Landslide susceptibility prediction mapping with advanced ensemble models: Son La province, Vietnam

Abstract

Landslide is a severe geohazard in many mountainous areas of Vietnam during the rainy season. They directly threaten human lives and properties every year. Landslide susceptibility maps are useful tools for risk mitigation, land-use planning, and early warning systems for local areas. It is necessary to update these maps continuously because of the complexity of landslide events. This fact requires further extending the approach techniques with practical implications. Therefore, this study aimed to develop landslide susceptibility prediction maps based on advanced machine learning (ML) techniques. Five state-of-the-art hybrid ML models were developed: bagging MLP, dagging MLP, decorate MLP, rotation forest MLP, and random subspace MLP with multilayer perceptron (MLP) as a base classifier. Sixteen causative factors were collected to build landslide susceptibility maps based on the relationship between historical landslide locations and specific local geo-environmental conditions. The model performance was verified using various statistical indexes. Based on the area under ROC curve (AUC) analysis results of the testing dataset, the rotation forest MLP model has the greatest predictive accuracy of AUC = 0.818. It is followed by the decorate MLP and bagging MLP (AUC = 0.804), the random subspace MLP model (AUC = 0.796), the dagging MLP (AUC = 0.789), and the single MLP (AUC = 0.698). The results of this study can be applied effectively to other mountainous regions to mitigate the risk of landslides.