Robust Arrival Time Uncertainty Estimation Using Gaussian Blurring

Date: 4/24/2019

Time: 09:30 AM

Room: Elliott Bay

The uncertainty of the arrival time of a picked seismic phase directly controls both the detection and characterization of the associated event, yet relatively little work has been done on developing robust uncertainty estimation methods, and the standard approach is to use a simple algebraic scaling relationship based on signal-to-noise ratio (SNR). We present a case study in seismic onset detection, comparing statistically computed multi-band distributions over possible signal onset times to a range of onset times determined by an expert seismic analyst (best time, earliest possible time, latest possible time). Importantly, the uncertainty distributions sometimes identify subtle changes in the seismic waveform that are missed by both point estimate calculations and by analysts.

We generate our statistically computed distribution for each band by first creating a large set of Gaussian Blurred samples of the band-passed version of our original signal. We then use each sample to determine the best onset detection time by fitting statistical models to noise and signal data separately, then finding the point in time where the summation of the noise and signal fit scores is best. Once we have determined the onset time for all our Gaussian Blurred samples within a band using this method, we aggregate the results into a likelihood distribution over possible onset times, which we then compare to the range of pick times estimated by our expert analyst.

Presenting Author: Matt G. Peterson

Authors