Flash flood susceptibility modelling using functional tree and hybrid ensemble techniques

The present research aims to assess and judge the capability of flash flood susceptibility (FFS) models considering hybrid machine learning ensemble techniques for the FFS assessment in the Gorgan Basin in Iran. Three novel intelligence approaches, namely, bagging–functional tree (BFT), dagging–functional tree, and rotational forest–functional tree are used for modelling, with consideration to 15 flood conditioning factors (FCFs) as independent variables and 426 flood locations as dependent variables. Three threshold-dependent and -independent approaches are used to evaluate the goodness-of-fit and prediction capability of the ensemble models with a single functional tree (FT). These approaches include the area under the receiver operating characteristic curve of the success rate curve (SRC) and prediction rate curve (PRC), efficiency (E) and true skill statistics (TSS). The random forest model is used to determine the relative importance of FCFs. Elevation, stream distance and normalized difference vegetation index (NDVI) have crucial roles in the study area during flash flood occurrences. According to the results of all threshold-dependent and -independent approaches (AUC of SRC = 0.933, AUC of PRC = 0.959, E = 0.76 and TSS = 0.72), the BFT ensemble model has the greatest accuracy in terms of modelling FFS. Results also show that the performance of the FT model is enhanced by three meta-classifiers. The seed cell area index technique is also used to check model classification accuracy and reliability. Results of this technique show that all the models demonstrate good performance and reliability. However, the FFS maps prepared by machine learning ensemble techniques have excellent accuracy and reliability, as per the results of validation methods. Thus, these FFS maps can be used as a convenient tool to reduce the effect of flood in flash flood-prone areas.