Reconstruction of Clipped Seismic Waveform
Summary
Seismic waveforms are clipped when ground motions exceed the dynamic range of seis
mometers. Discarding these clipped waveforms impedes high-resolution studies of Earth’s
structure and ground motion analysis. In this study, we develop and evaluate four meth
ods to reconstruct clipped waveforms: polynomial interpolation, projection onto convex
sets (POCS), similar relative source time function (rSTF) method, and machine learning
method.
The polynomial interpolation method is simple and computationally efficient, providing
satisfactory reconstructed waveforms. However, its low stability is a significant limita
tion. The POCS method combines both time and frequency-domain constraints, yield
ing robust results for moderate and slight clipping scenarios but fails to reconstruct large
clipped zones. The innovative similar rSTF method does not yield reasonable reconstruc
tions due to ill-posed issues. The machine learning method with the hybrid CNN+LSTM
model shows the best accuracy across various clipping scenarios. The main challenge of
this method is the underestimation bias in strongly clipped waveforms.
After identifying machine learning as the optimal approach, we apply it to reconstruct
the real clipped waveforms from the M6.6 Italy earthquake in 2017. The result shows that
when both the unclipped waveform and the reconstructed clipped waveform are utilized,
the Peak Ground Velocity (PGV) in the near field is lower compared to the PGV predicted
using only the unclipped data.
This research demonstrates the effectiveness of machine learning approaches in clipped
waveform reconstruction. Future work should focus on optimizing loss functions, iden
tifying more suitable training datasets, and developing new machine learning architec
tures.