Hybrid ensemble machine learning approaches for landslide susceptibility mapping using different sampling ratios at East Sikkim Himalayan, India

Abstract Landslide is a big problem in the mountainous region all over the world. Sikkim Himalayan region is also suffering from landslide problem. This study's main objective was to generate landslide susceptibility map (LSM) considering the hybrid ensemble of machine learning approaches using different sample ratios. Random Forest (RF) as the base classifier an ensemble with bagging, Rotation Forest (RTF), and Random Subspace (RS) Meta classifiers were used for spatial landslide modeling. First, collected 86 landslides locations through field investigation and from Sikkim district disaster office were mapped as a landslide inventory. Collected landslide locations were categorized into training and testing datasets randomly using four sample ratios (50:50, 60:40, 70:30 and 80:20). Based on the four sampling ratios and fifteen conditioning factors, a total of sixteen LSMs were prepared using RF, Bagging-RF (B-RF), RTF-RF and RS-RF in GIS platform. For assessing the modeling accuracy and comparison among these, the area under the receiver operating characteristics (AUROC) and other statistical methods such as root-mean-square-error (RMSE), mean-absolute-error (MAE) and R-index methods were used. The overall proficiency of RS-RF (AUC = 0.871, 0.847 of 50%:50%, AUC = 0.925, 0.931 of 60%:40%, AUC = 0.933, 0.939 of 70%:30%; AUC = 0.927, 0.933 of 80%:20%) was found to be substantially greater than the results of RF, B-RF, and RTF-RF. The RS-RF model and 70:30 sample ratio had the highest goodness-of-fit and accuracy as per the RMSE, MAE, and R-index methods. Furthermore, the model based on RS-RF is a promising and acceptable way of mapping regional landslides.