Inferring the Operational Status of Nuclear Facilities with Convolutional Neural Networks to Support International Safeguards Verification

摘要

International nuclear safeguards analysts use images in myriadways to support verification analysis tasks, from analyzing the designand construction of a facility to understanding the scope andcapacity of work performed therein. Potentially relevant groundbasedimagery from open sources has increased significantlyin the past 10 years as individual users with smart phones havebecome “citizen sensors,” posting geolocated content to socialmedia platforms in near real-time. While this is an exciting newsource of data for analysts, it is impractical to review unaided. Theauthors use machine learning to make image search and prioritizationmore efficient for safeguards analysts in three potentialworkflows. In this paper, the authors demonstrate the successfuluse of cooling towers and steam plumes as a signature that canindicate a facility’s operational status and describe a convolutionalneural network modeling approach that yields over 90 percentaccuracy for identification of cooling towers and steam plumesfrom open source ground-based images.