Classifying Worldwide Standardized Seismograph Network Records Using a Simple Convolution Neural Network

Abstract

Abstract The U.S. Geological Survey (USGS) maintains an archive of 189,180 digitized scans of analog seismic records from the World-Wide Standardized Seismograph Network (WWSSN). Although these scans have been made public, the archive is too large to manually review, and few researchers have utilized large numbers of these records. To facilitate further research using this historical dataset, we develop a simple convolutional neural network (CNN) that rapidly (∼4.75 s/film chip) classifies scanned film chip images (called “chips,” because they are individually cut segments of 70 mm film) into four categories of “interestingness” to earthquake seismologists based on the presence of earthquakes and other seismic signals in the record: “no interest,” “little interest,” “interest,” and “high interest.” The CNN, dubbed “Seismic Analog Record Network” (SARNet), can identify four types of seismic traces (“no events,” “minor events,” “major events,” and “errors”) in 200 × 200 pixel subcrops with an accuracy of 92% using a confidence threshold of 85%. SARNet then converts 100 random subcrops from each film chip into the overall classification of interestingness. In this task, SARNet performed as well as expert human classifiers in determining the film chip’s overall interest grade. Applying SARNet to 34,000 film chips in the WWSSN archive found that 21% of the images were of “high interest” and had an “indeterminate” rate of only 4%. Thus, the need for the manual review of images was reduced by 79%. Sorting of film chips derived from SARNet will expedite further exploration of the archive of digitized analog seismic records stored at the USGS.