Detection of volcanic eruptions using auto-encoder of infrasound

Infrasound often radiates when a volcano explodes, suggesting its use for detecting volcanic eruptions. In fact, there is an attempt to use infrasound for long-range monitoring of volcanic eruptions, using infrasound stations of the Comprehensive Nuclear-Test-Ban Treaty Organization (CTBTO). The problem here is the signal to noise ratio (SNR) which considerably fluctuates depending on the measurement environment and weather conditions. In this study, an acoustic anomaly detection using auto-encoder is applied to infrasound in order to improve SNR. An auto-encoder consisting of three layers is employed, and segmented infrasound signals of 12.5 s are directly fed to its input. The data obtained during the time period where no volcanic eruptions is reported by Japan Meteorological Agency (JMA) are used for training. After the training, the root mean squared error between the output of the auto-encoder and the input is considered as a signal (likelihood) of the occurrence of volcanic eruptions. Evaluation using the data obtained near Sakurajima shows that SNR is improved by approximately two times in some case. As for detection performance, while it shows a good agreement with those reported by JMA for explosions, there are some discrepancies for eruptions.