A new GIS-based data mining technique using an adaptive neuro-fuzzy inference system (ANFIS) and k-fold cross-validation approach for land subsidence susceptibility mapping

TLDR: In this article, an adaptive neuro-fuzzy inference system (ANFIS) was used for land subsidence susceptibility mapping (LSSM) in the Marand plain, northwest Iran.

Abstract: In this paper, we evaluate the predictive performance of an adaptive neuro-fuzzy inference system (ANFIS) using six different membership functions (MF). In combination with a geographic information system (GIS), ANFIS was used for land subsidence susceptibility mapping (LSSM) in the Marand plain, northwest Iran. This area is prone to droughts and low groundwater levels and subsequent land subsidence damages. Therefore, a land subsidence inventory database was created from an extensive field survey. Areas of land subsidence or areas showing initial signs of subsidence were used for training, while onethird of inventory database were reserved for testing and validation. The inventory database randomly divided into three different folds of the same size. One of the folds was chosen for testing and validation. Other two folds was used for training. This process repeated for every fold in the inventory dataset. Thereafter, land subsidence related factors, such as hydrological and topographical factors, were prepared as GIS layers. Areas susceptible to land subsidence were then analyzed using the ANFIS approach, and land subsidence susceptibility maps were created, whereby six different MFs were applied. Lastly, the results derived from each MF were validated with those areas of the land subsidence database that were not used for training. Receiver operating characteristics (ROC) curves were drawn for all LSSMs, and the areas under the curves were calculated. The ROC analyses for the six LSSMs yielded very high prediction values for two out of the six methods, namely the difference of DsigMF (0.958) and GaussMF (0.951). The integration of ANFIS and GIS generally led to high LSSM prediction accuracies. This study demonstrated that the choice of training dataset and the MF significantly affects the results.