|
Abstract
|
Landslide susceptibility assessment has always been the focus of landslide spatial prediction research. In the present study, Muchuan County was selected as the study area, and four well-known machine learning models were adopted, namely, rotation forest (RF), J48 decision tree (J48), alternating decision tree (ADTree) and random forest (RaF). They and their ensembles (RF-J48, RF-ADTree and RF-RaF) were applied to landslide spatial prediction in Muchuan County. Eleven landslide conditioning factors, including plan curvature, profile curvature, slope angle, elevation, topographic wetness index, land use, normalized difference vegetation index, soil, lithology, distance to roads and distance to rivers, were established. In addition, 279 landslide datasets were compiled and randomly divided into 195 landslide training datasets and 84 landslide verification datasets. The contributions of the eleven conditioning factors were analyzed by J48, ADTree, and RaF models, respectively. The results show that lithology, slope angle, elevation, land use, soil, and distance to roads were the six principal landslide conditioning factors. Then, the Jenks natural break method was used to divide the landslide susceptibility maps into five grades. In addition, the accuracy of the above six models was verified by implementing the receiver operating characteristic curve and area under the receiver operating characteristic curve. The RF-RaF model achieved the best performance, and the rest were ranked as follows: RF-ADTree model, RaF model, RF-J48 model, ADTree model and J48 model. The results could provide scientific references for local natural resource departments.
|