The determinant role of the modeling goal and end-user’s need in opting for the superior landslide susceptibility model (A case study: Tajan watershed, Mazandaran province)

Landslides are isolated processes which may not be very large, but they can occur frequently and cause sizable damages. In most areas, there is a vivid pattern of irrational reaction while confronting such events. Nonetheless, such actions as avoidance, prevention, or restoration are more feasible for landslides than all other natural hazards because many discernable morphological symptoms appear months and even years before landslide occurrences. To the date, inherent driving forces of terrain processes have been identified quite well. Therefore, if we optimistically identify the landslide-prone areas, we would be able to reduce the landslide driven accidents through landslide susceptibility zonation. Nowadays, landslide susceptibility assessment endeavors have made great progress. Nevertheless, concurrent with advancements in developing susceptibility models, end-users have had many challenges selecting the superior model.

This study is focused on the determinant role of the modeling goal and end-user’s need in opting for the superior model in the context of landslide susceptibility assessment and generally any endeavor with a spatial connotation. Hence, three widely used data mining models including artificial neural network (ANN), support vector machine (SVM), and maximum entropy (MaxEnt) were adopted for landslide susceptibility assessment in one of the pilot subbasins of the Tajan Watershed in Mazandaran Province. Models’ results were assessed using six performance criteria including 1) areal distribution of the susceptibility classes in each model, 2) distribution of landslides within the susceptibility classes in each model, 3) Error Type I (false positive), 4) Error Type II (false negative), 5) area under the success rate curve and 6) area under the prediction rate curve, based on which models were ranked.

The first criterion showed that the MaxEnt, SVM, and ANN, respectively, have the highest to the lowest performance. The second criterion showed that the SVM, MaxEnt, and ANN, respectively, have the highest to the lowest performance. The third criterion with economic losses connotation often associated with the modeling errors, indicated a good performance of the SVM model, while the MaxEnt and ANN were concurrently second-ranked. The fourth criterion with a connotation of casualties and economic losses often associated with the modeling errors indicated a good performance of ANN, followed by MaxEnt and SVM. The results regarding the fifth and sixth criteria both revealed a great learning and prediction power of the SVM model, followed by MaxEnt and ANN.

The findings of this study attests for the notion that models superiority is rather a relative matter and despite the fact that landslide susceptibility results are resultant of local properties and cannot be generalized to other areas. Therefore opting for the superior model should be also carried out on the basis of engaging a wide range of performance criteria as well as acknowledging the modeling goal and end-user’s need.