Optimizing the Application of Machine Learning Models in Predicting Landslide Susceptibility Using the Information Value Model in Junlian County of Sichuan Basin

Abstract Constructing accurate landslide susceptibility models is crucial for effective landslide prevention.This study explores methods to enhance the accuracy of landslide susceptibility models.This paper focuses on Junlian County, Sichuan, as the study area.Initially, a landslide inventory was created using field surveys and historical records.Eight environmental factors were identified via correlation analysis: elevation, slope, aspect, stratigraphic lithology, and distances from faults, roads, rivers, and areas of land use.Subsequently, we constructed an information value model.For training the IV-RF model, non-landslide points in areas of low susceptibility were randomly selected at various ratios (1:1, 1:2, 1:3, 1:4, 1:5).The optimal ratio was used to develop coupled models (IV-RF, IV-LR, IV-SVM, IV-BP), comparing their accuracy and discussing the impact of environmental factors on landslide susceptibility.Results indicate that: (1) the highest prediction accuracy was achieved with a non-landslide ratio of 1:1; (2) the IV-RF model achieved the highest AUC of 0.994; and (3) the most significant factors influencing landslide distribution were stratigraphic lithology and river proximity, followed by elevation and fault distance.