Landslide susceptibility mapping using logistic regression, random forests, and artificial neural networks: a case study in Mariana/MG, Brazil

The landslide susceptibility mapping (LSM) plays an important role in risk management. This study evaluated the predictive capabilities of three machine learning (ML) approaches applied to LSM: logistic regression (LR), random forests (RF), and artificial neural networks (ANN). The study was conducted in a mountainous region of Mariana/MG, Brazil. Initially, a point inventory with 364 landslides and 364 stable regions was randomly partitioned in a 70% training and 30% testing ratio for the models. Nine landslide conditioning factors (LCF), ranked by information gain (IG), were considered: slope angle (IG=0.486), geomorphology (IG=0.235), topographic wetness index - TWI (IG=0.138), lithology (IG=0.077), slope orientation (IG=0.067), topographic position index - TPI (IG=0.052), distance from drainage (IG=0.032), slope curvature (IG=0.029) and the distance from roads (IG=0.024). The evaluation of the area under the curve (AUC-ROC) and the classification efficiency rates in high () and low () susceptibility were used to compare the results of the approaches. The results demonstrated that although RF (AUC-ROC=0,947, =6,808, =0,030) slightly outperformed LR (AUC-ROC=0,936, =5,695, =0,050) and ANN (AUC-ROC=0,934, =6,495, =0,060), all the approaches exhibited high predictive capability in identifying areas susceptible to landslides.