Unsupervised probabilistic machine learning applied to seismicity
declustering: a new approach to represent earthquake catalogues with
fewer assumptions
Abstract
Many applications in seismology require to isolate earthquake clusters
from a background activity. Relative declustering methods essentially
find a 2D representation of an earthquake catalogue that distinguishes
between two classes of events: crisis and non-crisis events. However,
the number of statistical and/or physical parameters to be used is often
limited due to the difficulty of concatenating the information onto a
physically meaningful 2D grid. In this study, we propose to alleviate
the declustering task by using the ability of unsupervised artificial
intelligence to model complex spatio-temporal relationships directly
from data. Through a data-driven approach, we define an easily
transferable declustering model that provides declustering results with
fewer assumptions and no prior selection of thresholds. We first obtain
this model by training a self-organising neural network (SOM) that
learns to cluster data points according to their feature similarity on a
2D map. We then assign each SOM cluster a label (crisis or non-crisis
class) using an agglomerative clustering procedure. We quantify the
classification uncertainty by developing a probabilistic function based
on the projection learned by SOM. Our method is applied to a synthetic
dataset and to real catalogues from the Gulf of Corinth, Central Italy
and Taiwan. We discuss the validity of the method by estimating its
classification accuracy. For real data, we qualitatively compare our
results to previous declustering attempts. We show that our approach is
easy to handle, provides a fairly new representation of earthquake
catalogues and has the potential to reduce classification ambiguities
between nearby events.