Earthquake-Induced Landslide Susceptibility Assessment Using a Novel Model Based on Gradient Boosting Machine Learning and Class Balancing Methods

Predicting the susceptibility of a specific part landslide (SSPL) involves predicting likelihood that (e.g., entire landslide, source area, or scarp) will form in given area. When SSPL, samples are far less than non-landslide samples. This class imbalance makes it difficult to predict SSPL. paper proposes an advanced artificial intelligence (AI) model based on dice-cross entropy (DCE) loss function and XGBoost (XGBDCE) Light Gradient Boosting Machine (LGBDCE) ameliorate SSPL prediction. We select earthquake-induced landslides from 2018 Hokkaido earthquake as case study evaluate our proposed method. First, six different datasets with 24 influencing factors 10,422 established using remote sensing geographic information system technologies. Then, each datasets, four algorithms (XGB, LGB, random-forest (RF) linear discriminant analysis (LDA)) balancing methods (non-balance (NB), equal-quantity sampling (EQS), inverse landslide-frequency weighting (ILW), DCE loss) applied The results show non-balanced method underestimates susceptibility, ILW EQS overestimate while produces more balanced results. prediction performance XGBDCE (average area under receiver operating characteristic curve (0.970) surpasses RF (0.956), LGB (0.962), LDA (0.921). Our produce unbiased precise existing models, have great potential accurate general landslide) detailed combining run-out modeling) assessments, which can be further hazard risk assessments.