Statistical Interpretation of Webnet Seismograms By Artificial Neural Nets

We employed multilayer perceptrons (MLP), self organizing feature maps (SOFM), and learning vector quantization (LVQ) to reveal and interpret statistically significant features of different categories of waveform parameter vectors extracted from three-component WEBNET velocigrams. In this contribution we present and discuss in a summarizing manner the results of (i) SOFM classification and MLP discrimination between microearthquakes and explosions on the basis of single-station spectral and amplitude parameter vectors, (ii) SOFM/LVQ recognition of initial onset polarities from PV'-waveforms, and (iii) a source mechanism study of the January 1997 microearthquake swarm based on SOFM classification of combined multi-station PV-onset polarity and SH/PVamplitude ratio (CPA) data. Unsupervised SOFM classification of 497 NKC seismograms revealed that the best discriminants are pure spectral parameter vectors for the recognition of microearthquakes (reliability 95% with 30 spectral parameters), and mixed amplitude and spectral parameter vectors for the recognition of explosions (reliability 98% with 41 amplitude and 30 spectral parameters). The optimal MLP, trained with the standard backpropagation error method by one randomly selected half of a set of 312 mixed (7 amplitude and 7 spectral) single-station (NKC) microearthquake and explosion parameter vectors and tested by the other half-set, and vice versa, correctly classified, on average, 99% of all events. From a set of NKC PV-waveform vectors for 375 events, the optimal LVQ net correctly classified, on average, 98% of all up and 97% of all down onsets, and assigned the likely correct polarity to 85% of the onsets that were visually classified as uncertain. Optimal SOFM architectures categorized the CPA parameter vector sets for 145 January 97 events individually for each of five stations (KOC, KRC, SKC, NKC, LAC) quite unambiguously and stable into three statistically significant classes. The nature of the coincidence of these classes among the stations that provided most reliable mechanism-relevant information (KOC, KRC, SKC) points at the occurence of further seven statistically significant subclassses of mechanisms during the swarm. The ten neural classes of focal mechanisms coincide fairly well with those obtained by moment tensor inversion of P and SH polarities and amplitudes extracted from the seismograms interactively. The obtained results, together with those of refined hypocenter location, imply that the focal area consisted of three dominant faults and at least seven subfaults within a volume of not more than 1 km in diameter that likely were seismically activated by vertical stress from underneath.