Address-level flood risk scoring: zones, return periods and depth

Fluvial and pluvial flooding are two different perils

Fluvial (river) flooding is a watercourse overtopping its banks. It is the best-mapped peril in Europe because every member state must model it, and it has a clear physics: discharge for a given return period, routed through a channel and floodplain by a hydraulic model, gives an extent, a depth and a velocity. Pluvial (surface-water) flooding is rain falling faster than the ground and the drains can take it, pooling in the low points of a street with no river in sight. It is a large share of household water claims and is invisible on river maps, which only cover modeled watercourses. Flash flooding on small, steep streams sits between the two and is often missing from both.

The distinction matters for scoring because the two perils have different data, different resolution and different remedies. A river flood at an address is read from an official polygon or a modeled raster; surface-water susceptibility is derived from the terrain itself, cell by cell. A levee removes river risk up to its design standard and does nothing for surface water; a raised threshold helps against both. A vendor score that does not say which peril it describes cannot be checked.

The Floods Directive maps and their return periods

Directive 2007/60/EC obliged member states to complete a preliminary flood risk assessment by 22 December 2011, flood hazard and flood risk maps by 22 December 2013 and flood risk management plans by 22 December 2015, and to review all three every six years. The hazard maps are the authoritative state answer to "does this point flood, and how badly". Article 6 fixes what they must show.

Directive 2007/60/EC, Articles 4, 6, 7 and 14. Return-period grids vary by country; the data-source guides linked at the end list them per state.

The JRC European maps

Where official maps stop, the pan-European layer starts. The Joint Research Centre's river flood hazard maps for Europe and the Mediterranean Basin (Dottori et al., 2022) cover every river with an upstream area above 500 km2 at 100 m resolution for return periods of 10, 20, 50, 100, 200 and 500 years, with 30 and 1,000 years in limited areas for validation. Discharges come from a long-term run of the LISFLOOD hydrological model, the same model that drives the Copernicus European Flood Awareness System (EFAS), validated on more than 700 gauges, and inundation from the two-dimensional LISFLOOD-FP hydraulic model. The maps are licensed CC BY 4.0.

Two properties define their use. First, they are undefended: the authors state that the maps do not account for local flood defenses, because no consistent European data on levees exists. Second, their validation against official maps gives a hit rate of about 0.65 at the 1-in-100 event, meaning two thirds of the reference extent is captured, and a critical success index of about 0.35, rising toward 0.5 for the 500-year event. That is a regional screening tool, not an address-level answer; at 100 m one cell is a hectare, and a street can be wet on one side and dry on the other inside it. Its best use at an address is as the undefended counterpart to the official, defended map, and as the only layer on reaches the state has not mapped.

Zone membership versus depth, and why district scores fail

Being inside the Q100 polygon and having 0.75 m of water at Q100 are different facts, and only the second prices a loss. Official maps often carry depth as an attribute, but many publish only extents. Depth can then be estimated from the extent and a digital terrain model with the Floodwater Depth Estimation Tool (FwDET, Cohen et al., 2018): the polygon boundary is treated as the zero-depth shoreline, terrain elevation is sampled along it, a water-surface elevation is interpolated across the interior, and the terrain is subtracted. With a 1 m LiDAR terrain model the error on riverine floods is typically a few tens of centimeters; with a 30 m global terrain model it is not.

Depth is also where the building enters. Rasters measure water above the terrain; damage starts at the finished floor. A residential ground floor typically sits 0.30 to 0.45 m above grade, so one engine subtracts 0.25 m before reading a damage curve, and a flat on an upper storey drops out of the river-flood loss entirely. This is why a score based on zone membership alone over-rates half the buildings in a floodplain and cannot tell a bungalow from a third-floor flat. The expected annual loss guide walks the integration from depth to money.

A flood score for a municipality or postcode goes further in the wrong direction: it averages dry and wet ground into one number that is wrong for almost every building in it. Floodplains are hundreds of meters wide at most; a village on a river has houses in the Q10 extent, houses on the terrace above it and houses on the hillside, and the district score describes none of them. The same holds inside a city block for surface water. In two 5 m hydrodynamic simulations over 1,494 ground-floor addresses, only 3 to 5 percent of addresses received any water at the finished floor at all, and the expected loss rate rose only weakly with the neighborhood susceptibility score; the susceptibility located the wet neighborhoods, but whether water entered a particular building depended on 5 m micro-topography.

