Analysis on Structural Uncertainty of Ensemble Method for Landslide Susceptibility Analysis

Abstract:

Policymakers have prepared a countermeasure to reduce landslide damage. Landslide susceptibility analysis provides scientific evidence when establishing these measures. Accordingly, many researchers have applied the methodology using diverse models to various study sites. Recently, as it has shown that the method of 'ensemble', which synthesizes the results of several models, improves the performance by reducing bias and variance of a single model, the method has been frequently applied. However, when applying the ensemble method, it is necessary to consider the uncertainty caused by the structural differences of each model. This is because, although the accuracy of the ensemble model is high, from a spatial perspective, the variability of the predicted probability in a specific region could be very high.

Therefore, the purpose of this study is to increase the applicability of landslide susceptibility analysis by considering both the understanding of the spatial distribution and characteristics of these uncertainties. The area of study was Hadong, Republic of Korea. Five models and Ten variables were selected to analyze landslide susceptibility. the results of five models of GLM, ANN, MaxEnt, RF, and SVM were aggregated using an ensemble method. Subsequently, structural uncertainty was quantified for each pixel by calculating the deviation between the probability predicted by every single model and the probability predicted by the ensemble model. Also, by presenting the value of each pixel as a map, the structural uncertainty of the ensemble result was visualized, and the geographical and environmental characteristics of the region with high uncertainty were analyzed.

The uncertainty distribution of this study enables policymakers to identify the reliability of the landslide susceptibility analysis result spatially and furthermore, provides information to exclude areas with high uncertainty for efficient policies. In addition, the characteristics of regions with high uncertainty will serve as a reference for reducing uncertainty in the future.

※ This work was supported by Korea Environment Industry & Technology Institute(KEITI) through The Decision Support System Development Project for Environmental Impact Assessment, funded by Korea Ministry of Environment(MOE)(No. 2020002990009)