Landslide susceptibility mapping in the Loess Plateau of northwest China using three data-driven techniques-a case study from middle Yellow River catchment

Abstract

Accurate landslide susceptibility maps are relevant for stakeholders to take effective measures and perform land use planning. The present research aims at using three data-driven approaches to generate landslide susceptibility map for the middle Yellow River catchment (northwest China) and comparing their performances, including the weighted information value (WIV), where the IV model was integrated with analytic hierarchy process (AHP), the support vector machine (SVM) and the random forest (RF) models. A landslide inventory map including 684 historical landslides was generated first by visual interpretation of remote sensing images combined with a field survey. A total of 14 thematic layers were applied to serve as the landslides influencing factors. The Pearson correlation coefficient analyzed the correlation among these factors, and the C5.0 decision tree algorithm determined the factor importance. The results demonstrated the correlation between every two factors were all less than 0.5. Three factors (including distance to road, distance to river, and slope) were the most important contributions to the landslide occurrences in the region, whereas five factors (including NDWI, plan curvature, profile curvature, surface roughness, and aspect) had minor importance. All the models predict that most of the historical landslides are identified in moderate and high susceptibility areas. For the prediction percentage of landslides in high susceptibility areas, both SVM and RF models exceed 70%. The RF model represented the best performance, with high susceptibility zones accounting for 21.9% and landslide numbers accounting for 90.5%. The comparison among the receiver operating characteristic curves indicated that the accuracy was higher in the RF model than in the other two models: the area under the curve (AUC) for the RF was 0.904, whereas that for the WIV and the SVM were 0.845 and 0.847, respectively. Hence, the RF was proven suitable for assessing the landslide susceptibility in the region. Current results can provide valuable references for future studies and landslide risk mitigation strategies.