Geospatial liquefaction probability models based on sand boils occurred during the 2017 M5.5 Pohang, South Korea, earthquake

Abstract

The 2017 Pohang earthquake, triggered by an enhanced geothermal system, had a moment magnitude (M) of 5.5 and caused liquefaction in Pohang City, South Korea. Notably, approximately 600 sand boils were observed only in the soft Quaternary alluvial sediments. This contrasts with most historic liquefaction cases that occurred in earthquakes with M ≥ 6.0. In addition, ground and building settlements were reported in Pohang City. To assess the liquefaction hazard for the Pohang earthquake within the city, we define a 14.4 km2 area where the majority of liquefaction was observed. Using the liquefaction data, we develop eight geospatial liquefaction probability models using logistic regression (LR) and random forest (RF) algorithms (five LR- and three RF-based models). These models consider geospatial data (peak ground acceleration, peak ground velocity, slope-derived average shear wave velocity of the upper 30 m, compound topographic index, distance to the nearest water, and roughness) and geotechnical data (averaged standard penetration test N values up to 20 m depth, depth to rock). We evaluate the performances of our eight models and existing global and regional geospatial models using four indicators: accuracy, balanced accuracy, F1 score, and area under the receiver operating characteristic curve. The three RF-based models outperform the other models, followed by the best LR-based model and the global geospatial model from Zhu et al. (2015). Finally, we apply the best RF-based model to the entire Pohang City and evaluate its applicability to other regions and earthquakes. The estimated liquefaction probabilities agree well with the observed distributions of the sand boils and settlements. These geospatial liquefaction models provide a valuable tool for continuously mapping liquefaction hazards in areas with similar geological and geotechnical conditions to Pohang.