On Using Deep Learning for Automatic Classification System of Microseisms at Cotopaxi Volcano

This paper presents an automatic classification system of Long Period (LP) and Volcano Tectonic (VT) events from the Cotopaxi volcano, by using two Deep Learning techniques: Stack Autoencoders and Deep Neural Networks. We used a dataset provided by the Instituto Geofísico de la Escuela Politécnica Nacional (IGEPN) containing 1,044 LP and 101 VT events, in addition, we employed data augmentation process by considering a synthetic microseisms generator based on Conditional Generative Adversarial Network. Our dataset comprises 10,000 microseisms, including both real and synthetic instances. In the feature extraction stage, which involves computing Power Spectral Density (PSD) by using Welch's estimator with a Fast Fourier Transform (FFT) of 512 points, resulting in 257 frequency-based features. Additionally, we extracted coefficients from the Discrete Wavelet Transform using Symlet as the mother wavelet with six decomposition levels; then, PSD was computed on these coefficients, yielding an additional 257 frequency-scale features. Deep Learning algorithms were trained on this feature rich dataset and evaluated using standard performance metrics. The autoencoder technique yielded the best results, achieving an accuracy of 99.73% and a balanced error rate (BER) of 0.0027, indicating misclassification of only 3 out of 1000 microseisms. These results exceed the geophysical requirement of a BER below 0.01. This study demonstrates the effectiveness of Deep Learning techniques in classifying LP and VT events, offering promising avenues for improved volcanic risk monitoring and assessment.