Name
Improved earthquake clustering using a density-normalized DBSCAN algorithm: examples from Iran and eastern Canada
Description
The DBSCAN algorithm is a popular unsupervised machine learning technique for clustering and spatial data analysis. However, its accuracy is contingent upon the proper selection of two hyperparameters, MinPts and ϵ. In our study, we improve the algorithm by using an event density map to dynamically calculate MinPts, while calculating ϵ from the size of the density map cells. This modification reduces the number of hyperparameters from two to one, making optimization faster and simpler.
The variable MinPts results in improved accuracy, especially in regions with varying earthquake density. We tested the modified DBSCAN algorithm using the Iranian earthquake catalog and found results more consistent with the Mirzaei et al. 1998 seismotectonic model than those obtained using conventional DBSCAN. Additionally, when applied to the Eastern Canada earthquake catalog, the results are promising.
The variable MinPts results in improved accuracy, especially in regions with varying earthquake density. We tested the modified DBSCAN algorithm using the Iranian earthquake catalog and found results more consistent with the Mirzaei et al. 1998 seismotectonic model than those obtained using conventional DBSCAN. Additionally, when applied to the Eastern Canada earthquake catalog, the results are promising.