Regional landslide susceptibility assessment based on relevance vector machine

The main goal of this study is to assess regional landslide susceptibility using Relevance Vector Machine (RVM) at hilly area of Sichuan province, China. According to the developing features and distribution regularities of the landslide in the hilly area, this paper chooses eight disaster pregnant environmental factors (including altitude, lithology, slope, river distribution, etc) as input feature set of RVM and Support Vector Machine (SVM). And then, combined with GIS spatial analysis technology, these two models were used for regional landslide susceptibility mapping which shows that the map of RVM has better performance. For further proving of RVM model, both Receivers Operating Characteristic (ROC) curve and dot density of the landslide were used. For one thing, the area under ROC curve (AUC) of RVM and SVM reach 0.876 and 0.875 respectively, but the Root-mean-square Error (RMSE) of RVM (0.406) is slightly better than SVM (0.422). The results indicate that the predicted results of RVM have stronger discriminating ability for explaining landslide state than SVM. For another, this study uses the natural breaks classification method to divide the study area into four areas, which are general susceptibility area, slight susceptibility area, middle susceptibility area and high susceptibility area. The result that the sum of middle and high area predicted by RVM make up 50.69% of the total area but collect 68.27% of landslides which is better than SVM whose proportion of area and landslide are 51.79% and 59.73%. It means that the dot density of the landslide of RVM is more close to the actual situations. So, in conclusion, the RVM model is able to assess regional landslide susceptibility and the results of assessment will be useful for disaster early-warning. © 2015 by Binary Information Press.