Flash flood susceptibility mapping using stacking ensemble machine learning models

The objective of the present study was to introduce a novel methodological approach for flash flood susceptibility modeling based on a stacking ensemble (SE) model. Two SE models, Random Forest (RF) and Artificial Neural Network (ANN) were developed, whereas LDA, CART, LR, k-NN and SVM were the basic models of the two SE models. The performance of the developed methodology was evaluated at the Island of Rhodes, Greece. The database included 54 flash floods locations and 14 flood-related parameters. The SE-RF model produced slightly higher predictive results, in terms of accuracy (0.844), kappa index (0.687) and the area under the receiver operating characteristic curve (0.870), followed by the SE-ANN with values of 0.812, 0.625 and 0.773, respectively. Overall, the study provides evidence about the higher accuracy SE models can achieve since they are capable of combining in an intelligent way a number of weak predictive models.