← Back to Sessions

Toward Unbiased Volcano-Seismic Monitoring: Leveraging Weakly Supervised Learning for Comprehensive Insights

Description: 

Real-time monitoring of volcano-seismic signals presents a multifaceted challenge. Typically, automated systems are constructed through the assimilation of vast seismic catalogues, wherein each entry is accompanied by a label denoting its source mechanism. However, the creation of comprehensive catalogues is hindered by the prohibitive cost associated with data labeling. Although current machine learning techniques have demonstrated considerable success in crafting predictive monitoring tools, reliance on catalogue-based learning introduces potential bias to the system.

In our study, we demonstrate that while monitoring systems successfully identify nearly 90% of events annotated in seismic catalogues, they may overlook additional information crucial for understanding volcanic behavior. We discovered that weakly supervised learning approaches exhibit exceptional capabilities by simultaneously identifying unannotated seismic traces in the catalogue and rectifying mislabeled entries. By incorporating a system trained with a master dataset and catalogue as a pseudo-labeller within the framework of weakly supervised learning, we unveil and update information pertaining to volcanic dynamics.

Our findings suggest the potential for advancing more sophisticated semi-supervised models to enhance the reliability of monitoring tools. For instance, the exploration of advanced pseudo-labelling techniques involving data from multiple catalogues could be pursued. Ultimately, this research paves the way for the development of universal monitoring tools capable of accommodating unforeseen temporal changes in monitored signals across various volcanic settings.

Session: Multidisciplinary Approaches for Volcanic Eruption Forecasting - I

Type: Oral

Date: 5/2/2024

Presentation Time: 02:30 PM (local time)

Presenting Author: Manuel

Student Presenter: No

Invited Presentation: 

Authors