No abstract

Ð The Prototype International Data Center (PIDC) has designed and implemented a system to process data from the International Monitoring System's hydroacoustic network. The automatic system detects and measures various signal characteristics that are then used to classify the signal into one of three categories. The detected signals are combined with the seismic and infrasonic detections to automatically form event hypotheses. The automatic results are reviewed by human analysts to form the Reviewed Event Bulletin (REB). Continuous processing of hydroacoustic data has been in place since May 1997 and during that time a large database of hydroacoustic signals has been accumulated. For a two-year period, the REB contains 13,582 T phases that are associated to 8,437 events. This is roughly 25% of REB events after taking station downtime into account. Predicted travel times used in locations are based on the arrival time of the peak amplitude mode calculated from a normal mode propagation model. Global sound velocity and bathymetry databases are used to obtain reliable 2-D, seasonally dependent, travel-time tables for each hydroacoustic station in the PIDC. A limited number of ground-truth observations indicate that the predicted travel times are good to within 5 seconds for paths extending to over 7,000 km ± corresponding to a relative error of less than 0.1%. The ground truth indicates that the random errors in measuring arrival times for impulsive signals are between 1 and 6 seconds. This paper describes and evaluates the automatic hydroacoustic processing compared to the analyst reviewed results. In addition, special studies help characterize the overall performance of the hydroacoustic network.

A neural network employing the back propagation learning paradigm has been developed as an experiment in the automatic classification of small regional earthquakes and quarry explosions. The network has been used in the analysis of 66 events recorded by the NORESS array in southern Norway. The input vector consists of three broadband discriminants including the spectral ratios of Sn/Pn and Lg/Pn waves, and the mean cepstral variance of Pn, Sn, and Lg. Two hidden layers are used, consisting of 8 and 2 units. The output vector consists of two units which correspond to the classification of explosion or earthquake. The network was first trained using input vectors from the entire dataset. The network was able to perfectly model the training set with no classification errors. For comparison, an optimum linear classifier used with the same dataset resulted in 5 errors and 19 uncertain classifications. Next, the network was trained with half of the events and tested with the remaining half. This resulted in 5 errors and 2 uncertain classifications. This compares with 5 errors and 18 uncertain events for the optimum linear classifier. The apparent advantage of the neural network over the optimum linear classifier is the network's ability to model complex decision regions and in the reduction of the number of uncertain events.

In support of the proposed Comprehensive Nuclear Test Ban Treaty, a large database of hydrophone recordings including T-phases, explosions, and noise has been compiled and cross referenced with known seismic events at the Center for Monitoring Research. Using this database, an automated hydroacoustic arrival detection and classification system has been developed. Detection is accomplished with a long-term-average/short-term-average power detector operating in several passbands. Station specific tuning of SNR thresholds and passband bounds allows the detector to trigger reliably on T-phases and explosions while passing over the majority of noise events such as whale calls. For each detected arrival, features such as duration, energy moments, spectral ratio, and order statistics are measured in multiple passbands from 2–85 Hz. A neural network uses these features to classify each arrival as signal or noise. Declared signals are passed to a second-stage network which classifies them as T-phases, explosions, or unknown events. T-phases arriving within a 4-min window around the time predicted from a seismic location are associated with that seismic event. These associations reveal relationships among event parameters such as location, magnitude, depth, duration, and coupling region.

The objective of this work is to develop a surface modeling system capable of characterizing large gridded battymetric databases with potential applications to problems such as the extrapolation of survey data to higher resolution, the interpolation of bottom characteristics between surveyed regions, and the correlation of bottom features with other geophysical measurements. The system employs a two-dimensional stochastic seafloor model to represent blocks of gridded battymerry by a small number of model parameters that describe the physical features of the ocean bottom. This study focuses on the inversion component which is designed to estimate the model parameters quickly without iteration or starting values. Quality high-speed inversions are obtained using machine learning techniques, overcoming many practical limitaions imposed by conventional least squares techniques. This approach leads to the approximation of a direct mapping between the moments of the Fourier transformed data and the model parameter variables. Mapping is accomplished by the universal approximation capabilities of multilayer networks. Validation tests performed on synthetic examples spanning the entire model parameter space yield parameter estimation errors less than 5%. The modeling system is applied to a large northeast Pacific data set to examine correlations between bottom characteristics and sediment thickness. Errors in model parameters estimated from the battymetric data by Monte Carlo simulation are generally less than 10% dependent on data quality and conformance to model assumptions. The major advantages offered by this approach are clearly demonstrated by its error tolerance and the speed of inverson processing.