Cluster analysis of earthquake hypocenters in Azerbaijan and surrounding territories

TL;DR

This study analyzes Azerbaijan and surrounding seismicity (2010–2023) to identify spatial clusters of earthquake hypocenters. It combines a Morisita Index screening with density-based clustering (DBSCAN and HDBSCAN) and uses Adjusted Rand Index to compare partitions, selecting 4- and 6-cluster solutions that best reflect the regional fault network. For each cluster, depth distributions are modeled with two-component Gaussian mixtures, revealing bimodality in several coastal clusters and providing insights into lithospheric structure. The framework offers a data-driven approach to characterize regional seismic clustering and depth architecture, with practical implications for hazard assessment and faulting studies; the authors also provide data and code resources.

Abstract

The research focuses on seismic events that occurred in Azerbaijan and adjacent territories, regions known for strong seismic activity. We analyze a catalog of recorded earthquakes between 2010 and 2023, extracting the locations of the earthquake hypocenters for study purposes. Using statistical methods and cluster analysis tools, we developed a procedure for partitioning hypocenter clusters. The procedure begins with estimates of the Morisita Index, which is suitable for preliminary assessments of the statistical properties of hypocenter sets. Analysis of the Morisita Index indicates that the spatial distribution of hypocenters is heterogeneous, containing denser domains referred to as clusters. The next stage involves identifying spatial clusters using the DBSCAN and HDBSCAN algorithms. Due to the strong dependence of results on the algorithm's parameters, we selected several partitions with 5-8 clusters that provided maximal or near-maximal Silhouette Index values. The final stage assesses the similarity of the resulting partitions, using the Adjusted Rand Index to identify partitions with a specified degree of similarity. The final set of partitions was compared to the fault network of the region. Based on the selected partition, the earthquake depth distributions were studied. Specifically, approximate probability density functions were constructed in the form of mixtures of normal distributions, leading to the identification of several bimodal distributions.

Figures (8)

• Figure 1: a: Distribution of earthquake epicenters across Azerbaijan and surrounding regions. Earthquake depth is represented by color according to the accompanying colorbar. b: Normalized locations of the earthquake epicenters corresponding to panel (b).
• Figure 2: Morisita Index $I_\delta$ vs. $\delta$, represented in a logarithmic scale and evaluated for 2D (black dashed) and 3D (red solid) data, respectively.
• Figure 3: a: Silhouette Index vs. number of clusters. b: Two-parameter diagram showing the distribution of cluster counts. c and d: The 4- and 6-cluster partition marked in the Silhouette diagram by the labels "db 159" and "db 192", respectively. The noise points are removed. Filled bullets mark strong earthquakes with $M>5.5$. The results obtained via DBSCAN.
• Figure 4: a: Silhouette Index vs. number of clusters. b, c, and d: The 4- and 6-cluster partitions with noise points removed, marked in the Silhouette diagram by labels "hdb 7", "hdb 5", and "hdb 15", respectively. Filled bullets mark strong earthquakes with $M>5.5$. The results obtained via HDBSCAN.
• Figure 5: Adjusted Rand Index evaluation for 4-cluster (a) and 6-cluster partitions (b). The axis labels correspond to the selected partitions in Figs. \ref{['fig:3']}a and \ref{['fig:4']}a, numbered in descending order of SI. The labels with "db" and "hdb" stand for the results obtained via DBSCAN and HDBSCAN, respectively.