An automated, deep-learning-based method for investigating spatial-temporal evolution of seismicity

Abstract

Earthquake migration patterns are important to reveal various triggering mechanisms, including the tectonic process and those caused by anthropogenic activities. Mapping out the spatial-temporal seismicity pattern is traditionally conducted using reference marks either in spatial or time. However, such mapping is particularly challenging for induced earthquakes because most industrial records that provide reference marks are unavailable to the public. Moreover, advances in earthquake detection techniques proliferate earthquake catalogs and thus require labor-intensive investigation. Therefore, a new methodology is demanded to automatically investigate spatial-temporal patterns of seismicity without reference marks. Here, we present a deep learning-based method to automatically identify the timings and locations of anomalous seismicity, defined by the sudden change of earthquakes in a region. We first rasterize multi-dimensional earthquake catalogs into 2-D distribution maps. Then, we identify the maps with anomalous seismicities and extract their timings and locations to generate condensed catalogs to reduce the manual effort in further investigation. We choose Changning and Weiyuan in Sichuan Basin as our study areas due to their high seismicity rates in recent years. We use the Changning catalog to train the method and the Weiyuan catalog to test the method's spatial transferability. Our approach successfully condenses both the Changning and Weiyuan catalogs with the accuracy of 0.87 based on the F1 score. The anomalous seismicities identified by our network include both earthquakes associated with hydraulic fracturing and aftershocks following strong quakes. As such, our method could be applied to broader areas with more complex migration patterns, including natural earthquake sequences.