Evaluating the river-to-river transferability of deep learning-based fluvial flood extent predictions

Fluvial floods are one of the most common types of flooding worldwide. Therefore, accurate flood prediction is essential for effective flood preparedness and risk management. This study investigates the prediction of fluvial flood extent under three flood scenarios (Q₁₀, Q₁₀₀, and Q₁₀₀₀) using deep learning (DL), in particular, the U-Net model. The U-Net model was trained on official flood maps, created as part of the second cycle of the EU Flood Directive (2007), along with seven high-resolution predictors derived from the LiDAR DEM (1 m resolution), orthophotos (20 cm resolution), and ZBGIS spatial database: slope, stream power index (SPI), topographic wetness index (TWI), height above the nearest drainage (HAND), distance from river, roughness, and normalized difference vegetation index (NDVI). Multicollinearity among predictors was tested using the Pearson correlation and Variance Inflation Factor (VIF) with thresholds for Pearson correlation ≤0.7 and VIF ≤5. The model performance was evaluated using three quantitative metrics (Recall, Precision, and F1-score) and training time. The study focused on four river sections in Slovakia (Kysuca, Gidra, Torysa, and Topľa). In each U-Net application, three sections were used for training and one for testing and performance evaluation. The results indicate that the highest model performance was achieved when predicting flood extents on river sections that were similar in width and length. This was particularly evident in the cases where the training/testing river sections included combinations Torysa/Kysuca, Topľa/Kysuca or Kysuca/Torysa. When DL models were trained on narrow/short river sections and then tested on wider/longer sections, the number of false negative (FN) pixels tends to be high. Conversely, when these models are trained on wider/longer river sections and tested on narrow/short ones, the number of false positive (FP) pixels increases. Based on these findings, we recommend avoiding both of these training/testing strategies when transferring the prediction of fluvial flood extent across distinct river sections. Furthermore, the optimized U-Net model showed relatively fast training times with the maximum equal to 15 minutes. The findings of this study suggest strong potential for near-real-time or even real-time flood mapping and operational use. Acknowledgment: Funded by the EU NextGenerationEU through the Recovery and Resilience Plan for Slovakia under the project No. 09I03-03-V03-00085.