Planning a single low risk forest road based on artificial neural network model of landslide susceptibility (case study: Kojour watershed)

This research aimed at modeling the landslide susceptibility using the Artificial Neural Networks (ANN) in Kojour watershed and then planning the forest road based on the resulted map in Aghoozchal and Moor districts in the northern Iran. After recording the coordinates of 95 occurred landslides, six factors of terrain slope, aspect, curvature, distance to river, distance to fault, and geology formation were assumed as the effective factors in landslide occurrence. The digital layers of input variables were prepared in Geographic Information System. After data extraction, various backpropagation multi-layer perceptron ANNs with various setting parameters were developed and their performances were evaluated. Results showed that the best root mean square error (RMSE) and determination coefficient (R2) as model performance criteria for the most robust model were 0.1945 and 0.8676, respectively, in which 2 and 8 neurons have been obtained in the first and second hidden layers. Among the proposed variants, the variant No. 3 was selected as the most appropriate one with the least passing from very susceptible landslide classes and it was then implemented in the field.