Extracting continuous time series from scanned images of WWSSN seismograms using deep learning: Application to tropical storms

Abstract:

The 3+ decades of seismograms from the World-Wide Standardized Seismograph Network (WWSSN) represent a remarkable archive of continuous recording of the ambient field. Today, such data can be useful for exploiting new quantitative constraints on oceanic and atmospheric events happening during the beginning of the pre-satellite era and searching for climatically related changes. However, the millions of WWSSN seismograms are today "dark data", in the form of microfilm, microfiche and some scanned images, that cannot be easily analyzed using modern methods. For such data to be usable, we need to recover quantitative information from those analog records.

We take a deep learning approach to this problem, specifically through segmentation of the scanned photographic image into traces and subsequent conversion to time series. For this approach, training data is needed. To that end, we use data from modern digital instruments to synthesize image files that replicate the appearance of WWSSN records. This requires attention to the nuances of the analog records, including the width and slight blurriness of the traces (which, on the WWSSN records, were made by a light beam exposing photographic paper) and the minute and hour marks created by the periodic deflections of the analog light beam. During the segmentation model learning process, labels are needed to calculate the error of the model's predictions so that corrections can be made. In this case, the labels are images marked with the correct segmentation. Using synthetic training data makes labeling simple, as labels can be provided upon generation. We focus on seismic records at WWSSN stations in the Caribbean and extract the time series from periods with no earthquakes to access the seismic data generated by strong tropical storms and cyclones in the northern Atlantic Ocean.