Selecting optimal conditioning factors in shallow translational landslide susceptibility mapping using genetic algorithm

• Genetic algorithm was applied for the selection of optimum conditioning factors.
• GA-selected models outperformed the user's models selected from literature.
• 7-factor model was found to be the optimal model for the problem.
• Slope, lithology and distance to drainage were the effective factors.
• Superiority of GA-based models was confirmed by statistical tests.

Many landslide conditioning factors have been considered in the literature for landslide susceptibility mapping, but it is not certain which factors produce the best result for an area under analysis. With the availability of increasing number of landslide conditioning factors, finding the best combination of factors has become an important research issue. In this study, genetic algorithms (GAs) were applied to find the best factor combination among 16 factors available for the study area, Macka District of Trabzon, Turkey. Performances of the models including 4 to 15 factors were evaluated using logistic regression to investigate the effect of varying number of factors and the most effective factors were determined. Results showed that prediction accuracy of the models constructed with GA-selected factors increased to a certain level (up to 8 factors) and then showed a stable trend producing statistically similar results. In order to show the robustness of the GA algorithm, prediction performances of the models constructed with 4 to 8 factors determined by the GA were compared with those of models constructed with factor combinations applied in the literature. While slope, lithology and distance to drainage were found to be the most effective factors, soil depth, slope length and profile curvature were found to be the least effective ones. It was also found that the models determined by the GA generally produced better results than user's models. Also, the goodness of the GA-based models was confirmed by success rate curve analysis and McNemar's test. In summary, results revealed the robustness of the GA when searching the optimal landslide conditioning factors among a large number of factors.