Performance comparison of landslide susceptibility mapping under multiple machine-learning based models considering InSAR deformation: a case study of the upper Jinsha River

Abstract

Landslide susceptibility mapping (LSM) comprehensively evaluates the spatial probability of landslide occurrence by using different environmental factors. However, most of the evaluation methods ignore the dynamic characteristic factors of landslides, which makes it difficult to obtain reliable prediction results. Taking the upper reaches of the Jinsha River as the study area, this article introduces the deformation data into the landslide characteristic model and proposes an improved landslide susceptibility evaluation method. Four kinds of landslide susceptibility machine learning models were constructed by collecting 20 landslide related factors. The prediction accuracy of machine learning models is compared, and the performance of different models and the improvement of model performance by deformation information are evaluated. The results show that the performance of Random Forest and XGBoost model is better than SVM and logistic regression model. The prediction accuracy of Random Forest and XGBoost model is improved obviously after InSAR deformation is introduced. 96.9 and 93.19% of landslide areas were reasonably classified as high or very high risk levels. Compared with the calculation result of traditional model, the proportion of high and very high risk pixels in landslide area is increased by 2.97 and 1.13%, respectively. In addition, the percentage of high and very high risk areas in the susceptibility evaluation area increased from 15.45 to 16.23% and 18.73 to 21.89%, respectively. The accuracy of Random Forest and XGBoost models increased from 0.793 to 0.878 and 0.776 to 0.812, respectively, and the AUC increased by 0.9 and 1.7%, respectively. The SHAP and traditional feature importance analysis reveals that rainfall, aspect, temperature and NDVI are the main influencing factors of landslide in the upper reaches of the Jinsha River.

Keywords:

• Landslide susceptibility
• InSAR
• machine learning
• random forest
• Jinsha River

Acknowledgments

We are very grateful for the Sentinel-1A data provided by European Space Agency (ESA).

Authors’ Contributions

J.Y. conceived the manuscript; X.Y. provided funding support and helped improve the manuscript; Z.Z. and X.L. provided methods and data.

Disclosure Statement

The authors declare no conflict of interest.

Data Availability Statement

Sentinel-1A data used in this study were provided by European. Space Agency (ESA) through the Sentinel-1 Scientific Data Hub.

Additional information

Funding

The work was funded by China Geology Survey Project (DD20230433); National Key R&D Program of China (2018YFC1505002).