MENA · cyclones & flash floods
Global forecasts, local answers
A global model can tell you where a cyclone is going. It cannot tell a wilayat whether to close a wadi crossing tonight. This work is about that last step — turning cyclone forecasts into catchment-level decisions for the arid basins of the Arabian Peninsula.
Cyclone Shaheen, Oman — October 2021
A decision map replaying a real event: the storm’s track into Al Batinah, the AI forecast scenarios issued at each cycle, the impact score of every sub-basin, and the rainfall recorded at each wilayat. Drag the timeline and watch the forecast converge on what actually happened.
Opens full screen. Needs an internet connection for the basemap and satellite layers.This is a historical reconstruction of Shaheen (2021) for demonstration — not a live forecast service.
What the demo shows
A forecast you can interrogate
Playable event timeline
Step from 48 hours before landfall to 12 hours after. Every step re-issues the forecast from the storm’s real position at that moment — exactly how operational cycles behave.
Forecast vs. reality, side by side
The AI scenarios are drawn against what actually happened, with the landfall error in kilometres updating live as the lead time shortens.
Sub-basin impact scoring
63 HydroBASINS units classified Low → Extreme, so the map answers the operational question: which wadi catchments need action, not just where the storm goes.
Station rainfall convergence
Forecast rainfall per wilayat against the recorded gauge totals — early cycles under-estimate, then converge. The accuracy column makes the bias visible instead of hiding it.
Two model families compared
An ensemble view in the style of GenCast / WeatherNext, and a deterministic view in the style of Microsoft Aurora — the same event, two forecasting philosophies.
Live satellite layers
NASA GIBS IMERG precipitation and VIIRS true-colour imagery, wadi network, 3D terrain — toggled on demand over the decision map.
Why arid catchments need their own treatment
Global AI flood models are trained overwhelmingly on perennial, gauged rivers. Arid wadis behave differently: they are dry until they are not, they respond in hours rather than days, and they are barely gauged at all. That gap — documented in the literature and visible in the accuracy columns of this very demo — is the research question behind this work: how far can a global model be carried into an arid basin, and where does local calibration have to take over?
This system is at demo stage and looking for a pilot catchment and a partner agency — see open proposals or get in touch.