PREDICTION OF GROUND SUBSIDENCE DUE TO LIQUEFACTION BY APPLYING TRANSFER LEARNING TECHNIQUE

Abstract

 In this study, we aimed to construct a predictive model consistent with actual ground subsidence caused by liquefaction. Based on the analysis results from the Nankai Trough Great Earthquake Model Study Group, transfer learning was applied using vertical differences in the DEM (Digital Elevation Model) of Urayasu City, Chiba Prefecture, before and after the Great East Japan Earthquake as the actual subsidence amounts due to liquefaction. A neural network regression model was employed as the algorithm for this purpose. As a result, transfer learning using the pre-trained model improved the overall prediction accuracy of liquefaction-induced ground subsidence. This indicates that a realistic predictive model can be constructed by applying transfer learning based on actual subsidence data obtained from airborne laser surveys.