Comparison of landslide susceptibility models and their robustness analysis: a case study from the NW Himalayas, Pakistan

Abstract

Machine learning methods are considered as most effective approaches to accomplish landslide susceptibility analysis around the globe. Landslide susceptibility maps (LSMs) have been frequently executed by statistical models in NW Himalaya. However, the comparison and applications of the statistical models with modern machine learning techniques has not been fully explored in this region. Hence, this study aims to compare the predicted performance of statistical and popular machine learning models to explore robust landslide prediction model in the landslide-prone area of NW Himalaya and investigate the compensations and limitations of these models to grasp a more precise and consistent result. This study presented machine learning approaches based on the artificial neural network (ANN), support vector machine (SVM) and logistic regression (LR) and the statistical methods based on the frequency ratio (FR), information value (InfoV) and weight of evidence (WoE). For this purpose, first an inventory map of 1507 landslides was prepared and randomly divided into training (70%) and testing (30%) dataset. Furthermore, 12 landslide conditioning factors (LCFs) were extracted from geospatial dataset to prepare thematic layers in ArcGIS. Thereafter, factor analysis was performed to eliminate colinear and least important variables which can mislead the results. The results showed that all selected LCFs are noncolinear and have significant contribution on landslides initiation, however, lithology, slope angle, annual rainfall and landuse were most influential factors. For modeling purpose, landslide inventory was correlated against all LCFs and trained into six models to produce respective LSMs. Finally, the performance of produced LSM models was validated and compared through area under receiver operating characteristic curve (AUROC), Accuracy, Recall, F 1-score and Cohen’s Kappa coefficients to assess the robustness of employed models. The results exhibit that the performance scores of machine learning models were considerably superior than statistical models. While, the AUROC values based on validation dataset indicate that LR (0.89) has better prediction ability followed by SVM (0.86), ANN (0.84), FR (0.83), InfoV (0.82) and WoE (0.81) in this study. Therefore, it is reasoned out that the machine learning methods are more reliable in generating adequate LSMs. However, the LR is recommended as most efficient model for predicting landslide susceptible zones in study region and thus can be considered as robust model for landslide susceptibility assessment in similar geo-environmental regimes.