Intergration of Rotation Forest and MultiBoost Ensembles with Forest by Penalizing Attributes for Spatial Prediction of Landslide Susceptibility

Dien Bien province is considered one of the most prone to landslides areas of Vietnam. This study focused on developing landslide susceptibility maps (LSMs) for the study area using advanced machine learning prediction models, namely, Forest by penalizing attribute (FPA) and its optimized versions employing random forest (RF) and multibostAB (MAB) algorithms. For this purpose, distance to rivers, elevations, distance to roads, NDVI, faults, curvature, slope, flow accumulation, SPI, geological, TWI, and aspect were taken as input parameters. These attributes were used to calculate land susceptibility indices (LSIs) using the Frequency ratio (F r ) method, which was considered the target variable in this study. Landslide susceptibility was classified into five classes; very high, high, moderate, low and very low. The performance of the developed models was assessed using AUC curve, sensitivity, specificity, accuracy, PPA, and NPA. All the three models manifested reliable predictions with a good accuracy level, FPA-MAB being the most accurate one. The value of AUC for FPA-MAB was calculated as 0.985 and 0.814 for the training and testing data, respectively. The developed models were used to generate landslide susceptibility models that can accurately be sued for disaster risk management of the study area.