Utilizing deep learning approach to develop landslide susceptibility mapping considering landslide types

Abstract

Landslide susceptibility evaluation is pivotal for mitigating landslide risk and enhancing early warning systems. Current practices in developing Landslide Susceptibility Mapping (LSM) often overlook the diverse mechanisms of landslides, and traditional machine learning (ML) models lack the capability for autonomous feature learning in landslide contexts. This study proposes a methodology that precedes the application of deep learning algorithms for LSM by classifying landslides and selecting relevant factors based on their deformation mechanisms. In the Zigui-Badong section of the Three Gorges Reservoir area (TGRA), landslides are classified into rock landslides (RL) and soil landslides (SL) based on the geological conditions and historical landslide inventory. A comprehensive evaluation index system, comprising thirteen factors is established. To identify the most pertinent factors for each type of landslide, these factors are ranked according to their contribution to landslide occurrence. For susceptibility assessment, this study introduces a Convolutional Neural Network (CNN) model and benchmarks its performance to traditional ML models including Classification and Regression Trees (CART) and Multilayer Perceptrons (MLP). The efficacy of these models is evaluated using the Receiver Operating Characteristic (ROC) curve and various statistical analysis methods. The findings indicate that LSMs that consider different types of landslides yield more accurate and realistic outcomes. The CNN model outperformes its counterparts, with MLP being the second most effective and CART the least effective. Overall, this study demonstrates the superiority of an LSM approach that accounts for landslide diversity over traditional, monolithic methods.