Landslide Susceptibility Prediction: Improving the Quality of Landslide Samples by Isolation Forests

Abstract

Landslide susceptibility prediction (LSP) is the first step to ease landslide disasters with the application of various machine learning methods. A complete landslide inventory, which is essential but difficult to obtain, should include high-quality landslide and non-landslide samples. The insufficient number of landslide samples and the low purity of non-landslide samples limit the performance of the machine learning models. In response, this study aims to explore the effectiveness of isolated forest (IF) to solve the problem of insufficient landslide samples. IF belongs to unsupervised learning, and only a small share of landslide samples in the study area were required for modeling, while the remaining samples were used for testing. Its performance was compared to another advanced integration model, adaptive boosting integrated with decision tree (Ada-DT), which belongs to two-class classifiers (TCC) and needs a sufficient number of samples. Huangpu District, Guangzhou City, Guangdong Province in China, was selected as the study area, and 13 predisposing factors were prepared for the modeling. Results showed that the IF proved its effectiveness with an AUC value of 0.875, although the Ada-DT model performed better (AUC = 0.921). IF outperformed the Ada-DT model in terms of recognizing landslides, and the sensitivity values of IF and the Ada-DT model were 90.00% and 86.67%, respectively, while the Ada-DT model performed better in terms of specificity. Two susceptibility maps obtained by the models were basically consistent with the field investigation, while the areas predicted by IF tended to be conservative as higher risk areas were presented, and the Ada-DT model was likely to be risky. It is suggested to select non-landslide samples from the very low susceptibility areas predicted by the IF model to form a more reliable sample set for Ada-DT modeling. The conclusion confirms the practicality and advancement of the idea of anomaly detection in LSP and improves the application potential of machine learning algorithms for geohazards.