A GIS-based neuro-fuzzy procedure for integrating knowledge and data in landslide susceptibility mapping

A significant portion of the Mazandaran Province in Iran is prone to landslides due to climatic conditions, excessive rain, geology, and geomorphologic characteristics. These landslides cause damage to property and pose a threat to human lives. Numerous solutions have been proposed to assess landslide susceptibility over regions such as this one. This study proposes an indirect assessment strategy that shares in the advantages of quantitative and qualitative assessment methods. It employs a fuzzy inference system (FIS) to model expert knowledge, and an artificial neural network (ANN) to identify non-linear behavior and generalize historical data to the entire region. The results of the FIS are averaged with the intensity values of existing landslides, and then used as outputs to train the ANN. The input patterns include both physical landscape characteristics (criterion maps) and landslide inventory maps. The ANN is trained with a modified back-propagation algorithm. As part of this study, the strategy is implemented as a GIS extension using ArcGIS(R). This tool was used to create a four-domain landslide susceptibility map of the Mazandaran province. The overall accuracy of the LSM is estimated at 90.5%.