Real-time discrimination of battlefield ordnance using remote sensing data

One of the goals of the government-sponsored R&D program is to develop next generation algorithms to discriminate various types of battlefield ordnance in near real-time. Applications that could utilize this capability include early indication and warning of threats, support of battle damage assessment (BDA), level of conflict (LOC) assessment, and intelligence preparation of the battlefield (IPB). As part of this effort, the authors have previously investigated and reported on the performance of several classification algorithms applied to electro-optical data collected by a ground-based sensor. That study included evaluation of the authors' baseline algorithm OSCAR: Ordnance Statistical Classification And Recognition. This paper discusses enhancements made to the algorithm over the last year and evaluates algorithm performance as applied to data obtained from remote assets when remote assets may be ground-, air- or space-based. This remotely collected data has a larger noise component and higher intra-class variances than the ground-collected data, adding new challenges to the discrimination problem. Enhancements that the authors have made to the algorithm this year include 1) feature-based processing, 2) rejection of feature vectors from unknown classes, 3) addition of a confidence level in each classification result, 4) handling of multispectral data, and 5) handling of multiple input file formats. Enhancements the authors have made to the algorithm development workbench include analysis tools for displaying the feature space, the rotated feature space (via Principal Components Analysis (PCA)), and class boundaries/probability contours. These tools help the developer to understand the algorithm performance in insightful ways and help analyze class separability for various features, reveal why specific sample vectors get misclassified, highlight the normality of the data, identify data outliers, etc. Algorithm performance is evaluated for both broad and fine classes. A broad class is defined as a major weapon category such as tank muzzle flash, artillery muzzle flash, or explosion. Fine classes are defined as sub-classes of a broad weapon type, e.g., 105-mm, 155-mm, and 203-mm artillery muzzle flashes. Results are shown and compared for both profile-based and feature-based approaches. These results show that very good performance is obtained with both profile-based and feature-based discrimination versions of OSCAR for the broad classes and promising initial results are obtained for the fine classes.