Spatial Prediction of Landslide Hazard at the Yihuang Area (China): A Comparative Study on the Predictive Ability of Backpropagation Multi-layer Perceptron Neural Networks and Radial Basic Function Neural Networks

The aim of this study is to investigate potential applications of multi-layer perceptron neural networks (MLP Neural Nets) and radial basis function neural networks (RBF Neural Nets) for landslide susceptibility mapping in the Yihuang area (China). First, a landslide inventory map with 187 landslide locations was generated, and then the map was randomly partitioned into a ratio of 70/30 for training and validating models. Second, 14 landslide conditioning factors (slope, altitude, aspect, topographic wetness, sediment transport index (STI), stream power index (SPI), plan curvature, distance to rivers, distance to faults, distance to roads, land use, normalized difference vegetation index (NDVI), lithology, rainfall) were prepared. Using MLP Neural Nets and RBF Neural Nets, two landslide susceptibility models were constructed and two landslide susceptibility maps were generated. Finally, the two resulting landslide susceptibility maps were validated using the landslide locations and the receiver operating characteristic (ROC) method. The validation results showed that the areas under the ROC curve (AUC) for the two landslide models produced by MLP Neural Nets and RBF Neural Nets are 0.932 and 0.765 for success rate curve and 0.757 and 0.725 for prediction rate curve, respectively. The results showed that the MLP Neural Nets model is better than the RBF Neural Nets model in this study. The results may be useful for general land use planning and hazard mitigation purposes.