DevelNet: Earthquake Detection on Develocorder Films with Deep Learning: Application to the Rangely Earthquake Control Experiment

There exists over a century of instrumental seismic data; however, most seismograms recorded before the 1980s are only available in analog form. Although analog seismograms are of great value, they are underutilized due to the difficulties of making quantitative measurements on the original media and in converting them to digital time series. In this study, we present an alternative workflow, based on deep learning, to reconstruct an earthquake catalog from images of analog data without conversion to vector time series. We trained a convolutional neural network-DevelNet, using synthetic analog data to detect earthquakes on scanned multichannel Develocorder film images. We then developed an image-based processing workflow to measure arrival times, locate, and determine the magnitudes of earthquakes in the data. We demonstrate the performance of this approach on two years of continuous Develocorder film recordings from the Rangely earthquake control experiment in the mid-1970s. Our approach detects twice the number of events reported in the original catalog (Raleigh et al., 1976). This demonstrates that DevelNet efficiently detects earthquakes from Develocorder film scans, performs consistently over time, and is robust to changes in network geometry. Our locations generally agree with the original study, although the automatically measured arrival times are less precise than manual reading, leading to increased location scatter. Our automatic workflow of Develocorder films rivals the performance of skilled analysts in earthquake detection, but with minimal human intervention. This image-based processing offers a new approach for effectively and efficiently extracting earthquake information from analog seismic data.

Automated summary

This study introduces an alternative workflow based on deep learning to reconstruct earthquake catalogs from analog data images without conversion. By training a convolutional neural network, the approach successfully detects twice the number of events reported in the original catalog and achieves consistent performance in earthquake detection.