Abstract :
Now, there are many methods that have been used in landslide susceptibility analysis, but they all have some aspects need to be improved. Random forests methodology improves the accuracy of the model by aggregating multiple models. Especially when dealing with large data, it shows strong robustness. So, we plan to apply random forests methodology to landslide susceptibility analysis triggered by earthquakes. We made Lushan and its surrounding areas as our study area, which suffered from the earthquake in April 20, 2013. This area is located in fault zone in the Longmen Mountains, it shows guiding significance for the study of seismic landslide in southwest China. Based on seismic landslide physical mechanics, we chose slope, aspect, fault, river, Normalized Difference Vegetation Index (NDVI), waviness, lithology, seismic intensity and elevation as landslide factors. Then, we built the suitable seismic landslide model based on Random Forests. After that, we used Out-of-Bag estimates (OOB) to calculate the generalization error of our model, and we also used Receiver Operating Characteristic curve (ROC) error evaluation system to estimate the correctness of the model. When the number of sample data is greater than 50, the OOB generalization error result is less than 0.08, and the area under the ROC curve was 0.938 which means the model has a high reliability. Through this research we found that the random forests methodology showed a good performance when dealing with seismic landslide studies and should be spread to related research.
