Transfer-Learning-Based SVM Method for Seismic Phase Picking With Insufficient Training Samples

Abstract

Efficient seismic phase picking is fundamental to seismic signal processing. Phase picking methods based on neural networks show great potential in accurately picking signals with a low signal-to-noise ratio but require large training datasets. We present a transductive transfer-learning-based support vector machine (TTL-SVM) algorithm for seismic phase picking when the seismic dataset possesses insufficient training samples. An objective function of TTL-SVM, which is incorporated with a pretraining classification process in the source domain that possesses an adequate training dataset and quality labeling, is proposed for phase picking in the target domain with no quality labeling. Seismic compressional ( <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$P$ </tex-math></inline-formula> -) and shear ( <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$S$ </tex-math></inline-formula> -) phase picking is performed using two TTL-SVM processing steps: seismic phase and noise classification, and then <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$P$ </tex-math></inline-formula> - and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$S$ </tex-math></inline-formula> -phase classification from the picking phases. Experiments are performed to test the algorithm using a simulated dataset and two earthquake datasets from Jiuzaigou in China and New Zealand. The TTL-SVM results are remarkable compared with those obtained through traditional automatic and manual picking approaches. This algorithm provides an alternative approach for seismic phase picking when the dataset possesses insufficient training samples.