Landslide Susceptibility Mapping in North Tehran, Iran: Linear Regression, Neural Networks, and Fuzzy Logic Approaches

This study applies automatic linear regression (LINEAR), artificial neural networks (ANN), and fuzzy logic methods for landslide susceptibility mapping (LSM) based on controlling factors in the north Tehran region. We compiled a dataset of 225 landslides, including location, landslide area (LA, km2), and various controlling factors (e.g., lithology, topography, seismicity, hydrology, vegetation cover, and human activity). The findings showed that there are non-linear relationships between most of the controlling factors and the landslide area percentage (LA in factor classes/total LA). Therefore, we used diverse fuzzy membership functions (e.g., the Gaussian, MS-large, and MS-small) to enhance the consistency between the fuzzified layers and the actual conditions. The validation of Quality Sum (Qs) revealed the fuzzy gammas (0.75-0.95), ANN model, fuzzy AND, and fuzzy PRODUCT have the best prediction currency for LSM, respectively. Whereas, the fuzzy SUM and OR operators unrealistically classified 95-90% of the region in the very high to high category for the LSMs. Compared to the linear model, the neural network model had a more accurate prediction of LA (km2) and better performance for LSM. The reason for this is the high flexibility of the ANN for modeling the nonlinear relationships between independent parameters and the target. As a result, the most important factors affecting the LA are topography (the elevation, slope aspect, and slope), seismic activity (the distance from faults and acceleration of critical [Ac]), and lithology (factor of safety [Fs]).