Answers are grounded in this paper and cite their sources.
Study region Carlisle, a flood-prone city in northwestern England within the River Eden catchment. Study focus Flood inundation forecasting with sufficient lead time and spatial detail is essential for proactive flood risk management, existing approaches remain focused on nowcasting or simplified outputs. This study develops a unified deep learning framework that couples upstream flow forecasting with inundation simulation, enabling grid-scale prediction of flood depth up to 4 h ahead. The model integrates a gated recurrent unit (GRU)-based flow forecasting module with a spatiotemporal inundation module combining convolutional layers, recurrent units, and multi-head attention (MHA). Model performance is evaluated based on the LISFLOOD-FP model using leave-one-out cross-validation, event-based analysis, multi-stage forecasting, and stress test. New hydrological insights for the region Results highlight heterogeneous flood responses among three upstream inflows in discharge magnitude, peak timing, and rising-limb characteristics. Performance declines at longer lead times, particularly for the Caldew River inflow. For the 2005 flood event, flow forecasting component reproduces upstream hydrographs with Nash–Sutcliffe efficiency (NSE) values exceeding 0.97 at short lead times, while inundation forecasting achieves mean root mean square error (RMSE) of 0.24 m and NSE of 0.91 against the LISFLOOD-FP reference. Across stages, RMSE ranges from 0.06 to 0.30 m and NSE remains mostly above 0.95, with larger errors in transition and core-flooded areas and along the Eden River during rapid evolution. A 4-hour inundation forecast is generated in approximately 80 s on a GPU, compared with 36 min for the physics-based LISFLOOD-FP.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: