Landslide susceptibility mapping in southern China: a study based on remote sensing data and random forests

Landslide susceptibility maps (LSM) assess the geospatial distribution of potential landslide hazards by analyzing the intrinsic and extrinsic factors affecting landslides, and provide decision support for urban construction planning and landslide hazard prevention. In order to obtain a more accurate LSM, we combine the information model (I) and random forest (RF) methods. First, we combined 41,965 landslide data records and eight types of basic environmental factors (digital elevation model (DEM), slope, aspect, lithology, fault distance, normalized difference vegetation index (NDVI), annual precipitation, and land cover type) in southern China, and calculated the informativeness values (IV) of each factor by the I model, and then used these IV as the input variables of the RF. Subsequently, we divided the landslides and randomly selected non-landslide samples as training set and test set, predict landslide susceptibility using the trained model and make graphs. Finally, evaluate the I-RF model by using indicators such as accuracy (ACC) and area under the curve (AUC). The results show that: 1. The ACC of the I-RF model is 0.803 and the AUC is 0.894, which is a significant improvement over the I model alone; 2. The LSM obtained by the I-RF model show that the high and very high susceptibility zones account for 21.92% and 3.77% of the area, respectively, but contain 56.01% and 26.80% of the landslide hazards.