Jason N. Goetz

Evaluating machine learning and statistical prediction techniques for landslide susceptibility modeling

Abstract Statistical and now machine learning prediction methods have been gaining popularity in the field of landslide susceptibility modeling. Particularly, these data driven approaches show promise when tackling the challenge of mapping landslide prone areas for large regions, which may not have sufficient geotechnical data to conduct physically-based methods. Currently, there is no best method for empirical susceptibility modeling. Therefore, this study presents a comparison of traditional statistical and novel machine learning models applied for regional scale landslide susceptibility modeling. These methods were evaluated by spatial k -fold cross-validation estimation of the predictive performance, assessment of variable importance for gaining insights into model behavior and by the appearance of the prediction (i.e. susceptibility) map. The modeling techniques applied were logistic regression (GLM), generalized additive models (GAM), weights of evidence (WOE), the support vector machine (SVM), random forest classification (RF), and bootstrap aggregated classification trees (bundling) with penalized discriminant analysis (BPLDA). These modeling methods were tested for three areas in the province of Lower Austria, Austria. The areas are characterized by different geological and morphological settings. Random forest and bundling classification techniques had the overall best predictive performances. However, the performances of all modeling techniques were for the majority not significantly different from each other; depending on the areas of interest, the overall median estimated area under the receiver operating characteristic curve (AUROC) differences ranged from 2.9 to 8.9 percentage points. The overall median estimated true positive rate (TPR) measured at a 10% false positive rate (FPR) differences ranged from 11 to 15pp. The relative importance of each predictor was generally different between the modeling methods. However, slope angle, surface roughness and plan curvature were consistently highly ranked variables. The prediction methods that create splits in the predictors (RF, BPLDA and WOE) resulted in heterogeneous prediction maps full of spatial artifacts. In contrast, the GAM, GLM and SVM produced smooth prediction surfaces. Overall, it is suggested that the framework of this model evaluation approach can be applied to assist in selection of a suitable landslide susceptibility modeling technique.

Modelling Landslide Susceptibility for a Large Geographical Area Using Weights of Evidence in Lower Austria, Austria

Knowledge of the spatial distribution of landslide susceptibility supports and enhances decision-making involved in land-use planning for mitigating the impacts of landslide hazards. Our study focuses on the application of the weights of evidence (WOE) method to statistically model landslide susceptibility for a large area in the Austrian province of Lower Austria. Approximately 16,000 km2 was modelled with a 10 m × 10 m spatial resolution. To complete this task, a new implementation of WOE in R was developed, a free open source software for statistical computing, to handle the computation of weights for large geospatial datasets. Furthermore, the challenge of modelling diverse landslide conditions for a large area was addressed by modelling WOE separately for each lithology unit. The final susceptibility map was compiled by mosaicking these models. The performances of the models were estimated with a repeated cross-validation approach that measured the area under the receiver operating characteristic curve (AUROC). The results showed good WOE model performances; the median AUROC values ranged from 73 to 93 %, with an average performance of 86 % for the entire study area. Also, this study demonstrated the successful application of WOE for a large geographic area with a high spatial resolution.

Analysis of isotopic signals in the Danube River water at Tulln, Austria, based on daily grab samples in 2012

Results of stable isotope measurements (δ2H, δ18O) of daily grab samples, taken from the Danube River at Tulln (river km 1963) during 2012, show seasonal and short-term variations depending on the climatic/hydrological conditions and changes in the catchment area (temperature changes, heavy rains and snow melt processes). Isotope ratios in river water clearly reflect the isotopic composition of precipitation water in the catchment area since evaporation influences play a minor role. Average δ2H and δ18O values in 2012 are−78‰ and−11.0‰, respectively, deuterium excess averages 10‰. The entire variation amounts to 1.8‰ in δ18O and 15‰ in δ2H. Quick changes of the isotopic composition within a few days emphasise the necessity of daily sampling for the investigation of hydrological events, while monthly grab sampling seems sufficient for the investigation of long-term hydro-climatic trends. 3H results show peaks (half-width 1–2 days, up to about 150 TU) exceeding the regional environmental level of about 9 TU, probably due to releases from nuclear power plants.

Indicators for Earthquake-Induced Soil Slides in the Flatlands of an Alpine Fringe Area

In Austria’s most eastern province Burgenland, an ongoing study of the regional distribution of mass movements has been performed. In this alpine fringe area more than 280 previously unrecognized active mass movements were mapped, most of them are identified as creep or soil slides. We assume that the majority of these landslides were triggered by intense rainfall, but we also found strong indications that at least two large scale movements might be earthquake-induced. These two movements, each covering an area of more than 1 km2, are situated in the north of the province. In other parts of the province the dimension of mapped landslides averages only 0.03 km2. Compared to the other areas, the hill slopes in the north have the lowest steepness and the lowest annual precipitation. Underground geologic conditions, however, are overall comparable. Earthquakes are only documented in the north of the province and they could have acted as additional trigger for these comparatively large scale landslides with their noticeable different pattern in dimension and volume.