DisperPy: A machine learning based tool to automatically pick group velocity dispersion curves from earthquakes

Abstract

Seismology has made significant progress in high-resolution Earth imaging, largely driven by the increasing volume of freely available data. As a result, automated tools and machine learning algorithms are becoming essential for processing this vast amount of information. We present DisperPy, an open-source Python library developed to automatically extract group velocity dispersion curves from earthquake data. The analysis framework of DisperPy is structured around two primary tasks: (1) assessing the quality of waveforms to determine if dispersion extraction is feasible, and (2) measuring the group velocity dispersion curve for suitable waveforms. To address the first task, DisperPy uses a convolutional neural network trained on dispersion spectrograms to classify waveform quality. The model, based on the ResNet-34 architecture, is initialized with ImageNet-pretrained weights and fine-tuned using the fastai deep learning library. In the test set, the network achieves an accuracy of 92 % in distinguishing between high- and low-quality dispersion images. For the second task, DisperPy employs unsupervised learning techniques, starting with a Gaussian mixture model to separate dispersion energy from background noise, followed by k-means to separate the dispersion energy into clusters, making it easier to track amplitude maxima and then construct initial dispersion curves. Finally, a refinement of the initial dispersion is achieved using both the density-based spatial clustering of applications with noise algorithm and data quality criteria to remove possible outliers. To further test DisperPy, we conduct a surface wave tomography experiment across the contiguous United States using freely available vertical-component broadband waveforms. After processing the data with DisperPy and removing low-quality waveforms, the final dataset consisted of 194,325 unique dispersion curves. Consistent with previous studies, our maps reveal a prominent velocity dichotomy, with low velocities in the tectonically active western US and high velocities in the stable central and eastern US.