Improvement of Best First Decision Trees Using Bagging and Dagging Ensembles for Flood Probability Mapping

Development of zoning and flood-forecasting models is essential for making optimal management decisions before and after floods. The Komijan watershed of Markazi Province, Iran is often affected by floods that have caused great material damage and loss of life. The main objective of this study is to use a new machine-learning method to create three models: best-first decision tree (BFT), a bagging best-first decision tree (BBFT) ensemble and a dagging best-first decision tree (DBFT) ensemble to spatially predict flood probability. Twelve conditioning-factor measures for 272 locations of past floods were used to train and test three models. Receiver operating characteristic (ROC), positive predictive value (PPV), negative predictive value (NPV), sensitivity (SST), specificity (SPF), accuracy (ACC), kappa (K), and root mean square error (RMSE) are applied to compare and validate the models. The results are that all three models performed well in mapping, flood probabilities (AUC > 0.904). The BBFT model was best, however, with an AUC = 0.96. Based on the results of the Relief-F attribute evaluation method, two soil and slope factors were weighted highest among the parameters, indicating that they are the most important flood-conditioning factors. These models may improve identification of zones that are most susceptible to flooding, improving the capacity for risk management and providing more detailed information for managers and decision-makers.