Automatic Earthquake Declustering Using the Nearest‐Neighbor Distance
Abstract
In the widely adopted description of seismic occurrence, earthquakes are categorized as either background or triggered events. In this work, we present a fully automated, non-parametric algorithm for distinguishing between these two categories, a process known as seismic declustering, based on the widely used nearest-neighbor (NN) metric. We introduce a new measure, the susceptibility index, which identifies an optimal threshold to discriminate between background and triggered events within the NN metric. Through statistical testing on simulated epidemic type aftershock sequence catalogs, we demonstrate that our method yields classification metrics exceeding 90%, outperforming state-of-the art algorithms. Notably, we show that a single threshold is sufficient for reliable discrimination within a given data set. The identification of this threshold requires memory capacity and computational time that scale linearly and quadratically with the data set size, respectively, making the method particurarly suited for large earthquake catalogs. We also apply our method to the relocated Southern California catalog and the GeoNet catalog of New Zealand (NZ). Our method effectively adapts across the different tectonic settings, capturing the variability of background seismicity rates between the shallow crustal events of Southern California and the tectonically diverse seismicity of NZ.
Plain Language Summary Earthquakes are typically categorized into two types: background events, which occur randomly in time, and triggered events, which are caused by earlier earthquakes. Separating these two types is important for estimating the expected number of earthquakes in the future decades. In this study, we introduce a new method that automatically separates background earthquakes from triggered ones using a metric that combines time, space and magnitude. By testing this method on simulated earthquake data sets, we show that it can classify earthquakes with over 90% accuracy. Our approach provides an optimal threshold to make this distinction and is computationally scalable, making it suitable for use on large earthquake data sets. We applied our method to real earthquake data from Southern California and New Zealand, and obtained reasonable estimates of the background rate.
Related
- Is the Number of Foreshocks Underestimated? The Role of Declustering Symmetry
- Graph-based probabilistic earthquake clustering and forecasting in Sumatra
- Predicting rate of penetration (ROP) with machine learning models: A physics-validated comparative case study of Gradient Boosting Regressor (GBR), Random Forest (RF), and K-Nearest Neighbours (KNN)
- Sustained crustal helium enrichment associated with postseismic viscoelastic relaxation following the 2025 Myanmar earthquake
- Lithium and strontium isotope systematics of hot spring waters from the Noto Peninsula: Constraints on the origin of geofluids associated with the 2024 M7.6 earthquake
BibTeX
@article{bountzis2026automatic,
title = {{Automatic Earthquake Declustering Using the Nearest‐Neighbor Distance}},
author = {P. Bountzis and E. Lippiello and S. Baccari and G. Petrillo},
journal = {Earth and Space Science},
year = {2026},
month = oct,
doi = {10.1029/2025ea004539},
url = {https://hal.science/hal-05773436v1},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Automatic Earthquake Declustering Using the Nearest‐Neighbor Distance
AU - P. Bountzis
AU - E. Lippiello
AU - S. Baccari
AU - G. Petrillo
PY - 2026
DA - 2026/10/01
JO - Earth and Space Science
DO - 10.1029/2025ea004539
UR - https://hal.science/hal-05773436v1
AB - In the widely adopted description of seismic occurrence, earthquakes are categorized as either background or triggered events. In this work, we present a fully automated, non-parametric algorithm for distinguishing between these two categories, a process known as seismic declustering, based on the widely used nearest-neighbor (NN) metric. We introduce a new measure, the susceptibility index, which identifies an optimal threshold to discriminate between background and triggered events within the NN metric. Through statistical testing on simulated epidemic type aftershock sequence catalogs, we demonstrate that our method yields classification metrics exceeding 90%, outperforming state-of-the art algorithms. Notably, we show that a single threshold is sufficient for reliable discrimination within a given data set. The identification of this threshold requires memory capacity and computational time that scale linearly and quadratically with the data set size, respectively, making the method particurarly suited for large earthquake catalogs. We also apply our method to the relocated Southern California catalog and the GeoNet catalog of New Zealand (NZ). Our method effectively adapts across the different tectonic settings, capturing the variability of background seismicity rates between the shallow crustal events of Southern California and the tectonically diverse seismicity of NZ. Plain Language Summary Earthquakes are typically categorized into two types: background events, which occur randomly in time, and triggered events, which are caused by earlier earthquakes. Separating these two types is important for estimating the expected number of earthquakes in the future decades. In this study, we introduce a new method that automatically separates background earthquakes from triggered ones using a metric that combines time, space and magnitude. By testing this method on simulated earthquake data sets, we show that it can classify earthquakes with over 90% accuracy. Our approach provides an optimal threshold to make this distinction and is computationally scalable, making it suitable for use on large earthquake data sets. We applied our method to real earthquake data from Southern California and New Zealand, and obtained reasonable estimates of the background rate.
SN - 2333-5084
SN - 2333-5084
ER -