Year
2012
Abstract
In an effort to address the growing concern surrounding illegal nuclear activities around the world, the Oak Ridge National Laboratory has developed a software toolbox aiming at providing analysts with interactive and smart searching capabilities thru aerial and/or geospatial image libraries. A potential safeguards application for the technology resides with the International Atomic Energy Agency (IAEA). The IAEA relies on all-source information acquisitions and broad data fusion and analysis coupled with effective inspection implementation for the detection of undeclared nuclear activities. Geospatial imagery has become an important asset to safeguards—providing spatial, spectral, thermal, and temporal information over extended periods of time. Geospatial imagery is currently the only effective way to conduct non-cooperative localization, identification and monitoring of potential nuclear proliferation-related structures and activities. However, searching for nuclear proliferation activities in a given country requires visualizing thousands of images daily and comparing them with anterior datasets. Aiming at drastically increasing the search efficiency, our toolbox uses an original technique to narrow down the image library to a subset of region-of-interest, to identify the nature of the facility with a good level of confidence, and to display the results using an interactive user interface allowing to refine searches. The keystone of our approach relies on the specific arrangement of buildings that best describes a facility. By translating arrangements of buildings, smokestacks, parking lots, and roads (compiled from ground truth and simulated data) into a nuclear facility taxonomy we can distinguish one facility from another with a higher confidence level. Building upon this concept we have implement a multilevel segmentation algorithm to convert a large scale image down to a list of target compound objects, their associated characteristic features and their interconnections. Using ground truth data these features are compiled to create a semantic labeling framework that can be used to classify new target sub-images. The user interface of our toolbox has been developed around the Microsoft Pivot Viewer to browse quickly through the data. This graphical interface provides a novel way to look at the data and also provide ways to improve the classification results.