Algoritmos avanzados de procesamiento de señal basados en técnicas de deep learning para descripción y caracterización de señales sismo-volcánicas

Resumen

In highly populated areas where people are living near active volcanoes, volcano monitoring is an important task which helps to quantify the risk of potential eruptions. The myriad of seismic processes deep beneath the surface of volcanoes requires the development of early warning systems that can inform human societies about threats from volcanic hazards. Given the social and high economic impact generated by active volcanoes, there are several scientific disciplines to study volcanoes. Based on the study and analysis of volcanic regions using seismic data, volcanic seismology describes and characterizes the wide range of signals related to the surface manifestation of complex (physical and chemical) processes occurred in the Earth’s interior [#chouet1986dynamics, #alguacil1999observations, #ibanez2003recent]. Even though each eruptive scenario has different characteristics (magma rheology, morphology of the volcanic structure, position and origin of the magmatic source) that derive on a large variety of different seismic signals, there is a remarkable observation: many volcanoes show comparable seismic signal characteristics that can be associated to different seismo-volcanic sources [#chouet2013multi]. Therefore, one of the fundamental and most challenging objectives in volcano-seismology is to determine what always happens before an eruption in any scenario, which is differentiating between eruptions and that is only characteristic of some types of eruptive scenarios. The objective of seismic analysis is not only limited to identify volcano seismic signals, but to determinate the source that originates them. Early warning systems are mainly based on the analysis of the considered prior precursor events. By definition, a seismic precursor is any seismic signal prior to potential eruption and whose origin is related to the dynamic processes (like fluids) that cause it. In this sense, the development of an automatic system able to evaluate potential seismic sources and their relationship to the present volcanic process in real time will allow a more effective management of volcanic risk. Whilst direct applications of machine learning and signal processing could serve to improve monitoring and eruption forecasting, nature itself imposes several limitations that need to be taken into consideration. Hence, the implementation of automatic and robust monitoring systems is a difficult task: 1. The seismic signals are not entirely clear: seismic waves radiated by volcanic sources contain information not only about volcanic dynamic but also about the inner complex structure of the volcanic edifice affecting the seismic wave-field (source and attenuation effects), increasing the difficulty of geophysical interpretation from recorded signals [#palo2009analysis, #neuberg2000effects]. 2. Generalization capabilities from machine learning models are inferred from acquired data previously analyzed by geophysicists. However, data completeness plays an essential role when building volcano monitoring systems, as results will be strongly influenced by a human factor, including, but not limited to, lack of unified criteria, labelling fatigue, and time needed to analyze continuous streams of data. [#bishop2007pattern]. The vast amount of seismic data registered during an eruption requires robust and reliable systems able to operate in real time and tackle the mentioned drawbacks. The main objective of this thesis is to propose a deep learning volcano-seismic recognition system that can exploit the information contained within recorded seismic signals and improve generalization capabilities during real eruptive scenarios. Inspired on recent advances in the field of neural networks and volcano-seismology [#carniel2006user, #boue2015real, #ham2012neurocomputing], its main contributions are: 1. Development of an automatic recognition system for isolated volcano-seismic events: The monitoring of active volcanoes generates huge amounts of data that vulcanological observatories can hardly manage in the short term. Often, each recording is analyzed by geophysicists based on their experience and knowledge of a particular volcano. As a result, current datasets are composed by most relevant seismic signals that can be encountered in each volcano. Seismic observatories analyze the rate, type and location of seismic events. However, this wealth of data requires high recognition rates at a lower computational cost, with enough reliability and robustness to support the study. The approach followed in this work is to parse raw waveforms into a set of descriptive features using signal processing algorithms that can be further used in geological modelling frameworks. Deep learning could encode geophysical knowledge through hierarchical features that can enhance current early warning systems at scale. 