Non-landslide sample for landslide susceptibility prediction modeling: A review of selection strategies and their influence rules

Abstract

A proper non-landslide sample selection strategy can improve landslide susceptibility prediction (LSP) accuracy. However, there may be uncertainties regarding the compatibility between different selection strategies and machine learning models, as well as in the extent of LSP performance enhancement after their coupling. To overcome these uncertainties, this study takes Wuning county of China as a case area, collecting 24 conditioning factors and 379 landslides data. Four non-landslide sample selection strategies, namely random selection, low-slope, buffer zone, and semi-supervised strategies, are then combined with landslide samples in a 1:1 ratio to serve as input variables for constructing LSP models using support vector machine (SVM), logistic regression (LR), random forest (RF) and extreme gradient boosting (XGBoost). Finally, the uncertainty of semi-supervised machine learning coupled models with a 1:2 ratio of landslide to non-landslide samples is analyzed and compared. The results show that: (1) The semi-supervised and low-slope strategies demonstrate higher prediction accuracy compared to the buffer zone and random selection strategies. Moreover, the RF coupled models are the most reliable, followed by the XGBoost, SVM, and LR coupled models; (2) Compared to a 1:1 ratio, a 1:2 ratio of landslide to non-landslide samples significantly improves prediction accuracy, suggesting that appropriately increasing the proportion of non-landslide samples helps to mitigate overfitting and enhance the identification of landslide samples; and (3) LSP is more sensitive to non-landslide sample selection strategies than to the choice of machine learning models. In conclusion, prioritizing reliable non-landslide samples is crucial for improving accuracy of LSP. • Reveal the effect of four non-landslide sample selection strategies including random, low-slope, buffer zone, and semi-supervised on landslide susceptibility modeling. • Compare the prediction performance of four machine learning models including SVM, LR, RF, and XGBoost coupling with different non-landslide sample selection strategies. • Evaluation of uncertainty through AUC accuracy, F1 score, distribution rules of landslide susceptibility indexes, and sensitivity analysis. • Analysis of the impact of landslide-to-non-landslide sample ratios of 1:1 and 1:2 on landslide susceptibility prediction accuracy.

Topics & Concepts