Automated slowness measurement and uncertainty estimation of multiple seismic arrivals using bootstrapping, cluster analysis and array methods

Abstract:

Observations of backazimuth (direction) and horizontal slowness (inclination) of seismic arrivals have been used to analyse mantle structures such as Large Low Velocity Provinces (LLVPs), subducting slabs or mid-mantle reflectors. These observations have been essential to constrain the morphology and seismic properties of mantle heterogeneity and therefore how they contribute to whole mantle dynamics. These studies are limited in scale because of the time-consuming nature required in visual inspection of observations. In addition, uncertainty estimates are also challenging to make and are typically ignored.

We address these limitations by developing an algorithm to automatically identify one or multiple (multipathed) seismic arrivals and measure their direction and inclination properties with uncertainty estimates. To do this, we perform bootstrapping on the traces recorded for each event at a seismic array. Then, on each bootstrap sample, we use beamforming to recover data points describing directions and inclinations of potential arrivals. From this, the clustering algorithm DBSCAN is applied to all the recovered points and identify the arrivals. The number of clusters represents the number of arrivals and the cluster size is indicative of the uncertainties of the arrival properties such as its direction and inclination.

We apply this algorithm to a dataset sampling the African LLVP to identify arrivals perturbed to arrive off the great circle path and identify multiple arrivals, which are indicative of strong lateral velocity gradients. We compare the predicted number of arrivals (1,2 or 0) to the number of arrivals identified manually. We find this method can correctly identify 90% of the observations with a single clear arrival and approximately 70% of the observations with two arrivals.