Advanced Algorithms for Aerosol Plume Characterization and Differentiation

In a battlefield or other active environment where hundreds or thousands of aerosol plumes may be generated every day, it is a challenge to identify the rare plume resulting from a deliberate dissemination. Machine learning (ML) algorithms can be trained to identify patterns and features based on training data and have the potential to distinguish between intentionally disseminated plumes that may represent a chemical/biological threat and typical environmental plumes or those generated by common movement of events that are prevalent in a battlefield setting. During the Phase II project, S3 developed and demonstrated the utility of a plume anomaly detection algorithm using simulated data, data collected at Dugway Proving Ground (DPG), as well as data collected from an operational port in Beaumont, TX. This anomaly detection system provides the capability to ingest lidar data from any environment, identify the statistical characteristics of typical plumes within that environment, and produce alerts to anomalous plumes in real time. The real-time data feed provides the location and other statistical features associated with the plume of interest along with an anomaly score (i.e., confidence level). This initial anomaly detection system was integrated into the Real-time Eyesafe Visualization Evaluation and Analysis Lidar (REVEAL) real-time processing system and includes an automated process for retraining the algorithm in an operational environment. S3 proposes to build upon the previously demonstrated anomaly detection system by both enhancing and optimizing the demonstrated anomaly detection system, as well as generating real-time estimates of hazardous surface contamination based on the anomalous plume data products. These newly developed capabilities would extend beyond the S3 REVEAL system to any third-party backscatter lidar that produces typical lidar azimuth versus range intensity images. The anomaly detection algorithm will be enhanced by 1) improving ways in which the algorithm differentiates hard targets from persistent plumes, 2) better matching of plume identification from scan to scan, 3) integrating a data validation and quality control methodology to eliminate poor data inputs from affecting the training database, and 4) automatically adapting the algorithm to changes in the lidar scan patterns. Given the dynamic nature of the typical deployment environment, this anomaly detection algorithm requires periodic retraining to continue to identify anomalous plumes. This retraining will be optimized by considering additional trigger criteria beyond the time-based approaches previously demonstrated. Ultimately, this modified algorithm will be deployed into a stand-alone, real-time computational system capable of ingesting any third party backscatter lidar data that conforms to the API developed as part of the proposed project.