Feature elimination and comparison of machine learning algorithms in landslide susceptibility mapping.

Landslide susceptibility assessment was adopted for the Idukki region using 6 machine learning models viz., Adaptive Boosting (AdaBoost), Naïve Bayes (NB), Neural Network (NNET), Random Forest (RF), Support Vector Machine (SVM) and eXtreme Gradient Boosting (XGBoost) models. Due to the severe calamity caused by landslides in the Idukki district during the August 2018 monsoon, the district was listed as one of the top hotspots for landslides, which is why this study focuses on them. Eighteen conditioning factors that influenced the occurrence of the landslides in the Idukki region were considered primarily in this study. However, the Recursive Feature Elimination (RFE) approach was adopted to eliminate the variables with low importance. The research focuses on the sensitivity of the causative factors and explores the conditioning parameters that are considered necessary by each machine learning algorithm. Furthermore, the Accuracy, Kappa index, ROC-AUC, Sensitivity, Specificity were computed to estimate the performance of the models. The RF model portrayed the best prediction accuracy in training and testing cases, followed by the AdaBoost and XGBoost models. However, while determining the consistency of the model based on its practical applicability by generating the landslide susceptibility maps, the AdaBoost and NNET model showed better consistency, and the XGBoost model displayed inconsistency. Thus, the AdaBoost model was the best in accuracy and practical consistency compared to the other five models. [ABSTRACT FROM AUTHOR]