• Citer

On the use of the machine learning algorithms for automatically classifying cross-correlation functions: Application to the seismic data obtained around the Vatnajökull icecap, Iceland

par Forestier, Justine
Président du jury Mattielli, Nadine ;Debaille, Vinciane
Promoteur Lecocq, Thomas
Co-Promoteur Nowé, Sylvain
Publication Non publié, 2022-08-29

Mémoire

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

In Iceland the sources of seismic noise are numerous and diverse (glacial, volcanic, hydraulic, oceanic ...) and thus are complex to identify and classify. Yet, these sources can be associated with natural disasters such as floods. Their monitoring could therefore prevent these hazardous events. These are the objectives of the IS-noise and IS-tremor projects, in parallel of which this work takes place.Interferometry of the ambient seismic noise is used to generate cross-correlation functions. In these functions, trends in the amplitudes of the correlations over time are observed. These trends have been interpreted by Nowé et al. (2021) as different sources of noise. One is seasonal, occurring each year in spring and ceasing just before winter. It is assumed to be related to glacier melting, supplying rivers and waterfalls. Its detection in the cross-correlation functions is restricted by the large amount of data, difficult to be processed by a human operator. Machine learning, a branch of artificial intelligence, is a technique increasingly used in geosciences that can overcome this difficulty. This work proposes a workflow involving several machine learning algorithms to discriminate this seasonal seismic noise source from other noise sources appearing in the cross-correlation functions automatically and quickly. It consists of data exploration using a hierarchical clustering algorithm, features extraction via the algorithm of Hibert et al. (2017), and automatic classification with a Random Forest classification algorithm. A secondary objective to this work is to examine the features and information extracted during these steps to gain a better understanding and characterization of the seasonal ambient noise.This work shows that machine learning algorithms can identify the difference between seismic noise sources in cross-correlation functions and proposes points of reflection for further exploration of this data. This workflow can be used on other data sets to discriminate noise sources. The presence of a potential precursor sign preceding the arrival of the seasonal noise signature discovered in this work shows that it can be used as a basis for the implementation of a natural disaster monitoring tool.