Assessing the importance of variable selection in land subsidence susceptibility mapping

Abstract

In the domain of land subsidence risk science, Land Subsidence Susceptibility Mapping (LSSM) aids in spatially identifying regions prone to subsidence. This study used a multi-collinearity analysis through the variance inflation system (VIF) and tolerance (TOL) and a machine learning (ML) model of Random Forest (RF) for LSSM in the Damaneh Plain, Isfahan Province, Iran. The study investigated the importance of the conditioning variables in predicting Land Subsidence occurrences using an ML model. An ML model's prediction capabilities and performance were evaluated using conditioning variables in this paper. Using VIF, we eliminated the least "important" variables related to the LSSM. Conclusively, we found that removing the least "important" variables improves the accuracy of the resulting LSSMs. Based on the results of our study, using VIF could increase the predictive performance of the RF model by three percentage points in the applied accuracy assessment metric.