[HTT14-P03] GIS-based frequency ratio, logistic regression and artificial neural network models for landslide susceptibility analysis in central Shenzhen, South China
キーワード：地すべり感受性、GIS、ロジスティック回帰、ニューラルネットワーク
As major geological hazards, landslides account for a great number of human casualties and an enormous amount of property loss, and cause significant damage to natural ecosystems and human-built infrastructures, especially in urban areas. Landslide susceptibility analysis is regarded as a semi-quantitative risk assessment and has become an important part in landslide studies and risk management engineering. This study investigates landslide susceptibility in the new district of Shenzhen, South China. Shenzhen is a mountainous area, lying in the East Asia monsoon region, with a subtropical monsoon climate. The district has been rapidly developing so that there are broad construction areas and lots of living people. Some previous studies have shown the results of landslide susceptibility analyses for this area. This study aims to further develop such studies by analyzing landslide susceptibility using three different methods: frequency ratio, logistic regression, and artificial neural network (ANNs) models. The latter includes Back Propagation ANNs and Radial Basis Function ANNs. These models were established using elevation, slope, aspect, lithology, human activities, vegetation, and precipitation as independent variables. Land use data are utilized to represent human activities, and the Normalized Difference Vegetation Index (NDVI) is employed to indicate the level of vegetation coverage. Landslide susceptibility maps of the study area are created based on the three methods to show the relative risk of landslides. There are 429 landslide points and 400 points without landslides in the sample data. For the logistic regression model, the overall accuracy is 0.743, the recall is 0.799, and AUC (area under the receiver operating characteristic curve) is 0.790. For the BPANNs model, the overall accuracy is 0.724, the recall is 0.762, indicating that the two methods have similar high performances. The performance of the frequency ratio model is lower.