Answers are grounded in this paper and cite their sources.
This study focuses on estimating fluvial flood extent and flow depth using random forest (RF), extreme gradient boosting (XGB), and neural network (NN) algorithms for a 17.0 km section of the Teplica River and a 3.1 km section of the Gidra River, Slovakia. The machine learning (ML) models were trained on hydraulically derived data using MIKE+ and five scenarios: Q10, Q20, Q50, Q100, and Q1000. Unlike standard pixel-based or randomly sampled approaches commonly used, an innovative ML training approach was applied that integrates river-based longitudinal segmentation with segment-wise balanced sampling, which explicitly preserves hydrological structure during ML training. Additionally, different training/testing splits of river segments were evaluated. The results demonstrate that all ML models are capable of producing acceptable and computationally efficient flood extent and flow depth maps across all the segment splits and flood scenarios. The differences between the ML models in terms of flood extent classification and flow depth estimation were minimal for the respective study domain. The results suggest that the flow depth prediction was more successful in the Gidra domain than in the Teplica domain. The maximum training time was 14.7 min and 18.2 min for ML classification and regression, respectively, which is acceptable for rapid flood inundation mapping.
No takes yet. Share an insight, caveat, or question.
Vojtek et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: