Flood Intelligence

Cities need to know where water will collect — and what to do about it — before the storm, not after. The natural first model is a regression: rainfall in, flooded area out, fitted across the world's worst-affected regions and applied to Gulf cities. That requires both numbers, sourced, for the same flood.

• Scope: twelve flood regions — ten global hotspots and two in the UAE — over 2000–2025.
• Users the system is ultimately for: city operators who decide where pumps, tankers and road closures go.
• Standard: every figure must be auditable by a third party. A risk model built on numbers nobody can trace is not something an authority can act on.

48 documented flood events. For each, the dataset needs peak 24-hour rainfall and inundated extent, each with a resolvable source, an evidence tier and a measurement method. Dam- and levee-failure events are excluded (rainfall doesn't drive them), as are extents reported for whole administrative areas.

| Field | Events with a sourced value | |---|---| | Peak 24 h rainfall | 12 of 48 (25%) | | Inundated extent | 14 of 48 (29%) | | Both | 2 of 48 (4.2%) |

The shipped component is the provenance-enforced dataset. The diagram is the target architecture for an urban flood twin that this dataset is the foundation for — the risk-model stage is explicitly blocked until paired data exist.

Everything in the public repository: research design, the validator, the dataset, the coverage audit, the statistics and the write-up. The validator was written — and tested against deliberately broken inputs — before any data existed:

def _check_provenance(df: pd.DataFrame, figures, key: str, table: str) -> list[Problem]:
  """A figure is admissible only with a resolvable URL, a tier, and a method
  where the schema demands one. This is the rule the whole project rests on."""
  problems: list[Problem] = []
  for _, row in df.iterrows():
      rid = str(row.get(key, "?"))
      for fig in figures:
          if fig.value not in df.columns or not _present(row.get(fig.value)):
              continue
          url = row.get(fig.source_url)
          if not _present(url):
              problems.append(
                  Problem(table, rid, fig.value, "no_source", "figure present, source URL empty")
              )
          elif not URL_RE.match(str(url).strip()):
              problems.append(
                  Problem(table, rid, fig.source_url, "bad_source", f"not a URL: {url!r}")
              )

The model that was planned is a leave-one-site-out regression of inundated extent on rainfall. It is implemented, and it refuses to fit when it has too few paired observations — raising NotEstimable instead of returning a confident line through two points.

With the regression ruled out, the analysis turned to why the data are missing: a coverage audit by site and by flood type, and a rank correlation between rainfall coverage and extent coverage across sites.

• 57 tests, including the validator against broken fixtures.
• Rainfall coverage and extent coverage are anti-correlated across sites: Spearman ρ = −0.76, p = 0.0042, n = 12.
• The mechanism: met agencies gauge rain where cities are; satellite mapping activates where floods are wide, flat and rural. Urban sites (Mumbai, Jakarta) have rainfall and no extent; riverine sites (Lokoja, the Yangtze) have the reverse. The pattern holds within a site too.
• Only 3 of the 12 sourced rainfall figures are top-tier evidence.

For an urban twin in a flat Gulf city, two things decide feasibility before any model does:

• Elevation data. Across a test area, global DEMs report relief several times larger than the true relief of the city — the error is as large as the hydraulic signal. A survey-grade DEM is a prerequisite, not an upgrade.
• The handover format. Building to the city authority's own stormwater design standard makes handover a format match rather than a migration.
• The decision is drawdown. Operators care when water clears and what speeds it up — not just peak depth — so the twin must be time-resolved.

• A public, provenance-validated dataset of 48 events across 12 regions.
• A finding that changes what to build: the bottleneck in flood ML is paired measurement, not model choice.
• The only published inundated area for any UAE event is 23.8 km² (Dubai, April 2024). Nothing is published for any Fujairah event.

• Let the data veto the plan. A regression on two points would have produced a chart and a false conclusion.
• Write the validator before the dataset. It turned "I think this is sourced" into a test that fails.
• Next step: reproduce the Dubai 23.8 km² figure independently from Sentinel-1 radar — the first paired observation the Gulf has.