Landslide susceptibility mapping: application of novel hybridization of rotation forests (RF) and Java decision trees (J48).

In this study, we presented a novel hybrid artificial intelligence method for landslide susceptibility mapping (LSM) in the Kalaleh Watershed, Golestan Province, Iran. This method uses rotation forests (RF) as a Meta classifier based on Java decision trees (J48), as a base classifier named RF–J48. The created model was compared to some benchmark models, such as the J48, the open-source naive Bayes tree, and the reduced error pruning tree (REPTree). Locations of 260 different landslides were found and mapped. Additionally, it was split into training data (70%) and testing data (30%) for spatial modeling and validation analysis. Twelve landslide effective factors were chosen for the landslide susceptibility mapping process based on a literature review and the results of the multicollinearity evaluation (LSM). The approximate percentage of high LS over the Kalaleh Watershed catchment for four models is RF–J48 = 19.92%, J48 = 15.87%, NBTree = 20.13% and REPTree = 16.76%, respectively. The LS model is evaluated through the methods of TSS, efficiency, and kappa index methods on the training and validation datasets, shown in Fig. 8. Efficiency values computed using validation data/training data for each model is NBTree = 0.855/0.908, J48 = 0.868/0.911, REPTree = 0.862/0.877, and RF–J48 = 0.875/0.992, respectively. The Kappa index for each model has shown significant accuracy. The index values are maximum for the ensemble model RF–J48 of 0.537 and 0.88 for validation and training data. [ABSTRACT FROM AUTHOR]