Flash Flood: how HazardNet classifies and scores it

Short-duration, high-intensity flooding — the burst case of the rainfall proxy.

How the model arrives at this class

The classifier decides between Flood and Flash Flood from the driver tensor alone. On the committed snapshot the two classes are almost never both plausible for the same unit, and the physics track cannot arbitrate between them.

Same softmax caveat as Flood: uncalibrated, and published with that label.

Nothing above the WATCH level can be published automatically while no calibration map exists, and nothing at all can be published while the pipeline stamps no model version — the alert engine records that as `publication_blocked` rather than issuing the alert.

The independent physics cross-check

Identical to Flood in this implementation. The intensity term is what makes a burst score higher than a drizzle of the same total, but nothing in the proxy separates a flash flood from a slow one.

Drivers: precipitation accumulated over the horizon (P_total) · wettest 24 hours in that horizon (P_peak).

clip(x) clamps a term to [0, 1] — max(0, min(1, x)) — which is how every formula in the pipeline bounds its terms. Each term is clipped before it is weighted, so no single driver can run away with a score.

  • Expression: 0.5*clip(P_total/300) + 0.5*clip(P_peak/100)
  • That expression is executed against scripts/physics_severity.py by scripts/tests/test_content_engine.py — this page cannot silently describe a formula the pipeline no longer runs.

What the current run says about Flash Flood

In the run this deployment ships (prediction date 2026-09-16), 71 district-horizon units are classified as Flash Flood: 49 at 15 days · 22 at 7 days.

Highest model severity: Sunamganj (1.00, 15 days), Sylhet (1.00, 7 days), Khulna (1.00, 15 days), Sunamganj (1.00, 7 days), Kishoreganj (1.00, 15 days).

These are classifications from the current run, not a forecast of impact. The model version behind them is not stamped yet, and PRODUCT_SPEC §1.6 keeps publication closed above the WATCH level until one exists.

Season and geography

Typical season: March–May (pre-monsoon convection) and June–September. This describes when the hazard is climatologically plausible, not when this run flags it.

What this class cannot tell you

  • No terrain or drainage: catchments, slope and urban surfaces are absent, which is exactly what decides whether a burst becomes a flash flood.
  • Haor and hill-district flash floods are driven by upstream rainfall outside the district boundary; the drivers are district-aggregated.
  • A 24-hour peak inside a 15-day horizon is a coarse instrument for an event that lasts hours.
  • Confidence bins (Certain ≥ 0.85, Probable 0.70–0.85, Uncertain < 0.70) describe the model's own certainty in the class it chose. They are not accuracy, and no accuracy figure is published because no held-out evaluation has been run on real labels.

Inputs behind it

Ground-truth reports (district, hazard, horizon, date, what was observed) are the only route by which this project can publish skill metrics. The contact page reaches the maintainers.

  • Open-Meteo forecast (rainfall intensity)
  • ERA5-Land daily aggregates
  • FFWC bulletins (hydrological context, ingested as text)

Questions and answers

Is a high Flash Flood severity a prediction of damage?

No. The severity index is the model's continuous score for how strongly the drivers resemble this class, on a 0.00–1.00 scale. It is not a probability, not a percentage, and it does not model exposure. The physics cross-check is a second, independent estimate; where the two diverge, both numbers are shown rather than averaged.

Why is the confidence not a probability of Flash Flood?

The published confidence is the model's own softmax over its eight classes, labelled uncalibrated_model_softmax in the data. It says how certain the classifier is about the class it picked, not how often that class actually materialises. No calibration map has been fitted yet, which is also why nothing above the WATCH level can be published automatically.

Which districts does this page cover?

The district pages cover all 64 districts. This page summarises the districts the current run covers — a run can be partial, and the coverage stamp on the snapshot says exactly how partial, which the status page reports.

Content reviewed 2026-09-16. HazardNet is decision support, not an official warning service — see the methodology for scope and limitations.

Loading the interactive HazardNet application…