Performance Evaluation of GIS-Based Novel Ensemble Approaches for Land Subsidence Susceptibility Mapping

Land subsidence (LS) is difficult to predict due to limitations in monitoring, proper surveys and knowledge related to functioning and behavior of LS. However, the lack of susceptibility maps to identify areas prone to LS results in severe economic and human losses. LS susceptibility mapping (LSSM) can help to prevent natural and human catastrophes and reduce the economic damages. Machine learning (ML) techniques are becoming increasingly proficient at modelling such occurrences and they are increasing used for LSSM. This study compares the performances of single and hybrid ML models to predict and map susceptible areas to land subsidence. For the spatial prediction of land subsidence, in this study, land subsidence susceptibility (LSS) was assessed using four machine learning models; maximum entropy (MaxEnt), general linear model (GLM), artificial neural network (ANN) and support vector machine (SVM) and the possible number of novel ensemble models integrated through the mentioned four machine learning algorithms (MLAs). The new eleven ensemble models are created integrating MLAs. Previous occurrences of land subsidence areas were mapped and using as the training dataset (70%) and validating dataset (30%) in the modelling process. Each applied model produced a land subsidence susceptibility map (LSSM) and we got a total of fifteen LSSMs for this study area. To identify the robust model and best LSSMs, area under the receiver operating characteristic (AUROC) curve was employed. The AUROC result indicated that ANN model had the highest AUROC (0.924) prediction accuracy. The highest AUC (0.823) of the LSSMs was determined based on validation datasets identified by SVM followed by ANN-SVM (0.812).