagery Analysis

Abstract:

Much of what we know today about global events - like Russia's invasion of Ukraine or North Korea’s advancement towards nuclear capability - has come from satellite sensors that collect optical images of the earth's surface. Optical satellite images document the light spectrum visible to humans and reveal objects like facilities and vehicles that help us understand the scope, scale, and pace of events on earth; yet other key clues of activity pertaining to heat have long remained a mystery because no high-resolution thermal imaging satellite has ever existed in the commercial space industry. But just as the British company Satellite Vu launched the world's first commercial high-resolution thermal satellite - the HotSat-1 - in the summer of 2023, Stanford University faculty and students, supported by HAI, embarked on a first-of-its-kind effort to evaluate this new thermal medium and to establish an analytical approach towards implementing AI into satellite imagery analysis. 

AI is a critical-yet-untapped component of space-derived insight. As of 2024, commercial high-resolution satellites image millions of square kilometers of the earth’s surface each day, yet a human analyst can only review about 100 square kilometers’ worth of this imagery. The upshot is a vast delta between what is known, and what is knowable in images collected from space. Though imagery analysis has long been assisted by processes like the automated detection of simple objects like ships, aircraft, and railcars, there is very little actual AI-assisted analysis that can deliver unprompted insight about global security. In other words, visual signatures of nuclear weapons proliferation are so subtle, nuanced, deliberately obfuscated, and varied from facility to facility, there is no extant means to automatically quantify whether such a site is actively engaged in weapons manufacturing or not. But as more and more satellites collect images, and different types of sensors (like optical and heat) collect images of the same facilities simultaneously, we are finally beginning to amass the data and methods needed to build an AI capability for detecting nuclear proliferation. 

In our experiment, imagery analysts worked in a teaching environment with their student-researcher colleagues to leverage thermal satellite imagery against conventional optical imagery analysis and open-source investigative practices to detect nuanced, otherwise-imperceptible clues of nuclear proliferation. This experiment resulted in several discoveries about the nature of the tradecraft, a confirmation that a dormant North Korean nuclear reactor was now likely operational, and increased understanding on maximizing insight yield through AI-enabled satellite imagery analysis.