Flood susceptibility estimation using randomization-based machine learning models. A case study at the Putna river basin, Romania

Abstract

Floods represent the natural hazards that generate the most damage at the international level. A very important stage in the flood risk management activity is the mapping of areas susceptible to these hazards. In this context, in the present study, the following 3 hybrid models were applied to determine flood susceptibility in Putna river basin, Romania: Randon Committee-Weights of Evidence (RC-WOE), Random SubSpace-Weights of Evidence (RSS-WOE) and Randomizable Filtered Classifier–Weights of Evidence (RFC–WOE). 14 flood predictors and 192 flood locations (divided into 70% training sample and 30% validating sample) were used as input data in the 3 models. The applied models confirmed the fact that the most important flood predictors are: slope angle, distance from rivers and elevation. At the same time, around 24% of the study area shows a high and very high susceptibility to floods. The ROC Curve method along with other statistical metrics, used to validate the applied models, showed that the accuracy of the models generally exceeded 80%, which represents a very good performance. The obtained results provide useful information for the authorities responsible for reducing the flood risk. Also, the future planning of the territory can obviously take into account the zoning of flood susceptibility.