Rapid Prediction of Earthquake Ground Shaking Intensity Using Raw Waveform Data and a Convolutional Neural Network

Date: 4/24/2019

Time: 11:30 AM

Room: Elliott Bay

Effective early-warning, emergency response and information dissemination for earthquakes and tsunamis require rapid characterization of an earthquake's location, size, ground shaking and other parameters, usually provided by real-time seismogram analysis using established, rule-based, seismological procedures. Powerful, new machine learning (ML) tools analyze basic data using little or no rule-based knowledge, allows us to process the information we have about the earthquake in a new way. A ML deep convolutional neural network (CNN) can operate directly on seismogram waveforms with little pre-processing and without feature extraction.

Here we examine the application of CNN's to rapidly predict ground shaking intensity, using the Engineering Strong Motion Database ( Luzi et al. (2016)) of accelerometric waveforms and event meta-data for local earthquakes (up to 700 km epicentral distance). Lomax et al. (2019) used a CNN to predict earthquake parameters (distance, azimuth, depth and magnitude) using three-component single station waveforms for local, regional and teleseismic distances.

We extend the algorithm of Lomax et al. (2019) by exploring new machine learning techniques and data transformations. We find several ways to improvement the CNN algorithm for predict ground shaking intensity. Firstly, we separate the input of the normalized waveforms from the metadata, and then analysed these by separate networks which were then concatenated in an ensemble. Secondly, we use a LSTM (Long-short term memory network) which gave similar results to a CNN. Thirdly, we use an ensemble of a CNN and a LSTM to allow the network to learn local patterns (CNN) and the order of the patterns (LSTM) on the waveform. Additional improvement may be obtained with the use of two or more station waveforms to give even more information to the machine learning model. The ML CNN and classical methods will be compared for algorithm efficiency and accuracy of shaking intensity predictions.

Presenting Author: Dario Jozinovic

Authors