Spatial prediction of landslides for the Wanzhou District (China) applying a hybrid intelligent method based on random forest and cluster algorithms

In the Three Gorges Reservoir area of China, landslides have caused considerable losses of lives, environmental and social economy during the last decade. Hence, landslide susceptibility mapping is an urgent task that could help local decision makers in disaster risk assessment and management. This study aims at generating a regional landslide susceptibility map for the Wanzhou District in the Three Gorges Reservoir (China), based on random forest (RT) and cluster algorithms. Specifically, our objectives mainly include: (i) comparing the performances among different machine learning approaches, and (ii) validating the accuracy of a novel susceptibility reclassification method which used cluster algorithm. First, nine GIS-based thematic maps presenting landslide causal factors were prepared, including elevation, slope angle, aspect, lithology, land use, topographic wetness index (TWI), distance to rivers, distance to roads, and distance to geological structures. Total 441 landslides in a landslide inventory map were divided into two subsets: 75% landslides were used as training data, and 25% landslides were validation data. To establish the hybrid intelligent method, random forest was employed to calculate the landslide occurrence probability at every raster cell whereas the cluster algorithm was used to perform landslide susceptibility zonation. The analysis results of receiver operating characteristic (ROC) curve pointed out the prediction performance of random forest was 92.8%, better than that obtained from popular artificial neural network (ANN) (81.9%) and support vector machine (84.7%) models. Meanwhile, compared with traditional GIS-based reclassification methods, in the susceptibility zonation map obtained from cluster algorithm, more historical landslides distributed in the high susceptibility zones. Hence, the proposed approach is a promising tool for spatial prediction of landslides at the study area.