← Back to Buffalo River Dashboard
A small neural network watches the last 48 hours of 1 km radar rainfall over every square of the watershed above these gauges — plus the gauges' own recent behavior — and projects the level forward hour by hour. It was trained on 12 years of storms and validated against every year it never saw.
Based only on rain that has already fallen. If more rain falls, these numbers become the floor, not the forecast — the model re-issues every hour as new radar arrives. Experimental: trust the main dashboard and your own judgment first.
| Time | Predicted (cfs) | Likely range |
|---|---|---|
| +1 hr (6:00 AM) | 2 | 2 – 2 |
| +2 hr (7:00 AM) | 2 | 2 – 2 |
| +3 hr (8:00 AM) | 2 | 2 – 2 |
| +4 hr (9:00 AM) | 2 | 2 – 2 |
| +5 hr (10:00 AM) | 2 | 2 – 2 |
| +6 hr (11:00 AM) | 2 | 2 – 2 |
| Time | Predicted (cfs) | Likely range |
|---|---|---|
| +1 hr (6:00 AM) | 1 | 0 – 1 |
| +2 hr (7:00 AM) | 1 | 0 – 1 |
| +3 hr (8:00 AM) | 1 | 0 – 1 |
| +4 hr (9:00 AM) | 1 | 0 – 1 |
Model notes this hour: Ponca gauge last posted 12 min before this run; Boxley gauge last posted 42 min before this run.
Predicted value is the model's median, nudged toward “no change” in dry spells (the net wobbles when nothing is happening) — except while the crossing model is at its watch point, when the table shows the model's full projected move so the two never contradict each other; below 10 cfs with no rain the table simply holds the current level, and the likely range is never narrower than the gauge's own reporting step. The likely range spans the model's 25th–75th percentile, sized so it contains the truth about half the time. When the nudge changes a number by more than half, the unshrunk “model raw” value is shown beside it. Yellow cells are above 800 cfs (mid-optimal on the main dashboard), red cells are at or above the dashboard's Flood level for that gauge. Crossing chances are calibrated against 12 years of held-out storms and only shown once the model's signal passes a validated watch level (for 800 cfs: watch from about 20%, alert from about 67%; for flood: watch from about 48%, alert from about 56%); Boxley's crossing heads were never event-validated, so Boxley shows none. Boxley gets a shorter window because its smaller upstream basin gives the radar less lead time. Model: pixel-v5, running on-site — no cloud services. Want the full story — the architecture, the training, the failed experiments? Understand the math →