Landslide susceptibility mapping using an automatic sampling algorithm based on two level random sampling

doi:10.1016/j.cageo.2019.104329

Return

Landslide susceptibility mapping using an automatic sampling algorithm based on two level random sampling

2019-12-01

PRE

Hakan Aktaş

Bekir Taner San *

DOI：10.1016/j.cageo.2019.104329

Original

Original request for help

Save

Abstract

En 中文

The aim of this study was to develop and use an algorithm to generate the automatic landslide susceptibility map. The proposed algorithm based on the two-level random sampling (2LRS) strategy and machine learning classification was generated using MATLAB. Performing automatic susceptibility mapping using machine learning classification requires an automatic and robust algorithm for the training of the constructed model. The proposed algorithm contains 20 steps and most of them have novel solutions for sampling. The user could also change the ratio of training and testing with this algorithm to perform automatic landslide susceptibility mapping. During the study, the 28 parameters used as input data sets which are the Digital Elevation Model (DEM), slope, aspect, plan curvature, profile curvature, the convergence index, the Topographic Wetness Index (TWI), the LS factor, the Normalized Difference Vegetation Index (NDVI), the Kaolinite Index, the Calcite Index, the OH Index, distance to fault lines, distance to channels, and 14 decorrelation stretched Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) bands. In this study, a total of 12 susceptibility models with different numbers of samples were automatically generated and tested using the produced algorithm in the study area, which contains active deep-seated rotational landslides (Alakir catchment area [Western Antalya, Turkey]). In addition, derivatives of the random forest (RF), support vector machines (SVM), and decision tree (DT) algorithms were compared in the MATLAB Classification Learner Toolbox according to their accuracies. The Medium Gaussian SVM has the highest level of accuracy (84%) among them. The constructed 12 models with different numbers of samples were also tested according to their spatial performance and their processing times. Then, the obtained area under curve values (AUCs) were obtained between 0.90 (in 360.009 s) and 0.84 (in 78.307 s).

Keywords:

Automatic landslide susceptibility

SVM

MATLAB

Image processing

Sampling algorithm

AI Summary

Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.