Spatial consistency assessment and landslide susceptibility prediction optimization

xiaohong zhou¹Yin Xing^2*

• ¹Jiangsu Geological Bureau, Nanjing, China, Nanjing, China
• ²Suzhou University of Science and Technology, Suzhou, China

At present, although a variety of landslide susceptibility models can achieve high prediction accuracy, their results are significantly inconsistent in spatial distribution, resulting in high prediction uncertainty, which brings challenges to the optimization of assessment methods suitable for such complex geological disasters. In order to reduce the uncertainty, this study proposes a machine learning ensemble modeling method with spatial consistency analysis. Taking Ruijin City of Jiangxi Province as the study area, based on the selection of 12 impact factors and hyperparameter optimization, three algorithms of XGBOOST, Random Forest (RF) and Support Vector Machine (SVM) were used to generate landslide susceptibility maps. All the models performed well, and the AUC value was between 0.84 and 0.93. However, the spatial consistency analysis showed that the spatial correlation across maps between the models was only 0.78 to 0.84, indicating that despite the high accuracy, there was still obvious spatial heterogeneity and uncertainty in the prediction. Therefore, a Logistic Regression (LR) fusion model was constructed with the historical landslide compilation as the dependent variable and the results of the three models as the independent variables. It was found that XGBOOST contributed the most, followed by RF and SVM. By integrating the three prediction results, the comprehensive susceptibility map was finally obtained, which was better than the single model in spatial consistency (correlation coefficient 0.87-0.91) and prediction accuracy (AUC=0.95). This research framework effectively reduces the uncertainty in landslide prediction and improves the reliability and accuracy of the assessment results.