Executive Summary

FloodCat — what it is, how it works, susceptibility map

Catalonia flood intelligence

32,108 km² · 7.8M people · 8 validated events · 138 historical episodes

What is FloodCat?

The project: FloodCat is an end-to-end flood-intelligence platform for Catalonia. It fuses four decades of historical flood records with Sentinel-1 SAR and Sentinel-2 optical imagery, a terrain-based susceptibility map, and a Random Forest impact model — turning raw geoscience data into decisions for insurers, civil protection, and farmers.

Three datasets, three roles:

  • 8 validated modern events (2019–2021) — Sentinel-1/2 satellite-confirmed floods. Used to train the houses-affected regressor and to validate the FloodPotential susceptibility map (59% mean agreement between satellite-detected flood extents and our High/Medium zones).
  • 82 AGORA / INUNGAMA episodes (1996–2020) — insurance-grade economic loss records. Used to train the economic-loss regressor (€ per event).
  • 56 DesInventar episodes (1981–2003) — UN-OCHA disaster inventory with houses and people impacted. Used together with the modern events to expand the houses-affected training set.

The model: a scikit-learn Random Forest (200 trees, 20 features). Rainfall intensity, antecedent soil moisture, and FloodPotential exposure are the dominant predictors. Detection is delegated to IBM/NASA's Prithvi-EO-2.0 foundation model (Sentinel-2) and adaptive SAR thresholding (Sentinel-1).

Use the sidebar: Farmers to subscribe a parcel · Insurance for loss curves & impact estimation · Civil Protection for seasonality & climate trend · Technical for satellite detection & model performance.

Validation (8 events)

59%

Episodes (40 yrs)

138

Economic loss

€1.78B

Results computed from the FloodCat algorithm and trained .pkl Random Forest models (feature importances, 64 Catalonia training rows, FloodPotential zones) imported directly from the project repository.
All flood events (146)
Every event from the three datasets, plotted across Catalonia.
Loading map…
Validated modern (8)AGORA economic loss (82)DesInventar (56)
FloodPotential susceptibility map
Terrain-based flood-risk zones (Copernicus GLO-30 DEM, 30 m). Markers show validated events colored by example zone class.
Loading map…
1

Zone 1 — High

Floods first — lowest FloodOrder quartile

2

Zone 2 — Medium

Second quartile of FloodOrder

3

Zone 3 — Low

Third quartile

4

Zone 4 — Very Low

Highest FloodOrder quartile

Susceptibility (where)

A map of where water naturally wants to go. We take a detailed elevation model of the land (3D terrain, ~30 m resolution), simulate how rain would flow downhill, and highlight the urban areas where it would pool first.

Each city block gets a risk grade from 1 (floods first) to 4 (floods last). This is based purely on the shape of the land — it doesn't know about past floods.

Impact model (how bad)

Random Forest · 200 trees · 20 features. Three regressors: AffectedHouses (64 events), AffectedPeople (64), EconomicLoss (126 AGORA).

Rainfall intensity and FloodPotential exposure dominate Gini importance.

Pipeline (5 phases)

① Feature extraction (20 vars)

② RF training (64 + 126 events)

③ Impact prediction

④ Sentinel-2 + Sentinel-1 detection

⑤ Validation against FloodPotential — 58.6%