Convolutional Neural Networks & Deep Learning on Spectrograms for Earthquake Detection

Date: 4/24/2019

Time: 02:45 PM

Room: Elliott Bay

Continuous seismic waveform data recorded by a seismometer may potentially contain many events undetected by conventional techniques like STA/LTA or cross correlation. It is challenging to identify low signal-to-noise ratio seismic events when using conventional methods. The benchmark technique, autocorrelation, requires exponentially higher run time over longer time periods - it is infeasible to apply autocorrelation to even weeks of data. With recent advances in computing power and widespread access to large seismic datasets, machine learning can be applied to develop more robust seismic detection algorithms.

Data has been growing exponentially, with access to global seismic databases like IRIS enabling big data innovation. Our novel seismic window detection technique leverages advances in deep learning in order to achieve state of the art seismic event detection. Specifically, we use convolutional neural networks (CNN) to discriminate seismic events from noise. Instead of using raw seismic waveforms as the input, a series of preprocessing strategies constructs modified spectrograms. CNNs are well established in the image domain, having essentially “solved” image classification problems like handwriting recognition. This makes CNNs well suited to be used with spectrograms.

Our technique achieves very high detection accuracy even with low signal-to-noise ratio events. When the training data is from the same source as the test, performance is excellent, achieving 99%+ accuracy with linearly scaling runtime. Furthermore, this technique has potential to be generalized. Using event training data from a broad curation of different event sources can be used to train a model that can detect events in test data originating from a completely different source. We use 100+ stations in 30 locations in order to build a more generalized model, albeit with less consistency than the single source model. Current work involves increasing accuracy of the generalized detection model.

Presenting Author: James Audretsch

Authors