Application of machine learning methods for earthquake detection from high-density temporary observation seismic records on a volcanic island

We applied two machine learning models to detect earthquakes from records observed with seismometers temporarily installed on a volcanic island. The two models are based on different principles: one regards seismic waveforms as images, using a convolutional neural network (CNN) to determine the first arrival times of P-waves, S-waves. The other model regards seismic waveforms as series data. The model processes seismic waveforms as data in a specific order of noise, P-wave, and S-wave, similar to natural language.

The purpose of this study is to present the results of using machine learning first arrival times identification models with two principles for noisy seismic waveforms, caused by sea waves and strong winds in volcanic islands, and to evaluate the effectiveness of machine learning models for noisy observation records.

We created a Confusion Matrix using first arrival times determined by an expert and evaluated the detection performance of these two models using some metrics of the matrix. Additionally, we assessed accuracy of the model-identified first arrival times by generating a frequency distribution of the difference from the expert's detecting time.

The study discovered that the model treating data as series had superior detection ability for noisy data compared to the one treating data as images and the accuracy of the first arrival time detection was also better for the series data model too.

We compared the results obtained on this island with those obtained at the permanent station, which is considered to have less noise interference, described in Mousavi et al., 2020. It was found that the difference in detection ability between the two models is slight for data obtained at permanent stations with low noise interference, but that the difference in detection ability between the algorithms of the two models is significant in noisy environments.