Novel ensemble machine learning models in flood susceptibility mapping

Abstract

The research aims to propose the new ensemble models by combining the machine learning techniques, such as rotation forest (RF), nearest shrunken centroids (NSC), k-nearest neighbour (KNN), boosted regression tree (BRT), and logitboost (LB) with the base classifier adabag (AB) for flood susceptibility mapping (FSM). The proposed models were implemented in the central west coast of India, which is vulnerable to flood events. For flood inventory mapping, a total of 210 flood localities were identified. Twelve effective factors were selected using the boruta algorithm for FSM. The area under the receiver operating characteristics (AUROC) curve and other statistical measures (sensitivity, specificity, accuracy, kappa, root mean square error (RMSE), and mean absolute error (MAE)) were employed to estimate and compare the success rate of the approaches. The validation results of the individual models in terms of AUC value were AB (92.74%) >RF (91.50%) >BRT (90.75%) >LB (89.07%) >NSC (88.97%) >KNN (83.88%), whereas the ensemble models showed that the AB-RF (94%) was of the highest prediction efficiency followed by, AB-KNN (93.33%), AB-NSC (93.02%), AB-LB (92.83%), and AB-BRT (92.64%). The outcomes of the ensemble models established that the AB is more appropriate to increase the accuracy of different single models. Therefore, this study can be useful for proper planning and management of the study area and flood hazard mapping in alike geographic environment.

Keywords:

• Flood hazard
• geographical information system
• boruta approach
• adabag
• central west coast of India

Acknowledgements

We acknowledge the support from the University Grant Commission (3160 NET-June 2015) to the first author. The authors are thankful to the Director CSIR-NIO for encouragement from time to time. The authors are also grateful to the anonymous reviewers for their critical comments and constructive suggestions to improve the manuscript. Field support from the survey team members is thankfully acknowledged. The NIO contribution number is 6674.

Disclosure statement

All authors are declared no actual or potential conflict of interest including any financial, personal or other relationships with other people or organizations.