Comparative Analysis of K-Means, K-Medoids, and DBSCAN for Spatial Clustering of Positive Cloud-to-Ground Lightning in Tangerang City

Abstract

Cloud-to-Ground (CG) lightning is one of the most destructive atmospheric phenomena due to its potential to damage electrical infrastructure, communication networks, and public facilities. Understanding its spatial distribution is essential for disaster mitigation, infrastructure protection, and urban planning. This study compares the performance of the K-Means, K-Medoids, and Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithms in identifying the spatial distribution of Positive Cloud-to-Ground (CG+) lightning strikes in Tangerang City, Indonesia. The analysis utilizes CG+ lightning observations collected by the Tangerang Geophysical Station of the Indonesian Agency for Meteorology, Climatology, and Geophysics (BMKG) during 2025. Prior to clustering, the dataset underwent data cleaning, CG+ event selection, administrative boundary filtering, feature selection, coordinate transformation from WGS84 to UTM, and Min–Max normalization. The clustering performance of each algorithm was assessed using the Silhouette Coefficient and Davies–Bouldin Index (DBI). The evaluation results indicate that both K-Means and K-Medoids achieved the best performance with K = 3. K-Means obtained the highest Silhouette Coefficient (0.445640) and the lowest DBI (0.789620), while K-Medoids produced comparable results with a Silhouette Coefficient of 0.444728 and a DBI of 0.794027. For DBSCAN, the optimal parameter combination (ε = 200 m and MinPts = 6) generated 64 clusters and 170 noise observations, yielding a Silhouette Coefficient of 0.059880 and a DBI of 1.452474. These findings demonstrate that partition-based clustering approaches provide a more reliable representation of the spatial distribution of CG+ lightning in Tangerang City than the density-based DBSCAN approach.