Tool · concept demo
From a cyclone forecast to local impact
Global centres forecast where a cyclone is going and how much it may rain. National hazard maps show where floods can reach in a design event. What is missing is the layer that connects the two for a specific storm: which sub-basin, which wadi crossing, when, with what probability — and a record of how often such probabilities were right. This demo shows the shape of that layer on a real event.
Cyclone Shaheen, Oman — October 2021
A decision map replaying a real event: the observed track into Al Batinah, the rainfall recorded at each wilayat, 63 HydroBASINS sub-basins ranked by a simple impact index (event rainfall and distance to the track), satellite rainfall and imagery for the event dates, and a timeline from 48 hours before landfall to 12 hours after.
The demo is part of the Oman proposal package and is shared with the people involved, by arrangement; it opens with a password. Request access
Opens full screen; needs internet for basemap and satellite layers. Historical reconstruction for demonstration — not a live forecast service.
What is real, and what is illustration
| Real | Cyclone track (8 observed fixes), wilayat rainfall totals, HydroBASINS sub-basins and HydroRIVERS wadi network, NASA GIBS satellite rainfall and imagery for 1–4 October 2021. |
|---|---|
| Heuristic | The sub-basin impact index is a simple rule of event rainfall and distance to the track. It is a ranking aid, not a hydrological or hydraulic result. |
| Synthetic | The ensemble scenarios, the single-track view and the station-by-station "forecast" numbers are generated around the observed track to illustrate how forecast updates look. They are not output of any model and say nothing about the skill of GenCast, WeatherNext, Aurora or any other system. |
How the proposed system differs from this demo
The proposal builds a two-tier system. The operational tier uses established methods only: verified and post-processed rainfall forecasts (M1), an event-based hourly rainfall–runoff model with explicit arid processes (M2), a pre-computed hazard library on the ministry's national hazard maps (M3), probabilistic impact per sub-basin and wadi crossing (M4), and a verification loop that scores every issued probability against what happened (M6). Machine learning is used where it is defensible — error correction, parameter regionalisation, fast emulation — and graph neural networks stay in the research tier, tested against a simpler baseline with a stop criterion.
What is not claimed
- Better weather forecasts than the global centres — their forecasts are the input and are verified, not replaced.
- A deterministic flood extent at metre precision — the output is a probability of exceeding a threshold, with an uncertainty band.
- Graph neural networks as an operational method — they are a research topic with a stop criterion.
- Issuing public warnings — the product is decision support for the competent authorities.
The next gate
The demo is a concept prototype (TRL 4–5). The first real step is Phase 0 of the proposal — a data audit and a data-access agreement — followed by a hindcast of past cyclones with the forecasts that were available at the time, on one pilot wadi. Detailed working material is shared by arrangement.