Recognizing explosion sites without seismogram readings: neural network analysis of envelope-transformed multistation SP recordings 3–6 Hz

Seismic-waveform similarities for closely spaced earthquakes and explosions in particular are well established observationally. In many industrialized countries of low seismicity more than 90 per cent of seismic event recordings stem from chemical explosions and thus contribute significantly to the daily analyst workload. In this study we explore the possibility of using envelope waveforms from a priori known explosion sites (learning) for recognizing subsequent explosions from the same site excluding any analyst interference. To ensure high signal correlation while retaining good SNRs we used envelope-transformed waveforms, including both the P and Lg arrivals. To ensure good spatial resolution we used multistation (network) recordings. The interpolation and approximation neural network (IANN) of Winston (1993) was used for teaching the computer to recognize new explosion recordings from a specific site using detector output event files of waveforms only. The IANN output is a single number between 0 and 1, and on this scale an acceptance threshold of 0.4 proved appropriate. We obtained 100 per cent correct decisions between two sets of 'site explosions' and hundreds of 'non-site' explosions/earthquakes using data files from the Norwegian Seismograph Network.