Detection of Aseismic Geodetic Transients Using k-Nearest Neighbors

Observation and detection of aseismic transient signals in geodetic data is an important part of improving understanding of the earthquake cycle. Recent work, conducted as part of the Southern California Earthquake Center's transient detection exercise, showed that transient detection results were improved by using multiple detection methods and by localizing results to specified geographical regions. While geographical regions can be defined on the basis of a priori science domain knowledge, regions defined in such a way are vulnerable to edge effects (i.e., observations collected near the region boundary are less reliable) and are not easily extendable to new regions or measurement modalities where a priori understanding is limited.

We present an approach to overcoming these challenges that utilizes a k-nearest neighbor approach to defining geographical regions for GPS stations. In this approach, each station has its own individual region, defined by a chosen number k closest neighboring stations rather than any predefined geographical boundary. In this general framework, a weighting function controls the contributions of neighboring stations to the transient detection, while a distance function (e.g., Vincenty's method) controls which k stations are considered neighbors. The "proper" k depends both on the transient detection method and the weighting function; it can be defined by the user directly or optimized using hyperparameter optimization. For transient detection using multiple detection methods, this framework nests readily within an ensemble classifier.