Predicting the Spatial Distribution of Rain-Induced Shallow Landslides by applying GIS and Geocomputational Techniques: A Case Study from North East India

Abstract

This study presents a case of statistical modelling, by applying GIS and geocomputational techniques, to predict areas that are susceptible to future rain-induced shallow landslides. The statistical prediction model is based on the observed relationships between the spatial distribution of past landslideevents and environmental (causal) factors that are associated with such phenomena. The study also evaluates the predictive performance of a nonlinear regression model, namely the Generalized Additive Model(GAM),applied for the analysis. The study area comprises a residual hill of ? 6 Km2 area situated in the heart of Guwahati (capital city of Assam in NE India). We exploited the geoprocessing functions of SAGA GIS to derive nine different terrain attributesfrom a digital elevation model (DEM) processed by synthetic aperture radar interferometry (InSAR). The terrain attributes along with land use classes, in raster grid format, constitute the predictor variables. An inventory of the locations of eighty-two past occurrences of shallow landslide events constitutes the response. We performed the modelling and statistical geocomputation entirely in the open-source R language and software environment. The procedure comprises the following three steps: (1) Collinearityanalysis to discard redundant predictors. (2) 100-fold bootstrap resampling to fit the GAM by a random selection of 2/3 of the landslide pixels ("training" subset) and validate the GAM by the remaining 1/3 ("test" subset). (3) Estimate model accuracy (true error rates) by a repeated 100-fold 'hold-out validation' method and evaluate the predictive performance of the model by the Area under the ROC curve (AUROC) computed for 100 independently trained models. The mean and standard deviation of accuracy on training sets are 0.80 and 0.01, and that on test sets are 0.79 and 0.02 respectively. The AUROC corresponding to the meanof landslide probabilities is 0.87, and that of the 95% Confidence Intervals (CI) is between 0.86 and 0.88. Thevalues of these quality measures indicate that a data-driven model, such as the GAM, is efficient regarding its predictive performance, to highlight the unstable areas in the study area. We subsequently used the mean values of the landslide probability (susceptibility) estimates corresponding to each mapping unit (grid cell) to construct the landslide susceptibility map, which can be used for land use planning and hazard mitigation.