Evaluation of different models in rainfall-triggered landslide susceptibility mapping: a case study in Chunan, southeast China

Abstract

This study applied the information value model (IVM), logistic regression (LR), artificial neural networks (ANN) and support vector machine (SVM) to rainfall-triggered landslide susceptibility mapping for a typical study area in southeast China. Their capabilities under 42 modelling scenarios with different combination of 9 conditioning factors were tested and compared using the performance metric of area under receiver operating characteristic curve (AUC). The results showed that all the models could acceptably produce landslide susceptibility maps (LSMs) with the achieved AUC of 86.05, 85.89, 87.65 and 85.63% for model training, and that of 86.03, 85.7, 86.4 and 85.11% for model validation. Slightly, ANN was best, followed by LR, IVM and SVM. The LR showed a better generalization capability under the circumstances of changing combination of the conditioning factors, whereas the performance of ANN and SVM appeared to be sensitive to that. Although the comparable performance (i.e. AUC) could be achieved by all the models, the hazard zones classified from the produced LSMs exhibited considerable variation in spatial patterns. The IVM tended to overestimate the landslide susceptibility compared to others, and predicted a less area at very low-level hazard. The associated uncertainty caused by the modelling methods will lead to different management costs and risks being taken when applying the produced LSMs in geological engineering design and planning.