No abstract

The objective of this study is the assessment of potential failure zones of landslides in unstable areas. For this purpose, two different stochastic classification models were used: A boosted decision tree approach with TreeNet (TN), and a bagging decision tree approach with Random Forests (RF). Both topographic and soil parameters were considered as predictor variables for training and testing the models. We assume that several predictor variables will lead to misclassification and incorrectness, especially soil parameters. Hence, the misclassification of these particular predictors should be avoided, using the strategy of tree boosting. The investigated area is the hydrological basin of Vernazza in Cinque Terre, Northwest Italy. A disastrous flash flood on the 25 th of October 2011 with numerous landslides caused fatalities and economic losses amounting to millions of Euros. We mapped landslide areas in the field and checked the resulting maps with high resolution remote sensing images. Furthermore, the relevant soil parameters were collected based on a geostatistical approach. We measured topographic parameters, and physical and hydrological soil characteristics such as maximum shear strength under saturated and unsaturated conditions, and hydraulic conductivity (Ksat), and attributed random points in three distinguished classes: i) initiation areas, representing the most likely failure areas for possible landslides, ii) transport areas which were considered as a mix of classes 1 and 3, and iii) stable areas, such as valley bottom, ridges, and unconditionally stable areas. We ran both models with a training dataset (0.8 of the total points Ntot) and a test dataset (0.2 of Ntot) and each with 2000 grown decision trees. We validated the models with a Receiver Operating Characteristic (ROC) curve integral. The regionalized results of the TreeNet dataset yielded potential susceptible landslide areas of a total area of 1.74 km², which is 29.74% of the total area. In contrast, the Random Forests model classified a much greater susceptible area (84.27% of the total area). The results show that Treenet is outperforming RF. The latter misclassifies especially the soil related variables, whereas TreeNet yields robust model results.

No abstract

No abstract