Rare Target and Anomaly Detection in Overhead Imagery using Spatial Feature Vectors

Accurate and timely detection and classification of chemical and/or biological threat agents (CBA) used by an adversary in a battlefield environment is a crucial task for reconnaissance and force protection systems. A key technology component in CBA plume sensing systems is the light detection and ranging (LIDAR) system. The LIDAR laser propagates through and backscatters off of aerosol plume droplets, producing a characteristic return related to the CBA dissemination mechanism. In the operational space of a high-intensity conflict, airborne particulates generated by vehicle movement and weapons effects create a cluttered background that reduces the effectiveness of LIDAR detection. There is a need for advanced automated tools that can separate cases of deliberate dissemination from more common kinetic events ubiquitous to the battlefield environment. Toward this objective, Spectral Sciences, Inc. (SSI) proposes to develop artificial intelligence algorithms built on an artificial intelligence (AI) framework to rapidly classify LIDAR returns as CBA or non-CBA. The proposed solution will utilize the localization of dissemination events to adopt a transfer learning approach to reduce a demand on training data. The Phase I work will prototype the basic approach and demonstrate the feasibility of achieving these goals.