Earthquake Detection in Develocorder Films: An Image-based Detection Neural Network for Analog Seismograms

Kaiwen Wang, Weiqiang Zhu, William L. Ellsworth, & Gregory C. Beroza

Published August 14, 2019, SCEC Contribution #9604, 2019 SCEC Annual Meeting Poster #066

From the late 19th century into the 1970s, seismograms were almost exclusively recorded in analog form. It was not until the 1980s that digital recording began to replace analog recording. Over 100 years of analog seismograms - about 50 million - still exist in seismic observatories worldwide, as estimated by IRIS Seismo Archives. This includes about one million film seismograms. Many of these analog films were recorded by a Develocorder, with as many as 18 seismic channels recorded in strip chart mode on each 16 mm microfilm. These films contain valuable information about historical seismicity and record unique experiments that are unlikely to be repeated today. Develocorder films are challenging to convert into time series because traces cross and fade in intensity when an earthquake arrives. Taking advantage of recent development in machine learning techniques for computer vision, we propose a new path to perform image-based seismic processing without the need to create vector digital time series. We trained an image-based neural network to detect earthquakes directly from film scans. One challenge we faced for training an analog film detector is the lack of labels for events and noise on the films. Thousands of labels are needed to apply deep learning methods. To solve this problem, we synthesize Develocorder film images using well-labeled digital waveforms. The synthetic images are carefully calibrated using a Long Valley, California swarm for which both digital seismograms and Develocorder films are available. The neural network takes film scan images as input and outputs the probability of an earthquake as a function of time (horizontal position) in the image. Once detected, earthquakes can be located by following steps of phase picking, association, and correlation of the earthquake images. We tested the performance of the neural network on the Develocorder films of the Rangely earthquake control experiment and demonstrated that our image-based method can achieve near human performance without manual intervention. This demonstrates the potential of our alternative approach to process historical analog data and to improve the accuracy and completeness of historical, pre-digital catalogs.

Wang, K., Zhu, W., Ellsworth, W. L., & Beroza, G. C. (2019, 08). Earthquake Detection in Develocorder Films: An Image-based Detection Neural Network for Analog Seismograms. Poster Presentation at 2019 SCEC Annual Meeting.

Related Projects & Working Groups