A GIS-based comparative study of hybrid fuzzy-gene expression programming and hybrid fuzzy-artificial neural network for land subsidence susceptibility modeling

Land subsidence is a complicated hazard that artificial intelligence models can model it without approximation and simplification. In this study, for the first time in land subsidence studies, we used and compared the accuracy and efficiency of hybrid fuzzy-gene expression programming (F-GEP) and fuzzy-artificial neural network (F-ANN) models in estimating land subsidence susceptibility modeling in Varamin aquifer of Iran. For this purpose, after selecting and gathering information from fifteen geo-environmental and hydrogeological effectual factors including specific yield, erosion, aquifer thickness, distance of fault, bedrock level, digital elevation model (DEM), annual rainfall, clay thickness, transmissivity (T), soil type, Debi zonation of pumping wells, slope based on DEM, groundwater drawdown in 20 years, land use, and lithological units event based on literature review in the GIS environment, they were first standardized with GIS fuzzy membership functions, and then GEP model was used to integrate the layers. For this step, using 70% of the data (2919 pixels) for the train and 30% (1251 pixels) for the test. Finally, using several statistical criteria and radar image data, the models were validated. We repeat the model on ANN, and our results showed that F-GEP model (with R2 = 0.99 and RMSE = 0.004) is more accurate than F-ANN model (with R2 = 0.94 and RMSE = 0.056) for land subsidence susceptibility modeling in the study area.