GLANCE: Graph Learned Anomalies in Changing Environments

While Wide Area Motion Imagery (WAMI) activity analysis has considerably advanced for continuous tracking—particular through adoption of improved machine learning (ML) methods, less research has been applied to using similar quality image sequences over sporadic intervals to characterize site activity. While pure still-image remote sensing collection can characterize basic change detection using daily imaging, using sporadic sequences or video clips for site activity characterization is a comparably underexplored area. Previous work in this area has focused on modeling and characterizing patterns-of-life of individual tracks in a WAMI scene. Comparably little attention has been given to modeling and characterization of site activity, such as facility usage patterns. Improving ML methods for characterizing site activity within an interpretable model is the focus of our research in this proposal. Quantifying the degree of statistical predictability in activity patterns from these sporadic sequences enables improved prioritization of both analyst and sensor attention. For sensing platforms whose communication bandwidth is less than the overall sensor’s field of regard, this statistical activity characterization can inform re-tasking decisions to maximize overall collection resource utilization without relying solely on strict rule-matching criteria as activity triggers.