SBIR-STTR Award

Elastic backscatter lidar has proven to be a powerful tool for high-sensitivity long-range detection and tracking of airborne aerosol plumes and in its application to the characterization of chemical/biological (CB) aerosol agent release events. Advanced sensor platforms and networks are now integrating lidar systems in order to improve situational awareness and increase early warning in the event of a CB attack. These system-of-systems architectures are part of an overarching strategy known as Integrated Early Warning (IEW). Lidar is playing a key role in IEW systems by providing a standoff detection layer that has the potential to provide the greatest impact on improving situational awareness and increasing early warning. Despite the sensitivity of elastic backscatter lidar to CB pathogens in the respirable range, it has proven difficult to discriminate threat plumes from variations in the aerosol background caused by kinetic events, vehicle movement and other conflict-related actions. A backscatter lidar that could differentiate between these ubiquitous background plumes and deliberately disseminated threat plumes would significantly enhance the capability of IEW systems to provide accurate and timely threat information to battlefield commanders. The Chemical, Biological, Radiological, and Nuclear Directorate of the Pentagon Force Protection Agency (PFPA-CBRN) deployed a continuously operating elastic backscatter lidar that monitored the urban atmosphere around the Pentagon. Statistical features (size, shape, location, intensity and duration) were calculated for naturally occurring plumes and compared to plumes generated from nearby simulant disseminations. This early work suggested that simulant plumes displayed spatial and temporal variations that could be exploited to discriminate them from typical background plumes. Since this initial PFPA program, a significant amount of research has gone into machine learning algorithms for a variety of tasks including image classification of remote sensing data. This work has demonstrated the potential and efficacy of active learning for challenging remote sensing data processing scenarios. Continued development of graphical processing units (GPU), along with increasingly larger datasets, has led to the resurgence of 2-D convolutional neural net (CNN) implementations for a variety of image classification and machine learning applications using deep learning. These machine learning implementations are now surpassing human-level performance for some image classification tasks. The deployment of elastic backscatter lidar systems will continue to expand the amount of data presently available for aerosol plumes. Using these larger datasets, coupled with recently developed machine learning algorithms, provides an opportunity to utilize the differences in spatial and temporal plume features to discriminate between potential threats and more benign environmental plumes.

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. S3 has previously shown that a Support Vector Machine (SVM) supervised learning ML model can distinguish disseminated from non-disseminated plumes at a high rate of accuracy in a stable, low-clutter environment such as that typically found at Dugway Proving Ground (DPG). However, the SVM requires data from plumes of both types in its training dataset, and in a real-world, high-clutter environment, disseminated plumes will generally be unavailable for training the SVM. An alternative approach using an unsupervised learning anomaly detection algorithm has demonstrated promising results for both simulated and real lidar data with accurate identifications of disseminated plumes more than 90% at false positive rates <2%. In Phase I of this project, S3 teamed with Aeris to utilize their Joint Outdoor-indoor Urban Large-Eddy Simulation (JOULES) model for generating synthetic plume data from a variety of simulated sources in a realistic atmosphere. Algorithms were benchmarked using the simulated data before down-selecting for evaluation with real lidar data collected at DPG by the West Desert Lidar. S3 proposes to continue working with Aeris to generate higher fidelity simulated plume data in more complex scenarios to further evaluate the anomaly detection algorithm. Following optimization with simulated data, the algorithm will be tested with lidar data collected continuously for multiple weeks outside the S3 Albuquerque facility. This collection will allow for a range of conditions and provide data for a variety of plumes common to an urban environment. This low-cost approach will produce valuable test data for algorithm refinement and aid in the development of a real-time autonomous and adaptable training methodology. Autonomous retraining based on observed conditions and data characteristics is critical to deploying a battlefield capability rather than a post-processing detection capability. Finally, S3 will integrate this real-time anomalous plume detection computational system with the REVEAL data stream and demonstrate its operation in a relevant environment such as a port, industrial area, or other highly cluttered background. Algorithm performance will be evaluated as a function of environmental conditions and retraining parameters and reported in terms of false positive and true positive rates.