Delineating flood-prone areas using advanced integration of reduced-error pruning tree with different ensemble classifier algorithms

Floods are the most frequent of extreme events and they can easily impact lives and property. Flood mapping and mitigation are prohibitively expensive and time-consuming. The need for accurate maps of flood probability is vital and meeting that need is a challenge. This study presents an effective flood-probability mapping framework that compares one standalone machine-learning model, the reduced-error pruning tree (REPT), to four hybrid models using meta-classifiers: Bagging (BA-REPT), Dagging (DA-REPT), AdaBoost (AB-REPT), and LogitBoost (LB-REPT). Nine continuous and categorical variables that reflect the conditions that influence flood probabilities—slope, elevation, aspect, topographic wetness index (TWI), distance to river, land use, lithology, rainfall, and valley depth—were used to prepare flood-probability maps of a catchment in western Isfahan Province, Iran. To train the five algorithms, 108 flood events were mapped in the catchment, and these locations were randomly separated into two subsets at a ratio of 70:30. The data (comprised of the nine flood-conditioning variables) for the larger portion (70% of cases) were used to build the models and the remaining 30% of cases were used to test them. The performances of the models were evaluated with several statistical and error tests (TP, TN, FP, FN, accuracy, sensitivity, and specificity), receiver operating characteristic (ROC) curve, and the area under the ROC curve (AUC). The results show that the hybrid (ensemble) methods boosted accuracy performance above the standalone REPT by between 2.56 and 15.38%. AUC statistics indicated that the predictive powers of models were REPT = 0.78, LB-REPT = 0.80, DA-REPT = 0.83, BA-REPT = 0.85, and AB-REPT = 0.90.