Application of classification coupled with PCA and SMOTE, for obtaining safety factor of landslide based on HRA

摘要

Machine learning algorithms have been recently applied to build a landslide susceptibility map. The objective of this studyis to find whether classification algorithms of machine learning are suitable for obtaining safety factor based on a highrisk-area (HRA) model, composed of eight geotechnical properties. Each property value is designated as an input value formachine learning, and the output value is determined as a safety factor. The data are transformed into continuous data afterpreprocessing with label encoding since the data have a discontinuous pattern. The DT, KNN, LR, RF, and SVM algorithmsare selected to perform the classification with train and validation ratio of 7:3. To improve the reliability of the results,the classification is also performed after applying the PCA technique, which can reduce eight dimensions to two principalcomponents. In addition, the number of data is equally oversampled using the SMOTE technique to solve the data imbalanceproblem for each class, and the results of classification are also compared. The PCA shows a limited ability to reflectthe characteristics of the original data, and the oversampled data by the SMOTE provides high reliability. The results showthat the RF is suitable for performing classification with high accuracy in the range of 1.2-2.4 of safety factors. This studydemonstrates that it is possible to classify even discontinuous data through a preprocessing technique, and SMOTE canimprove the accuracy of landslide risk mapping.