Event classification and quality assessment for local seismic events using machine learning

Automatic monitoring of local seismicity produces events of varying quality. Some events will be poorly located, and some event solutions will not represent a real seismic event, arising only due to noise. Noise events are removed and location quality improved during manual revision, but that is not always feasible for large catalogs. In cases with >10000 events manual review is not tenable, so an automatic quality score calculation is beneficial in improving catalog quality.

Machine learning methods are a useful tool for this purpose, both for classification and calculating a quality score. We explore machine learning methods for solutions to this problem, with a special focus on feature extraction from travel-time information.

The models are evaluated on data from three seismic networks in Iceland. The dataset contains both automatic and manual solutions for a large number of earthquakes, so direct comparisons between manual and automatic solutions can be made. The manual locations can then be used as a ground truth solution that the automatic solutions attempt to approximate. The models are ranked on their classification score as well as their ability to estimate the spatial distance from the ground truth solution.