Flood alerts gain a 90% band: quantile LSTM ensemble flags risk across Australian catchments
Synopsis
Flood warning for Australian catchments now comes as a 90% prediction band rather than a single flow curve: three independently trained LSTMs give the mean flow and the 5th and 95th percentiles, and the band's position against a 15% annual exceedance threshold becomes a high, moderate or low risk label. On CAMELS-AUS, the q=0.05 Quantile-LSTM reached the lowest RMSE of 0.0801 in South Australia against 0.0963 for a conventional LSTM, while peak timing error across states ran about 1.94 to 3.12 days with peaks underestimated by roughly 30%.
Interpretation
The forecast output changes from a single streamflow value to an interval whose position maps onto a risk level. Earlier ensemble and regional deep learning models targeted streamflow accuracy rather than the rare high-magnitude events that matter most for flood warning. In the South Australia comparison, the q=0.05 Quantile-LSTM reached the lowest RMSE at 0.0801, against 0.0963 for the conventional LSTM and 0.0976 for the CNN; the q=0.95 model performed best on SER 1% and 2% windows of extreme flow.
Three architecturally identical but independently trained LSTMs fit the mean flow and the upper and lower quantiles, writing uncertainty into the forecast explicitly. Quantile regression estimates conditional quantiles rather than the conditional mean and is more robust to outliers in the response, which suits extrapolation to extremes. Models were trained on the 222 catchments of CAMELS-AUS with daily cumulative precipitation, daily minimum and maximum air temperature and sine-cosine time of year as time series, plus static attributes such as mean streamflow and streamflow sensitivity to rainfall.
Comparing the predicted interval with a historical flood threshold yields a risk label that can be read directly. The threshold is the 15% annual exceedance probability of annual maximum flow, roughly a 7-year recurrence interval, and any flow above it counts as a flood event. Across states, empirical coverage of the 90% interval was 0.77 in New South Wales, 0.76 in Queensland and 0.83 in Tasmania, close to the nominal 0.90 target, while interval skill scores stayed above 0 but below 1.
Perspective
The evaluation covers Australian catchments in CAMELS-AUS from 1990 to 2020, and the cross-state part uses only the five highest runoff-ratio catchments per state, so reported accuracy is biased toward high-runoff basins. Forecasts extend up to five days ahead, and the individual-catchment strategy requires retraining for each new gauging station.
The model predicts about half as many extreme events as observed in most states, and detection is low overall, highest in the Northern Territory at 0.27 and 0.169 in Tasmania; detected peaks are underestimated by roughly 30%, with Victoria at −42.57% bias and a detection rate of 0.08. Empirical coverage of the 90% interval falls slightly below 0.90 in most states, worst in Western Australia, observed flows cross the predicted band in some basins, and the lower-quantile model is biased toward zero in low-flow basins.