R?zne sp?soby hodnotenia úspe?nosti máp zosuvného hazardu: bivaria?ny verzus multivaria?ny ?tatisticky model

摘要

Systematic studies of geological hazard and risks were generated by interest from insurance companies during the 20th century. The first studies were linked to individual building structures and later also to landuse and environmental impact assessment. Data collected were transformed to maps of seismic zonation and landslide hazard maps. The paper is devoted to landslide hazard map quantification and verification. The landslide hazard assessment is based on the assumption that landslides will occur in the future under the same conditions as occurred in the past. In the model area of the Myjava Upland (Western Slovakia) statistical methods - bivariate statistical analysis and conditional multivariate analysis were applied to assess the landslide hazard. The necessity to evaluate the informative value of final maps has arisen recently; practically it means to verify them. In the 80-ties, when the first landslide susceptibility maps were created, they were verified by visual comparison of the prognostic maps with a map of registered slope deformations. Here in, methods of statistical accuracy and ROC (Receiver Operating Characteristic) curves are used for evaluation of both statistical models. 285,004 pixels selected from raster of registered landslides were evaluated and an equal number of pixels randomly selected from raster of landslide hazard map prepared using bivariate statistical analysis? in the case of conditional multivariate analysis, there were 285,030 pixels. The results illustrate that, according to most of the methods of statistical success used to set model performance, both prognostic maps correspond to quality configured statistical models. This comparison shows that the difference between the accuracy of these two approaches has a value of about 5% in favour of multivariate statistical analysis. The difference between the statistical methods represents less than two percent using ROC curves for model verification.