where Y is in m∕s2, a = -2.1575, b = 0.8359, c = -1.9690 and σ = 0.3542.
Data from 61 stations on Eurocode 8 class B (360 ≤ Vs,30< 800m∕s) sites mainly from 33 stations in the
Irpinia Seismic Network (ISNet, located in Campania and Basilicata regions) with the addition of data
(ML> 3) from other networks. Originally collect data from sites of other classes but as 93% of these
accelerograms from class B develop model only using those records.
Exclude data with focal depth h > 30km. 1 ≤ h ≤ 28km with most 12–20km.
Model for use in earthquake early warning system.
Linearly detrend records. Then bandpass filter (4-pole Butterworth with cut-offs of 0.075 and 20Hz) and
2% cosine taper. Exclude records with signal-to-noise ratio < 10 on either component using the pre-event
portion as the noise estimate. Exclude some records from before 2002 due to late triggering.
Did not include term to account for focal mechanism (normal v strike-slip) because of difficulty in
estimating mechanism for small earthquakes immediately after occurrence.
Regression method gives higher weight (0.7) to data recorded at stations with geotechnical and geophysical
measurements compared with sites with only estimated B class (0.3).
Compare predictions to an independent dataset (data from ISNet from 2015, 262 records from 41
earthquakes, 1.5 ≤ ML≤ 3.2, 3 ≤ rhyp≤ 60km) using the LLH method of Scherbaum et al. (2009). Find
LLH of derived model is the lowest of all those GMPEs test using these data.
Compare predictions and observations for 2 well-recorded earthquakes (ML3.2 and ML2.1) and find good