Automated Toolset for Object Identification and Mapping (ATOM)

Intelligence analysts use a variety of tools to gather actionable information. A common medium of collection is satellite images. However, combing through the captured images for objects of interest (OoI) requires a tremendous investment in analyst-hours. With the implementation of advanced artificial intelligence (AI) and machine learning (ML), some geographic information systems (GIS) and mapping techniques have transitioned into fully-automated or semi-automated systems, dramatically decreasing data processing labor and improving the quality of results. Geospatial Artificial Intelligence, also referred to as GeoAI, is the application of AI and ML to GIS. Disruptive advancements in the field of GeoAI are currently being investigated and developed with the promise of expanding the utility of geospatial information. These advancements are being driven by open availability of geospatial data, advancements in ML methodologies, and increased accessibility of high-performance computing. In partnership with NGA, the Air Force Agency for Modeling and Simulation, as described in the first sub-topic found within the 20.3 Mission Focus Areas document published by AFWERX, is seeking the development of GeoAI methodologies with high potential to deploy in operational environments. As such, Bihrle Applied Research, Inc. proposes the Automated Toolset for Object Identification and Mapping (ATOM). ATOM will provide human analysts with interactive maps and reports that distill enormous amounts of satellite imagery into actionable intelligence. To this end, BAR has identified the following problems that ATOM seeks to research and solve: 1) AI/ML object identification in satellite imagery; 2) Geolocation of identified OoI in satellite imagery; 3) Reporting and mapping techniques that aid human analysts in identifying actionable intelligence. ATOM is designed to generate reports and maps of OoI based on collected data. In Phase I, BAR will utilize the xView dataset for algorithm training, validation, and testing as it contains geolocated satellite imagery that is already labeled with categories of various structures and vehicles. These images are used to train the object identification algorithms. This network is then used by ATOM to process new images from existing data collection processes and generate reports and OoI maps that are sent to analysts to review. Upon review, analysts either approve them or send them back to ATOM in a reinforcement learning feedback loop. If ATOM ever misidentifies an OoI or if analysts identify a new OoI, the analysts will indicate as such on the reports and ATOM will use this information and retrain the objective identification network so that it can continually adapt to ever changing missions. Additionally, analysts will be able to identify specific OoI to track in time and space to generate specific intelligence in time on those objects.