Flash flood forecasting in North East England through weak label-guided mixture of experts with multi-scale explainability

Study region: Five fast-response catchments in Northumberland, North East England, along the River Tyne and its tributaries. Rainfall and water-level records at 15-minute resolution from the UK hydrometric archive are used, covering 2016–2025 for most sites and 2022–2025 for Stocksfield. Study focus: We develop a Weak Label–Guided Mixture of Experts (WL–MoE) framework for 32-step water-level forecasting in flash-flood conditions. Historical water level, historical rainfall, and future rainfall forcing are transformed using continuous wavelet transforms and routed by a soft gating network to specialised convolutional experts. Weak labels from DTW-based time-series clustering guide early expert specialisation, and macro-level expert profiling together with Grad-CAM provide multi-scale interpretability. New hydrological insights for the region: Across the five Northumberland catchments, WL–MoE improves the mean Nash–Sutcliffe efficiency from about 0.83 to 0.90 on the full test set and from about 0.39 to 0.73 on the high-water subset relative to the strongest baseline models. The largest gains occur during rapidly rising high-water events, indicating that flood-onset and recession dynamics in these flashy rural catchments are better represented by specialised experts than by a single stationary model. The smooth soft-gating behaviour further suggests that hydrological states in the region are transitional rather than sharply separated, supporting more reliable and interpretable flood forecasting. • Weak-label guided Mixture-of-Experts improves flood peak forecasting accuracy. • Soft gating enables expert specialisation under severe regime imbalance. • Multi-horizon water-level forecasts outperform single expert baselines. • Interpretable routing and saliency reveal influential hydrological regimes. • Supports more reliable flood warnings for regional authorities in Northumberland.