Recognizing Spatial Seismicity Patterns in Earthquake Data Using Unsupervised Machine Learning Techniques

Abstract. In this study an unsupervised pattern recognition approach was used to identify spatial seismicity patterns in areas around the Kazerun and Kare- Bas fault systems, located in southern Iran. For this purpose, 5,546 earthquake events (with M ≥ 2.5) recorded between 2006 and February 2024 were extracted from the Iranian Seismological Center catalog. Then, the study region, extending from 50.5°–52.5°E and 28°–31°N, was divided into a 0.2° × 0.2° grid net, resulting in 225 cells, of which 182 contained earthquake data. After that, for each cell, three quantitative seismicity parameters were computed: (1) number of earthquakes, (2) maximum event magnitude, and (3) mean focal depth of earthquakes. The k-means clustering technique was then applied to the obtained dataset using Euclidean distance as the similarity metric, and also with the number of clusters (k) varied between 3 and 5 to generate different seismicity zoning maps. The seismicity pattern maps, obtained in this study, revealed spatially coherent zones corresponding to areas with distinct seismicity behavior, reflecting variations in both frequency and focal depth of earthquakes. The results obtained demonstrate that the k-means clustering technique, as an unsupervised learning technique, can effectively distinguish spatially significant seismicity zones and that this technique can provide a quantitative basis for data-driven seismicity zoning of active regions. In addition, this approach can be easily extended to other tectonically and seismically active regions or refined using higher-resolution grid cells and additional seismicity parameters and variables to improve pattern resolution and predictive capability.