Slovakia: Kysuca, Torysa, Topľa, and Gidra rivers Efficient and accurate prediction of fluvial floods is a necessary task for flood preparedness and risk reduction. This study focuses on predicting the extent of fluvial floods under three flood scenarios (Q 10 , Q 100 , Q 1000 ) using three machine learning models: random forest (RF), extreme gradient boosting (XGBoost), neural networks (NN) and one deep learning model (U-Net). The models were trained using official flood maps, created under the second cycle of EU Flood Directive (2007) implementation, and seven high-resolution physical-geographic and land cover predictors. Model performance was assessed using three metrics and training time. Overall, we studied four river sections in Slovakia (Kysuca, Gidra, Torysa, and Topľa) for model development, with three river sections being used for training and the remaining one for testing and calculating performance measures. Results suggest that transferability of the modeled fluvial flood extent on similarly long and large river sections provided the highest performance. This finding is demonstrated by lower and balanced numbers of FP and FN pixels, specifically, for the Torysa/Kysuca, Topľa/Kysuca or Kysuca/Torysa training/testing river sections. The RF and XGBoost models required the least time for training. The optimized U-Net model resulted in less training time than the NN model. The results have high potential for near real-time flood mapping or operational early warning. • Machine and deep learning were used to transfer fluvial flood extent among rivers. • Models were innovatively trained/tested on official flood maps and seven predictors. • RF, XGBoost, NN, U-Net models and Q 10 , Q 100 , Q 1000 flood scenarios were applied. • The highest performance was achieved on similarly long and large river sections.
Vojtek et al. (Fri,) studied this question.