Deep Learning Event Detector from Long-term Signal Variation for Seismic Activity Warning out of Schumann Resonance

Abstract

Deep Learning (DL) has shown capability in many areas of impact on everyday life. The paper proposes a DL architecture tailored for event detection from examining the time evolution of a signal. With temporal characteristics extracted by a Convolutional Neural Network (CNN) encoder and fed as input to a recurrent neural network, the model targets the detection of a possibly occurring investigated event in the given time interval. The utility of DL methodologies to solve physical problems is demonstrated for an application of the complex experimentally-studied existing interaction between Schumann Resonance (SR) and seismic activity. SR signals are electromagnetic waves propagating along the Earth-ionosphere cavity. Intense lightning activity is continuously present at the same locations around the world, being sensitive to physical perturbation. Seismic activity modifies this steady lightning pattern. The new DL model is applied to answer the research question of whether the variation of the SR signal is truly a verifiable forerunner of seismic activity. Several parameter configurations are explored, either model-related or linked to criteria for selecting seismic events. Results show preliminary evidence about the relation between distance-intensity space and SR perturbation, and provide valuable corroboration about the sensitivity of the sensor to a specific azimuth between the observatory and the Earthquake (EQ) epicenter, hence argumentatively supporting the SR temporal characteristics as an early seismic warning. This is the first generalization of seismic disturbance as a derivative of the SR, based only on its signal time series variation, as a hypothesized precursor of the EQ event.