Enhancing Infrasonic Signal Detection and Event Localization Through Advanced Computational Methods

The installation of worldwide infrasound and seismo-acoustic networks at both global and regional scales necessitates automated techniques and algorithms for accurate and efficient data processing and analysis. Signals recorded across the networks originate from a number of natural and anthropogenic sources. Data processing efforts focus on the separation of signals of interest from background noise, followed by the identification, or detection of signals of interest. Once signals are identified, association and location processing produces estimates of a signal\u27s source. This dissertation focuses on the evaluation of automated processes for identifying and locating sources of interest. Chapter two applies two state-of-the-art automated infrasonic signal detectors to real and synthetic waveform data. Comparisons between the two detectors are produced across a variety of background noise conditions. The first detector, the Adaptive F-Detector, accounts for coherent noise across an array through the application of a C-value, which effectively reduces the detection threshold (p-value) and decreases the number of noise-related detections by constantly re-mapping the conventional F-statistic based on moving estimates of the background noise. The second detector, the multivariate adaptive learning detector, applies three distinct time windows to adaptively alter detection thresholds in order to account for the changing background noise environment. Performance is ultimately quantified in terms of recall and precision through validation with an analyst-derived catalog, where recall is the proportion of true objects found from all true samples and precision is the proportion of true objects from all found objects. Comparisons using real waveform data document similar precision and recall rates for the two detectors of between 45-79\% and 28-100\% across the network, respectively, indicating that performance across both detectors is nearly equal. Comparisons against the synthetic catalog indicate that both automated detectors identify long period signals with the same success rate, while successful identification of short period signals varies based on the detector methodology and background noise level. One detector, the Adaptive F-Detector, successfully identifies most short period signals and performs well across all background noise levels, motivating further study of the detector success and failure points in chapter three. Conclusions from this chapter lead to recommendations for future automated detector development. Following the comparison of multiple signal detectors, in chapter three a single detection algorithm, the Adaptive F-Detector, is utilized to further assess factors that influence infrasound signal detection at regional networks. Signatures from repeating above-surface explosive events, denoted ground truth events due to the known location and origin time, are detected with the Adaptive F-Detector and by a human analyst. The number of automatic detections varies as a function of array distance from the source where the arrays closest and furthest from the source detect very few events (~10/45). Analyst review adds additional GT detections at all stations, and particularly increases the number of detections within the dataset at stations near the source of interest. Successful detection increases to between 24-90% of GT events depending on the stations, as compared to success automatic detection of between 14-80% of the GT events. Situations when automated methodologies fail are evaluated through a combined background noise quantification, atmospheric propagation analyses and comparison of spectral amplitudes. Results indicate that detection capability is primarily related to station proximity to the source, driven by the atmospheric propagation of tropospheric and thermospheric infrasound signals. Detection capability can also be related to background noise levels at individual stations. This analysis provides an estimate of detector performance across the network as well as a qualitative assessment of conditions that impact infrasound monitoring capabilities. Finally, accompanying the evaluation of successful signal detection rates across the network, in chapter four the series of explosive events are used to evaluate recent improvements to infrasonic source localization methodologies. These improvements apply modeling-based predictions for atmospheric propagation to better constrain arrival celerities and enhance temporal and spatial localization estimates at regional distances between 58-410 km from the source. The network utilized in this study has a sub-optimal geometry where six out of seven arrays are located to the SE-SSE of the source and only one station offers azimuthal resolution in the NE; this station distribution may increase bias in location results. Locations are produced using three distinct signal celerity and backazimuth deviation models; (1) a generalized celerity model, derived from ray-tracing; (2) monthly Path Geometry Models (PGMs) for celerity and backazimuth based on station range and azimuth; and (3) a data-based empirical celerity model. Application of the PGMs both underestimates and overestimates signal celerities, leading to large errors in both spatial and temporal location across the GT dataset. These errors correspond to perceived improvements in spatial accuracy, where the 90% EE areas decrease, coupled with reductions in spatial precision, although the 90% Error Ellipses (EEs) do not contain the true GT location for 14/45 or ~33% of the events in this dataset. Use of the generalized celerity model and the empirical UTTR celerity models produce spatial location estimates with similar accuracy, while use of the empirical model improves localization precision. Additionally, 90% EE for all 45 events produced with both the generalized celerity model and the empirical model contain the true source location within their bounds. Results indicate that atmospheric specifications need refinement in order to accurately predict infrasonic signal propagation at regional distances; unrefined propagation results lead to errors within both spatial and temporal location estimates. These errors are driven by the predominantly tropospheric arrivals at stations as the resolution of current atmospheric specifications cannot account for the variability of the atmosphere near the surface of the earth. Results additionally demonstrate that bias and accuracy in localization results are driven by both network detection capability as well as network design with large location errors related to azimuthal gaps between detecting arrays. This result indicates that the applied localization methodology weighs detection backazimuth observations more heavily than travel time estimates and further suggests that additional research into backazimuth corrections may offer the opportunity to improve spatial localization results. These three chapters provide a detailed analysis of signal characteristics from surface explosions, leading to better constraints on the performance of automated detection and location algorithms applied in a regional network setting. Results, in the form of successful event detections and locations across the network are driven by the network geometry, where stations are located between 84-410 km from the source and 6/7 stations are located towards the SE-SSE of the source. Work presented in chapter two motivates further research into optimal techniques for direct comparisons of signal detectors, particularly focused on tuning precision and recall rates rather than producing direct parameter-based detections. Results from chapters two and three indicate that further detector development is necessary in order to reduce false detections related to coherent noise sources. Additionally, a deeper understanding of repeating noise sources at individual stations is needed. Results from chapters three and four demonstrate that an understanding of infrasound signal propagation dynamics is limited by the resolution of available atmospheric models; advances in atmospheric predictions will significantly improve the ability to model infrasound propagation, particularly in the lower atmosphere. Finally, location results in chapter four strongly suggest that network design introduces bias; this bias is driven by algorithms that more heavily weight backazimuth measurements over travel time measurements. A combination of improved infrasound networks with increased azimuthal resolution and refinement of location algorithms should continue to reduce errors and bias in event location results