A study of non-landslide samples and weights for mapping landslide susceptibility using regression and clustering methods

Landslides are considered one of the most frequent natural disasters that cause human loss and damage property, affecting the sustainability of communities worldwide. Mapping landslide susceptibility has gained a lot of interest from researchers. Using machine learning methods to yield landslide susceptibility maps (LSMs) has been extensively studied in the last decade. However, the problems of sampling non-landslides and weighting classes of conditioning factors are not thoroughly investigated in landslide susceptibility models. Moreover, the process of zoning LSMs for different susceptible areas is often overlooked. To address this gap, we present a study on sampling non-landslides and computing appropriate weights, which are used to train machine learning models to obtain accurate outputs. The outputs are clustered into different susceptible levels by K-means clustering method. Our study was conducted in Nam Pam area, Vietnam. A landslide inventory of 71 landslide points and 44 delineated landslide polygons, and 13 conditioning factors were selected in this study area. Furthermore, three regression methods, i.e., Bayesian, K-Nearest Neighbors, and Support vector machine (SVM), were taken to generate LSMs. We assessed three models based on various metrics, including 5 statistical metrics and AUC. The reliability of LSMs was evaluated by MFR value. Empirical results have shown that SVM method and the 2:1 ratio of non-landslides to landslides are recommended to build landslide susceptibility models.