Time–Frequency characterization of microearthquakes based on Convolutional Neural Networks and explainability models

Given the potential destructiveness of volcanic eruptions, the study of volcanic microearthquakes is a key tool for improving our understanding of the relationship between a volcano and its environment. In the last decade, Convolutional Neural Networks (CNN) have shown great potential for the automatic classification of microearthquakes. However, one major limitation is their “black box” nature, it is often unclear which features drive their decisions. In this work, we propose the use of Explainability Models in conjunction with CNNs and two novel Time–Frequency representations: Adaptive Superlets (ASLT) and Spectrogram-based Periodogram with Window Switching (SPWS). The aim of this paper is to extract the Time–Frequency features leveraged by CNNs for microearthquake classification. This could be used to increase the reliability of CNN-based recognition systems and to identify possible new Time–Frequency characteristics that identify microearthquakes. This proposal verifies the frequency bands for Long Period (LP), Volcano Tectonic (VT), Tectonic (TC), and Tremor (TR) microearthquakes. Moreover, this could be useful for identifying frequency components that can be used to distinguish between LP and VT events and to determine the starting point of overlapping events within the same detection window.