Manifestation of LiDAR-Derived Parameters in the Spatial Prediction of Landslides Using Novel Ensemble Evidential Belief Functions and Support Vector Machine Models in GIS

Abstract

Landslide susceptibility mapping is indispensable for disaster management and planning development operations in mountainous regions. The potential use of light detection and ranging (LiDAR) data was explored in this study for deriving landslide-conditioning factors for the spatial prediction of landslide-susceptible areas in a landslide-prone area in Ulu Klang, Malaysia. Nine landslide-conditioning factors, such as altitude, slope, aspect, curvature, stream power index (SPI), topographic wetness index (TWI), terrain roughness index (TRI), sediment transport index (STI), and slope length (SL), were directly derived from LiDAR for landslide-susceptibility mapping. The main objective of this research was to propose a novel ensemble landslide susceptibility mapping method to enhance the performance of individual methods of support vector machine (SVM) and evidential belief function (EBF). SVM is time-consuming when various data types, such as nominal, scale, and ordinal, are used. This characteristic of the individual SVM method is not optimal for hazard modeling. This drawback can be resolved by assessing the effects of the classes of each conditioning factor on landslide occurrence through a data-driven EBF method. Hence, EBF was applied in this study, and weights were acquired for the classes of each conditioning factor. The conditioning factors were reclassified based on the attained weights and entered into SVM as a scale to evaluate the correlation between landslide occurrence and each conditioning factor. Four SVM kernel types [radial basis function kernel (RBF), sigmoid kernel (SIG), linear kernel (LN), and polynomial kernel (PL)] were tested to explore the efficiency of each kernel in SVM modeling. The efficiencies of the ensemble EBF and SVM methods were examined through area under curve (AUC). The RBF kernel obtained better results than the other kernel types. The success and prediction rates obtained from the validation results of ensemble EBF and RBF-SVM method were 83.04% and 80.04%, respectively. The proposed novel ensemble method reasonably accelerated the processing and enhanced the results by combining the advantages of both methods.