GIS-based spatial modeling of snow avalanches using four novel ensemble models

Snow avalanches can destroy lives and infrastructure and are very important phenomena in some regions of the world. This study maps snow avalanche susceptibility in Sirvan Watershed, Iran, using a new approach. Two statistical models – belief function (Bel) and probability density (PD) – are combined with two learning models – multi-layer perceptron (MLP) and logistic regression (LR) – to predict avalanche susceptibility using remote sensing data in a geographic information system (GIS). A snow avalanche inventory map was generated from Google Earth imagery, regional documentation, and field surveys. Of 101 avalanche locations, 71 (70%) were used to train the models and 30 (30%) were used to validate the resulting models. Fourteen snow avalanche conditioning factors were used as independent variables in the predictive modeling process. First, the weight of Bel and PD techniques were applied to each class of factors. Then, they were combined with two MLP and LR learning models for snow avalanche susceptibility mapping (SASM). The results were validated using positive predictive values, negative predictive values, sensitivity, specificity, accuracy, root-mean-square error, and area-under-the-curve (AUC) values. Thus, the AUCs for the PD-LR, Bel-LR, Bel-MLP, and PD-MLP hybrid models are 0.941, 0.936, 0.931 and 0.924, respectively. Based on the validation results, the PD-LR hybrid model achieved the best accuracy among the models. This hybrid modeling approach can provide accurate and reliable evaluations of snow avalanche-prone areas for management and decision making. [Display omitted] •Predicting of snow avalanche susceptibility using an ensemble of statistical and learning models•All models have very good accuracy (AUC > 90%) in predicting snow avalanche susceptibility.•EBF and PD methods were used to weight the predictive factors.•PD-LR showed practical and robust results compared to the other models.