permanent water (sea, port, Albufera)
new inundation
affected buildings
cloud-obscured (~¾ of corridor)
Headline results
≥ 15.06 km²
new inundation (lower bound — 73% cloud)
≥ 254
buildings flooded (of 31,463 mapped)
Paiporta #1
hardest-hit municipality — matches the real epicentre
±3%
permanent water: 90.2 km² (satellite) vs 87.8 km² (JRC)
95% < 2 m
flood pixels at lowest elevation — 0 above 100 m (DEM check)
55 mm / 72 h
MERRA-2 regional rain · stations recorded 400+ mm (DANA)
What the agent caught — errors a naïve run would have shipped
🛰️ Imagery missed the city
The provided Sentinel-2 scenes covered only the sea east of Valencia — the flood zone was nodata. The agent detected this and re-fetched AOI-cropped post-flood imagery.
🗺️ Baseline was the Sahara
The on-disk JRC surface-water tile covered 20–30°N (North Africa), not Valencia at 39.5°N. It caught the mismatch and downloaded the correct tile.
🌾 Farmland read as water
The first index (NIR−SWIR) flagged irrigated cropland as “water” (3,307 km² — impossible). It diagnosed the physics and switched to MNDWI.
Full step-by-step reasoning → agent_log.md
Affected structures by municipality
| Municipality | Buildings flooded (≥) | Risk |
|---|---|---|
| Paiporta | 89 | HIGH |
| Sedaví | 44 | HIGH |
| Massanassa | 35 | MEDIUM |
| Benetússer | 24 | MEDIUM |
| Catarroja | 22 | MEDIUM |
| Picanya | 21 | MEDIUM |
| Albufera / Sur | 13 | LOW |
Priority actions
- 0–6 h: Deploy rescue + medical to Paiporta and Sedaví — highest confirmed impact, dense, lowest ground.
- 0–6 h: Recon the cloud-obscured zones (~¾ of corridor, west/SW) — mapped extent is a floor; assume larger.
- 0–24 h: Open access to Massanassa, Catarroja, Benetússer; check Poyo-crossing road/rail for cuts.
- 24–72 h: Damage survey for recovery funding; acquire a clear-sky or SAR scene to finalise true extent.
What triggered it — MERRA-2 rainfall
NASA MERRA-2 precipitation at the Valencia grid cell (39.50°N, −0.625°E). Regional
reanalysis totals ~55 mm over 72 h with a 3.9 mm/hr peak — far below the 400+ mm recorded by local
stations, because the ~50 km grid averages out the hyper-local DANA cells. The flood was driven by
water routed down the Barranco del Poyo from extreme upstream rainfall, not rain on the plain itself.
How it works
- Flood extent — Sentinel-2 MNDWI change detection (pre vs post-flood), cloud-masked.
- Permanent water — subtracted and cross-validated against the JRC 38-year surface-water record.
- Impact — flood mask intersected with 31,463 real OpenStreetMap building footprints, by municipality.
- Cause — NASA MERRA-2 precipitation context, with the DANA grid-averaging caveat.
- Plausibility — Copernicus GLO-30 DEM confirms flooding sits on the lowest ground.
- Verification — a Claude sub-agent grades the brief against a rubric until it passes (result: PASS, 11/11).
Read the full output
⚠️ Quantities are lower bounds: the 31 Oct post-flood scene was 73% cloud, so flooding under cloud is unobservable. The agent reports floors, not guesses — and says so.