Bayesian networks model for identification of the effective variables in the forecasting of debris flows occurrence

Abstract

Comprehensive assessment of debris flow hazards is a challenging issue due to the complexity and uncertainty of its factors. For this reason, the practical forecasting of debris flows requires developing a reliable and realistic forecasting model. In this paper, a Bayesian Networks (BNs) model is proposed for identification of debris flows events in the northern basins of Iran. BNs model illustrates the uncertainty of the results as probability percentage of debris flow occurrence in different categories (non-occurrence, occurrence with low-intensity and occurrence with high-intensity). In this research, average basin elevation, average basin slope, watershed area, the current rainfall, antecedent rainfalls of 3-day ago and discharge of 1-day ago were used as the predictor variables. Moreover, K-means clustering method was applied in modeling by the BNs model. To identify the effective predictors in debris flow occurrence, sensitivity analysis was performed. For this purpose, scenarios which employ various predictor variables were tested. The scenario which uses all predictor variables has a forecasting accuracy of 91%. This scenario was selected as the best scenario. However, a scenario which employs only effective predictors also proposed for practical uses. The results of the various forecasting scenarios showed that average basin elevation, watershed area, current rainfall and discharge of 1-day ago are the effective predictors in the forecasting debris flows. The BNs model may be proposed for future tests in the other debris flow prone regions.