Tornyai R., Bednarik M. & Havlín A., 2016: Application of neural network to assess landslide hazard and comparison with bivariate and multivariate statistical analyses. Acta Geologica Slovaca, 8, 1, 109–118.

Application of neural network to assess landslide hazard and comparison with bivariate and multivariate statistical analyses

Rudolf Tornyai¹, Martin Bednarik¹ & Aleš Havlín²

¹Department of Engineering Geology, Faculty of Natural Sciences, Comenius University in Bratislava, Ilkovičova 6, 842 15 Bratislava; tornyai@fns.uniba.sk
²Czech Geological Survey, Branch Brno, Leitnerova 22, 658 69 Brno, Czech Republic

Abstract

Landslide hazard in the Žilina area in northern Slovakia is assessed using neural network analysis. Four input parameters are evaluated, they are presented as a result of statistical processing in the form of parametric maps. Statistical evaluation was executed in ArcGIS environment; neural network was calculated in Matlab. The output of this study is a prognostic landslide hazard map. Further, the result was compared with the hazard map created using bivariate and multivariate statistical analyses through ROC curves. Area below curve (AUC) calculated from ROC curve shows accuracy of individual models. It can be stated that the NN´s AUC is equal to 0.924, what represents the rate of success 92.4%; bivariate multivariate analyses AUC is equal to 0.852 and 0.919.

Key words: landslide hazard assessment, neural network, bivariate analysis, multivariate conditional analysis, Žilina region, ROC curves

Manuscript received: 2016-03-16

Revised version accepted: 2016-05-04