Commercial and government organizations are now collecting 100 terabytes or more of overhead imagery via satellites and drones on a daily basis. Specifically, within the Department of Defense (DoD), analysts spend enormous amounts of time sifting through these data to detect events of interest, categorize them, and report them through the appropriate channels. Analysts are well-trained in their ability to sift through data, however, the amount of analysts available to perform this work is limited. The volume of the data is increasing rapidly, and as this increase continues, it will be more difficult for analysts to find the bandwidth to support this activity. As a result of this limited bandwidth and lack of equivalently increasing number of personnel, the DoD community will have to prioritize which data to analyze. This will inevitably lead to unintentionally missing significant events. As this trend continues, how can the DoD improve this process and alleviate the workload? One solution is to employ machine learning; more specifically, deep learning to the imagery to perform change detection as well as object detection and localization such that further methods for drawing higher-level insights on Patterns of Life can be enabled. Employing these methods could ensure that all of the data gathered could be analyzed without requiring more analysts. The paper will describe the background associated with Patterns of Life analysis and discuss in detail the development of a modified U-Net architecture adapted to do change detection for military applications with overhead imagery. The described architecture will show the DoD community a viable approach for dealing with the problem of scalability that comes with collecting increasing amounts of data. Additionally, the paper will explain how the method was tested with publicly accessible satellite imagery datasets, and finally, describe conclusions about the work in terms of warfighter applicability.
Keywords
AI,DEEP LEARNING,MACHINE LEARNING,PATTERN OF LIFE
Additional Keywords