Geography Aided Inference ATR (GAIA)

The amount of data collected from the suite of current and future sensors far surpasses the bandwidth of analysts to processes the data streams into actionable intelligence. This pixel to pupil ratio problem is a forcing function for developing robust algorithms which accurately find non-cooperative objects while minimizing the false alarm rate.  As exploitation algorithms are tasked with performing wide area searches, the ability to accurately train on relevant regions of clutter is limited, and techniques to reduce false alarms are paramount.   These false alarms can be generated by both manmade (buildings, trailers, towers, windmills, dams, refineries, bridges, roads, fences, powerlines), and natural objects (tree clusters, riverbanks, cliffs, ponds).  Even though the AI/ML workflows for SAR detection/classification are currently leveraging state of art advancements to provide useful analytics to the community, we are still hindered by blindly treating each cluster of pixels as a potential target – without including the geographic context.  In this proposal, we will present a framework for augmenting detection likelihood based on proximity to meta data (from Open Street Maps and elevation contours/pointclouds from United States Geological Survey data).  Under this effort we will enhance  ATLAS/GLOBE ML workflow to include geography metadata. As a complete ML system, the geography aided inference ATR (GAIA) will enhance the ability of both commercial and IC partners to exploit sensor data in open world environments.