Turning the Open ECMWF Ensemble into Admin-Level Flood Risk at Metadata Cost: A Deployed Anticipatory-Action System for East Africa
Abstract
East Africa is a hotspot for hydrometeorological extremes, and the ensemble forecasts that could underpin anticipatory action there are already free. The European Centre for Medium-Range Weather Forecasts (ECMWF) publishes its 51-member Integrated Forecasting System (IFS) ensemble as openly licensed GRIB2 on public object storage, yet it is out of reach for most national and humanitarian analysts, because scanning it demands compute, egress, and engineering that few such agencies can absorb. We present a deployed, cloud-native pipeline that turns this archive into daily administrative-unit flood-risk guidance without copying a single forecast field. Byte-range manifests expose roughly 1,250 forecast days as one virtual dataset of 18.5 GB of metadata, about 13,000x smaller than a copy of the referenced fields, and cut the exceedance stage from 22 to 1 minute per forecast day. Exceedance probabilities are scored against generalized extreme value (GEV) return levels stratified by season and by the El Nino Southern Oscillation (ENSO) and Indian Ocean Dipole (IOD) phase, and a first-order dynamic Bayesian network fuses forecast tails, satellite antecedent wetness, and catchment routing across 238 admin-1 units. On the 2024 short rains, scored against independent satellite flood observations, catchment routing raises recall at the two highest alert levels from 0.29 to 0.51 at unchanged precision (0.96). The panel is strongly imbalanced (91 positives, 14 negatives), so we report operating-point counts with interval estimates and alert-rate-matched paired tests rather than a threshold-free AUC whose 95% interval spans 0.61 to 0.86. Marginal cost is on the order of $0.02 per forecast day.