Standardization of Noisy Volcanoseismic Waveforms as a Key Step toward Station‐Independent, Robust Automatic Recognition

This article addresses the automatic volcanoseismic recognition (VSR) in a noisy scenario studying the robustness of a classifier based on hidden Markov models (HMMs). The system learns recognition models analyzing signals recorded in 1995 to automatically detect and classify noisy events of 2009. Both datasets were acquired in different locations at Deception Island and with a different type of sensor showing a variety of site effects and noises. To deal with the inherent waveform variability of this setup, we propose to reconstruct the seismograms to achieve both modeling standardization and noise reduction goals. We analyze a set of empirical mode decomposition (EMD) algorithms jointly with static and dynamic reconstruction criteria to evaluate their impact on the robustness of the recognition process. This machine-learning focus on real time, continuous, unsupervised VSR paradigm is able to increase by 16% the global VSR accuracy using an adaptive reconstruction compared to the scores obtained without any standardization.