Use of decision tree ensembles for crustal structure imaging from receiver functions

Abstract

SUMMARY Mode conversion of P waves at the boundary between Earth's crust and upper mantle, when analysed using receiver functions (RFs), allows characterization of Earth structure where seismic station density is high and earthquake sources are favourably distributed. We applied two ensemble decision tree algorithms—Random Forest (RanFor) and eXtreme Gradient Boost (XGBoost)—to synthetic and real RF data to assess these machine learning techniques' potential for crustal imaging when available data are sparse. The synthetic RFs, entailing both sharp increases in seismic velocity across the Moho and gradational Moho structures, calculated with and without added random noise, correspond to idealized crustal structures: a dipping Moho, Moho offset by crustal-scale faults, anti- and synform Moho structures and combinations of these. The RanFor/XGBoost algorithm recovers input structures well regardless of event-station distributions. Useful crustal and upper mantle seismic velocities can also be determined using RanFor and XGBoost, making it possible to image crustal thickness and P- and S-wave velocities simultaneously from RFs alone. We applied the trained RanFor/XGBoost to RFs determined from real seismic data recorded in the contiguous United States, producing a map of the Moho and P- and S-wave velocities of the lowermost crust and uppermost mantle. Use of XGBoost, which evaluates residuals between input RFs and ground-truth to update the decision tree using the gradient of a penalty function, improves the crustal thickness estimates.