Does machine learning adequately predict earthquake induced landslides?

Abstract

Machine learning (ML) has been used for landslide susceptibility analysis for a while; however, studies using real-time earthquake induced landslide data are barely used. We used the data from the 2015 Gorkha earthquake in Nepal to assess adequacy of various machine learning models and segregated the importance of various landslide conditioning factors in this study. We used five supervised machine learning (ML) algorithms, namely, Logistic Regression (LR), Random Forest (RF), Extreme Gradient Boosting (XGB), Support Vector Machine (SVM), and Extremely Randomized Trees Classifier (ET) to predict earthquake-induced landslide susceptibility. Due to the Gorkha earthquake and its aftershocks, around 25,000 landslides were triggered, among them 23,217 were located within the study area. Using landslide polygons, 23,217 landslide points, and 23,213 randomly selected non-landslide points were generated. The data was randomly divided into 70:30 ratio to create training and testing data sets for model training and evaluation. Results of the SHapely Additive exPlanations (SHAP) analysis and factor importance analysis using ET infer that the first five crucial factors governing earthquake induced landslides are topographic ruggedness index, slope, distance to fault, peak ground acceleration, and distance to rupture. We evaluated the ML models by comparing their performance using accuracy, precision, recall, F1 score, Kappa value, and Matthew's correlation coefficient (MCC). The results indicated that the accuracy varies between 86.60% for ET to 74.40% for LR. The area under the ROC curve for five machine learning algorithms occurs between 0.866 and 0.744 for actual label prediction and 0.935 to 0.819 for probabilistic prediction. Based on the comparative performance assessment, ET is designated to predict earthquake induced landslides more efficiently than the other ML algorithms. The sum of the results highlights that ensemble learning outperforms other ML based classifiers to assess earthquake induced landslide susceptibility.