GIS-based rainfall-induced landslide susceptibility mapping: a comparative analysis of machine learning algorithms and a numerical method in Kvam, Norway

Abstract

Analysis and prediction of climate-driven geohazards, such as rainfall-induced landslides and slope failures, are becoming more challenging given the changing climate where extreme events are inevitable. Therefore, there is a need to move beyond conventional sources of data and consider multiple types of data for more accurate analysis and prediction of landslides. In recent years, Data Fusion and Machine Learning techniques have played an important role in paving the path towards a better understanding of the problem and finding more accurate models at regional and local levels that incorporate several contributing factors for slope failures. The purpose of the study is thus to evaluate the capacities of machine learning models in landslide susceptibility prediction and analyze their model performance in comparison of a numerical method, Transient Rainfall Infiltration and Grid-Based Regional Slope-Stability Model (TRIGRS). Classic machine learning models, namely Multi-Layer Perceptron Neural Network (MLP), Random Forest (RF), Gradient Boosted Regression Tree (GBRT) and Extreme Gradient Boosting (XGBoost) are selected and developed respectively. The study is carried out based on a preliminary field survey of rainfall-induced landslides near Kvam village, Norway, in June 2011. A methodology workflow of landslide susceptibility modeling is proposed, in which effective data processing approaches including feature selection, data resampling, data splitting, and feature scaling are discussed and summarized. The optimal hyperparameter optimization method is determined by performing a comparative time efficiency analysis of Bayesian and Grid Search methods. It is concluded that GBRT is the optimal method for landslide susceptibility mapping in the study case of Kvam based on seven popular model evaluation metrics. Other tree-based machine learning algorithms (RF and XGBoost) also show an overall outstanding performance and computational efficiency in comparison to MLP and TRIGRS models. The landslide susceptibility maps developed by prediction results from five models are also presented and statistically analyzed. Corresponding model performance ranks are found with results from model evaluation metrics.