Large-scale landslide susceptibility evaluation in upper reaches of Minjiang River based on different selection methods for landslide negative samples and random forest algorithm

Abstract

[Objective] Accurate landslide susceptibility results enable precise prevention and control of landslide hazards and risks. In landslide susceptibility evaluation, selection methods for different landslide negative samples represent a critical uncertainty factor that affects the prediction accuracy of landslide susceptibility. [Methods] Taking the mountainous area of the Minjiang River basin in Sichuan Province as the study area, data on 881 large landslides(>106 m³) were compiled through remote sensing imagery. Thirteen landslide evaluation factors including topography and geomorphology, basic geology, hydrogeology, geological environment, seismic parameters, and human activities were selected, and factor redundancy was examined through collinearity analysis. Subsequently, landslide negative samples were selected using random sampling across the study area, sampling in slope zones below 10°, sampling in areas outside 1 km buffer zones around landslides, Information Value(IV) method, Support Vector Machine(SVM) method, and semi-supervised method, with the same proportion as landslide positive samples. These negative samples were further coupled with the Random Forest(RF) to establish Random RF, Low-Slope RF, Buffer RF, IV-RF, SVM-RF, and Semi-Supervised RF models for landslide susceptibility zoning. Finally, the prediction accuracy of different sampling models for landslide negative samples was compared and evaluated using the mean Area Under the Curve(AUC) value derived from the Receiver Operating Characteristic(ROC) curve. [Results] The results showed that:(1) the high and extremely high landslide susceptibility zones obtained by different sampling methods for landslide negative samples were predominantly concentrated on both sides of the river valleys from Wolong to Yingxiu, Miansi to Gu'ergou, Heihu to Musu, and Diexi to Songpan.(2) The prediction accuracy of landslide susceptibility using different sampling methods for landslide negative samples ranked as follows: Semi-Supervised RF ($\overline{A U C}=0.971$)>SVM-RF($\overline{A U C}=0.954$)>IV-RF($\overline{A U C}=0.945$)>Buffer RF($\overline{A U C}=0.902$)>Low-Slope RF($\overline{A U C}=0.895$)>Random RF($\overline{A U C}=0.882$). (3) The selection of landslide negative samples in low-susceptibility areas significantly enhanced the prediction accuracy of susceptibility. Compared to Random RF model, the values of the Low-Slope RF, Buffer RF, IV-RF, SVM-RF, and Semi-Supervised RF models increased by 0.013, 0.02, 0.063, 0.072, and 0.089, respectively. [Conclusion] The Semi-Supervised RF model exhibits the smallest standard deviation(0.004) and the highest mean AUC value(0.971), demonstrating optimal stability and prediction capability. This indicates that the semi-supervised sampling method offers the best optimization for the model. These research findings provide references for selecting landslide negative samples and establishing models in landslide susceptibility prediction, while offering theoretical support for landslide risk assessment and disaster mitigation strategies in the upper Minjiang River Basin.

Keywords

landslide susceptibility prediction / sampling strategy / random forest / semi-supervised method / model averaging method / Minjiang River Basin / landslides / influencing factors