Flash-flood Potential Index mapping using weights of evidence, decision Trees models and their novel hybrid integration

摘要

Flash-floods are among the natural risk phenomena that annually cause important material damages and losses of human lives worldwide. One of the main activities for mitigating the negative effects of these phenomena consist of the identification and spatial representation of the surfaces prone to surface runoff occurrence. Flash-Flood Potential Index (FFPI) is the common method used to assess the degree of susceptibility to flash-floods of a certain surface. The main drawback of the aforementioned method is represented by the fact that, in the majority of the studies, the geographical factors considered for FFPI calculation received equal weights, even if they do not influence in the same measure the surface runoff process. Moreover, within the methodologies developed in the previous studies, the areas affected by torrential phenomena have not been considered for FFPI computation. To address these shortcomings, in the present study, FFPI values are estimated by using a number of 4 stand-alone models (Alternating Decision Tree (ADT), Rotation Forest, Weights of Evidence (WOE), Logistic Model Tree) and 3 hybrid models generated by the integration of WOE model with each of the other decision tree-based algorithms. The first stage of the study consisted of the inventory of the areas where torrential phenomena occurred in the past, 70% of them being included in the training sample, while the others 30% in the validating sample. Further, 12 flash-flood conditioning factors, selected through the correlation-based feature selection algorithm, were used to train the 7 models applied for FFPI calculation. The results of the 7 models revealed that the surfaces with a high and very high flash-flood susceptibility occupy between 23.3 and 43.7% of the entire study zone. The ROC Curve method was involved in the models performance assessment and in the results validation procedure. From this point of view, the best results were obtained by the ADT-WOE hybrid model.