Landslide susceptibility modeling based on integration of confidence-weighted samples and factor coefficient of variation

Abstract

[Objective] To address the problem of insufficient model accuracy caused by data uncertainty in landslide susceptibility assessment, a modeling approach is proposed that integrates confidence-weighted samples with coefficient of variation(COV) of factors. [Methods] Taking Pingjiang County, Hunan Province as the study area, a three-level sample confidence evaluation system(1. 0, 0. 8, 0. 5) was established to weight landslide samples, and COV was used to quantify the uncertainties of key factors such as slope(COV= 0. 15) and rainfall intensity(COV= 0. 12). On this basis, Bayesian neural network(BNN) and XGBoost models were employed to develop a dual-dimensional uncertainty integration model. [Results] The results showed that the proposed dual-dimensional integration method significantly improved model performance, with the accuracy of the XGBoost model increasing from 0. 721 to 0. 867, and its AUC value rising from 0. 745 to 0. 892. The BNN model achieved an accuracy improvement from 0. 715 to 0. 861, with its AUC value increasing from 0. 738 to 0. 885. BNN visually demonstrated the uncertainty of landslide probability distribution through probability density function, while XGBoost reflected prediction deviations via confidence intervals. [Conclusion] The findings provide a new paradigm for landslide susceptibility assessment in data-scarce and geologically complex regions. In the future, model robustness can be further improved through multi-source data fusion and physical model coupling.

Keywords

landslide susceptibility / data uncertainty / confidence-weighted samples / coefficient of variation / dual-dimensional integration / machine learning / Bayesian neural network / influencing factors