The nature of statistical learning theory
The nature of statistical learning theory
Neural Computation
Expert Systems with Applications: An International Journal
| Hi-index | 0.00 |
The main purpose of this study is to compare the following six GIS-based models for susceptibility mapping of earthquake triggered landslides: bivariate statistics (BS), logistic regression (LR), artificial neural networks (ANN), and three types of support vector machine (SVM) models that use the three different kernel functions linear, polynomial, and radial basis. The models are applied in a tributary watershed of the Fu River, a tributary of the Jialing River, which is part of the area of China affected by the May 12, 2008 Wenchuan earthquake. For this purpose, eleven thematic data layers are used: landslide inventory, slope angle, aspect, elevation, curvature, distance from drainages, topographic wetness index (TWI), distance from main roads, distance from surface rupture, peak ground acceleration (PGA), and lithology. The data layers were specifically constructed for analysis in this study. In the subsequent stage of the study, susceptibility maps were produced using the six models and the same input for each one. The validations of the resulting susceptibility maps were performed and compared by means of two values of area under curve (AUC) that represent the respective success rates and prediction rates. The AUC values obtained from all six results showed that the LR model provides the highest success rate (AUC=80.34) and the highest prediction rate (AUC=80.27). The SVM (radial basis function) model generates the second-highest success rate (AUC=80.302) and the second-highest prediction rate (AUC=80.151), which are close to the value from the LR model. The results using the SVM (linear) model show the lowest AUC values. The AUC values from the SVM (linear) model are only 72.52 (success rates) and 72.533 (prediction rates). Furthermore, the results also show that the radial basis function is the most appropriate kernel function of the three kernel functions applied using the SVM model for susceptibility mapping of earthquake triggered landslides in the study area. The paper also provides a counter-example for the widely held notion that validation performances of the results from application of the models obtained from soft computing techniques (such as ANN and SVM) are higher than those from applications of LR and BA models.