2. Development of a continuous recognition system: The classification of isolated events, whilst useful to classify at scale, is still insufficient to manage eruptive crisis and issue early warnings in real time. During eruptive crisis, seismic data is registered as a continuous stream. Unlike isolated and segmented signals, these continuous seismic registers are temporal sequences with an indeterminate number of concatenated events. In this sense, the detection and classification of volcano-seismic events from real-time seismic data is a sequential problem which involves a complex and high dimensional dynamic signals which require efficient models able to capture the long temporal dependencies of seismic data. According to the requirements of each system, several artificial neural networks architectures and several data parameterization schemes have been proposed. For the isolated recognition systems, we have found that deep architectures based on pre-training initialization and multiple processing layers to learn abstract representations of data could be a particularly effective strategy to increase the overall generalization capabilities of the systems. In this sense, we tested two different DNNs, DBN (Deep Belief Networks) and SDA (Stacked Denoising AutoEncoder) with seismic data recorded at “Volcán de Fuego”, Colima (Mexico). In the case of continuous recognition system, we find that RNNs (Recurrent Neural Networks) can be applied as statistical models able to exploit temporal information. We tested three recurrent architectures with seismic data from Deception Island (Antarctica) over different seismic periods: vanilla-RNN, LSTM and GRU. In order to explore their generalization capabilities with data recorded in different time periods we further tested the models with data from a recent seismic survey in 2017. The classification results obtained in both approaches outperform the recognition rate obtained by classical architectures as Support Vector Machine, Random Forest, Hidden Markov Models and Gaussian Mixture Models. We find that sDA and DBN can classify seismic events with higher precision and recall than classical architectures. Moreover, deep architectures are more sensitive to detect events that occur simultaneously in time, such as explosions and tremors. With regards to continuous recognition systems, attained results have shown that vanilla-RNN, LSTM and GRU classify volcano-seismic events with good accuracy, and memory cells (LSTM and GRU) enhance the detection of long-term signals. In conclusion, classifiers based on deep neural networks can be deployed in real-environments to monitor the seismicity of restless volcanoes, and enhance current early warning systems. However, given the nature and size of volcanic snapshots (dataset), the use of raw volcanic events as training data results on non useful representations, and therefore, a direct application of state-of-the-art deep learning architectures is still a challenge. Structure The rest of this work is organized into three large blocks as follows: • First, Part 1 serves as an introduction to the classification of volcano-seismic signals and describes the theoretical framework of the proposed architectures. Chapter 1 provides the fundamental concepts of volcano-seismology and active volcano monitoring. Chapter 2 addresses the development of a recognition system based on supervised learning and introduces the related research in the field. Finally, Chapters 3 and 4 provide the theoretical background of deep neural networks architectures (DNNs and RNNs), describing and identifying the main drawbacks when deploying these architectures. • Part 2 describes the experimental methodology followed in this work, with detailed results. Chapter 5 stands as a small introduction about the volcanic signals that compose our data sets. Chapter 6 motivates the need to parameterize the data, and how to exploit the capabilities of deep learning architectures in order to extract features from our datasets. Taking into account that that systems performance is highly dependent on how accurately parameters of the model can be estimated, we propose two different parameterization schemes. Chapter 7 proposes a novel approach in the field of volcano seismology to classify volcano-seismic events based on fully-connected Deep Neural Networks (DNNs). Two DNN architectures with different weights initialization are studied: stacked Denoising Autoencoders (SDA) and Deep Belief Networks (DBN). Using a combined feature vector of Linear Prediction Coefficients (LPC) and statistical properties, we evaluate classification performance on seven different classes of isolated seismic events. The results obtained are compared to well-established techniques as Multilayer Perceptron (MLP), Support Vector Machine (SVM) Random Forest (RF), Gaussian Mixture Model (GMM) and Hidden Markov Model (HMM). Following this methodology, Chapter 8 proposes RNNs ( (LSTM, GRU and Vanilla) for detection and classification of continuous sequences of volcano-seismic events at Deception Island Volcano, Antarctica. The best configuration trained with data from seismic records obtained from 1995 to 2002, will be tested with data from recent seismic survey performed at Deception Island Volcano in 2017 by the XXX Spanish Antarctic scientific expedition. This experiment explores how RNNs can perform continuous monitoring of volcanic-activity when terrain and seismic sources changes. • Finally Part 3, discusses the interpretations of the final results and motivates future research lines in order to enhance effectiveness of the proposed systems.