QuakeFinder's Lessons and Challenges Using Big Data Searching for Earthquake Precursors
Abstract
QuakeFinder (QF) is a humanitarian R&D project attempting to detect anomalous electromagnetic activity occurring ahead of earthquakes. QF has acquired over 70 TB of data from nearly 14 years of observation of Earth's magnetic field with outstanding spatial and temporal resolution. Their observatory network consists of 150 stations along the major faults in California, Greece, Taiwan, Peru, Chile and Indonesia (Warden et al. 2018). Each station is equipped with 3 feedback induction magnetometers, 2 ion sensors, a geophone, a temperature sensor, and a humidity sensor. The data are continuously recorded at 50 samples per second with GPS antennas supplying reference timestamps and transmitted daily to the QF data center. QF has developed an algorithmic framework to support an automated approach for finding earthquake precursory statistical significance (initial results from this framework are currently in review).
There are many challenges applying machine learning to a dataset of this type. First, since large earthquakes are relatively rare, the dataset will contain many times more negative examples than positive ones which makes it critical to ensure the algorithm is not overfitting. Additionally, the earthquake truth data (obtained from earthquake seismic catalogs) isn't always reliable. Earthquakes are admitted as positive cases depending on their magnitude, depth, and distance from a station. For a particular earthquake, these values can vary across different catalogs and even within a single catalog the values might be updated or refined at any time. This can potentially affect the set of earthquakes included in a given analysis. These issues only compound the overall challenge of detecting faint, hypothesized precursory signals amongst the cultural and natural noise in the dataset. Since the source of these signals is not well understood and competing models have been proposed, this study provides a good opportunity to apply a data-driven approach to inform the development of physical models.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFMNH41B0926S
- Keywords:
-
- 1930 Data and information governance;
- INFORMATICS;
- 4315 Monitoring;
- forecasting;
- prediction;
- NATURAL HAZARDS;
- 4316 Physical modeling;
- NATURAL HAZARDS;
- 4339 Disaster mitigation;
- NATURAL HAZARDS