An intelligent VMD-WTC-GRU hybrid framework with uncertainty quantification for forecasting extreme flood events in semi-arid regions

Flood prediction in data-scarce semi-arid regions presents significant challenges. This study develops an innovative artificial intelligence (AI) framework for the Jamash watershed, utilizing comprehensive daily data (2000–2023) of meteorological, hydrological, and remote sensing variables. We evaluated two feature selection methods—Random Forest (RF) and Wavelet Transform Coherence (WTC)—and employed Variational Mode Decomposition (VMD) to process non-stationary time series. Twelve modeling scenarios compared Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) architectures, both standalone and hybrid forms, with AI-based uncertainty assessment. Results demonstrated the superior performance of the hybrid VMD-WTC-GRU model, achieving a determination coefficient of 0.83 during testing. Uncertainty analysis confirmed this model as both the most accurate and reliable, exhibiting the narrowest uncertainty band (32.26) with satisfactory confidence probability coverage (0.95). This integrated AI framework effectively overcomes challenges of limited data and hydrological complexity in flood forecasting.