Automatic Detection of Amplitude-Distorted Samples from Clipped Seismic Waveforms

Abstract

Abstract Seismic waveforms are essential for seismology but are clipped when their actual amplitudes are too high to be faithfully recorded by seismometers. The clipping effects are popular for both big earthquakes and small earthquakes within a short epicentral distance. Here, we illustrate potential risks of direct usage of clipped waveforms by examining the frequency leakage and show the failure of bandpass filtering for different clipping levels; then we summarize two characteristics of clipped records: (1) The temporal gradient is unusually large around the clipped segment compared with the unclipped portions, and (2) the clipped samples cluster into one segment or several if many samples are involved. Next, we propose three criteria for distinguishing clipped samples from the perfect samples based on these two characteristics. Finally, we design a numerical algorithm for automatic detection of clipped samples using constraints on the gradient, amplitude, and gradient-varying range. Numerical experiments show the excellent performance of our algorithm on automatically detecting the clipped samples. Our algorithm seamlessly integrates all necessary constraints for both flat-top type and back-to-zero type and thus can correctly recognize these two types simultaneously; in addition, it is basically data driven and thus can work well without considering seismometer configuration and instrument type, which would be helpful for real-time detection of clipped records without interruption from human operations. As a robust and swift tool of automatic detection on amplitude-clipped samples, our algorithm could identify most typical clipped records and reduce potential risks due to using unrecognizable clipped waveforms; furthermore, it would be helpful for fast detection and possible restoration of clipped waveforms in the presence of huge volumes of data.