GNGTS 2023 - Atti del 41° Convegno Nazionale

Session 1.1 GNGTS 2023 is) in use at INGV, while DB was chosen being the latest available updated for Italy and because it is going to be implemented soon as the routine INGV M L . For the event M L estimation, we also applied two different, both iterative, statistical approaches. (Fig. 2). The first method is " Outliers’ Removal ", where the median is iteratively re-calculated, and outliers are removed at each iteration based on the new median absolute deviation. No weighting is applied. The second method is the " Huber Weighted Mean " ( Huber, 2011 ) , where the mean is iteratively recalculated over all the stations' M L , and weighting is applied based on the distance from the latest mean, with full weight < 0.3, and down-weighting beyond. The comparison is based on the same attenuation law, the DB, and the largest differences are found for M L ≤ 2.5. We finally compared the results and the respective standard deviations of the two attenuation laws, basing on the Huber Weighted Mean, that is the method in use at INGV. The results highlight that, as expected, DB is slightly lower than HB for M L ≤ 2.5 while it is slightly higher for M L ≥ 2.5 (Fig. 3a). In addition, HB from PyML is frequently lower for small magnitudes with respect to INGV HB (Fig. 3b) and this is probably due to both different amplitudes’ datasets (we only use revised P-picked waveforms while INGV can include automatic earthworm amplitudes, especially in the past) and the use of adaptive filtering that preserves signal in noisy waveforms for smaller M L events. Issues Overlapping, or even too close, earthquakes, especially during seismic sequences, and saturation of the signal (maximum amplitude is out of the upper-limit dynamic range of seismometer) for stronger events at close stations, are the two main issues to handle in an automatic procedure. For overlapping earthquakes, we take advantage of the a-priori INGV Bulletin information, redesigning the search-window to avoid extending it in the next event signal. About the saturation issue, at least three kind of clipping are recognized (see for example Wang, S., and J. Zhang 2020 and references therein): flat-top, back-to-zero, and wrap-around. All this kind of clipping are present for larger magnitude events in our dataset at the closest stations. Conclusions The result of applying our on-purpose designed method for amplitude and local magnitude calculation to millions of digital waveforms from the Italian seismological network is a dataset consisting of ~3Mln stations amplitudes (x3 channels), and ~270k event M L between 2009 and 2018. This dataset is characterized by homogeneity (the same peak-to-peak search method has been applied to all the recordings and the same attenuation law to all the stations’ M L ), quality (the amplitudes are searched for when a P-pick is present, and the pre-filtering method is adaptive, preserving the seismic signal), reproducibility (we collect all the output parameters for each calculated amplitude), up-to-date (the latest attenuation law specific for the Italian region, Di Bona et al. [2016], is applied).