Evaluating River-Level Rate of Change as a Training-Data Selection for LSTM Flash Flood Forecasting in UK Catchments

Flash floods are challenging to forecast with data-driven models, as their rarity in hydrological records produces severe class imbalance during training. To mitigate this problem, we propose an event-based training data selection method that uses the rate of change (RoC) of river levels as a filtering criterion and assess how the choice of RoC threshold affects forecasting performance across three UK catchments. For each catchment, a cascaded LSTM architecture, which is a rainfall model whose forecasts feed into a river-level model, is trained on event datasets filtered at a range of RoC thresholds and evaluated against the 20 highest-magnitude testing events per site. The results reveal a trade-off in the use of RoC thresholds. Higher thresholds, which filter the training data toward rapid-rise events, improve the forecasting of high-magnitude events at the two smaller, faster-responding catchments, where the median RMSE improves from 0.18 m at the optimum low-magnitude RoC threshold to 0.05 – 0.09 m at the optimum high-magnitude threshold. Lower thresholds produce more accurate forecasts of low-magnitude events at all three sites; however, no single threshold is optimal across all ranges of event magnitudes. Cross-catchment comparison shows that the optimal threshold depends on both catchment characteristics and the target event magnitude. The improvement in high-magnitude forecasting at higher thresholds is substantial in the smaller, flashier catchments but weak in the larger, slower-responding areas. Higher RoC-based event filtering is therefore most effective for high-magnitude forecasting in small, fast-responding catchments.