Use of “Crowd-Sourcing” and other collaborations to solve the short-term, earthquake forecasting problem

Abstract:

QuakeFinder (QF) and its international collaborators have installed and currently maintain 165 three-axis induction magnetometer instrument sites in California, Peru, Taiwan, Greece, Chile and Sumatra. The data from these instruments are being analyzed for pre-quake signatures. This analysis consists of both private research by QuakeFinder, and institutional collaborators (PUCP in Peru, NCU in Taiwan, PUCC in Chile, NOA in Greece, Syiah Kuala University in Indonesia, LASP at U of Colo., Stanford, and USGS).

Recently, NASA Hq and QuakeFinder tried a new approach to help with the analysis of this huge (50+TB) data archive. A collaboration with Apirio/TopCoder, Harvard University, Amazon, QuakeFinder, and NASA Hq. resulted in an open algorithm development contest called “Quest for Quakes” in which contestants (freelance algorithm developers) attempted to identify quakes from a subset of the QuakeFinder data (3TB). The contest included a $25K prize pool, and contained 100 cases where earthquakes (and null sets) included data from up to 5 remote sites, near and far from quakes greater than M4. These data sets were made available through Amazon.com to hundreds of contestants over a two week contest period.

In a more traditional approach, several new algorithms were tried by actively sharing the QF data with universities over a longer period. These algorithms included Principal Component Analysis-PCA and deep neural networks in an effort to automatically identify earthquake signals within typical, noise-filled environments.

This presentation examines the pros and cons of employing these two approaches, from both logistical and scientific perspectives.