Resolution is therefore the first question about any score. An official extent drawn at a 1:10,000 working scale, a depth raster at 5 to 30 m and a 1 m terrain model can tell two adjacent houses apart; a 100 m European raster or a district aggregate cannot. The second question is the building: its floor height, its storey, whether it has a cellar. A score that does not read those cannot separate a ground-floor shop from the offices above it, yet those two policies carry completely different expected losses.

Pluvial susceptibility from the terrain

Surface-water risk has no official polygon at most addresses, so it is modeled from terrain. The physically grounded shortcut is height above nearest drainage (HAND, Nobre et al., 2011): the vertical distance from a cell to the nearest point of the drainage network, which normalizes topography to the local water table and separates low, flat, poorly drained ground from everything else. One national susceptibility composite combines HAND with four other terrain descriptors per 20 to 30 m cell and then amplifies for sealed surfaces.

Weights from the Bytero Shield scoring method. Where a city has a 5 m two-dimensional hydrodynamic simulation (SFINCS), depth and velocity per return period replace the susceptibility score.

How a score and a band are formed

Two representations coexist. A 0 to 100 hazard score ranks addresses on the most frequent return period that still floods them; a zone compresses the same evidence into a few classes that underwriting rules can act on. One published rule set derives the score as 100 − (ln(T/5) / ln(1000/5)) × 80 from the minimum flooding return period T, and the zone from four layers (river defended, river undefended, surface water and a claims-informed maximum-credible layer), taking the highest zone any layer supports and reporting which layer controlled and which layers were missing.

Score curve, bands and zone thresholds from the Bytero Shield scoring method. Insurer zoning conventions in the region likewise use return period for membership and depth only as a modifier.

Defenses: defended, undefended and the protection standard

A levee does not remove the hazard; it moves it to rarer return periods and adds a breach scenario. Official Floods Directive maps are built with the defenses in place, so they are the defended picture; the JRC maps model no defenses and are the undefended picture. Where the two disagree, dry on the official map but wet in the undefended model, the likeliest explanation is a defense, and one screening rule labels the address "likely behind defenses" with low confidence, falling back to a terrain proxy (height above nearest drainage of 2.0 m or less while officially dry) where no undefended raster is loaded. This is screening, not a hydraulic defended-undefended pair, and should be presented as such.

The design standard itself comes from FLOPROS (Scussolini et al., 2016), a global database of flood protection standards expressed as return periods, built from three layers (design data on existing defenses, policy requirements and a modeled estimate) merged into a best estimate per sub-national administrative unit. A label such as "protected to about 1-in-100" tells an underwriter that events up to Q100 are held and that the residual risk is the tail plus breach, which is exactly the shape the worked example in the expected-loss guide shows: removing Q10 to Q100 from a floodplain house cut its raw expected loss by 77 percent and left the Q200 and Q500 contributions intact.

What an insurer should ask a flood-data vendor

Ten questions that separate an address-level product from a re-labeled district map.

• Which peril? River, surface water, flash or coastal, scored separately, and which one controls the headline.
• Which source and which vintage? Official Floods Directive cycle (2013, 2019 or 2025), a JRC 100 m layer, or the vendor's own hydraulic model, per return period.
• What resolution reaches the address? The extent scale, the depth grid (5, 20, 30 or 100 m) and the terrain model behind the depth (1 m LiDAR or a 30 m global model).
• Depth or membership? Whether the output carries water depth per return period at the coordinate, or only which polygon contains it.
• Defended or undefended? Which scenario each layer represents, and the protection standard used, with its provenance.
• Where does the building enter? Floor threshold, storey, cellar, occupancy; whether an upper-floor flat is scored differently from the ground floor below it.
• What happens on unmapped reaches? Whether "no data" is returned as such or silently scored as zero.
• How is the score built? The formula from return period and depth to score, band and zone, and the thresholds, in writing.
• How was it validated? Against claims, high-water marks or official maps, with a hit rate and a false-alarm rate, not a testimonial.
• What is versioned? A model identifier on every response, so a score can be reproduced after the next map cycle changes the answer.

In practice

An address-level score is a chain: official extents, an undefended counterpart, a depth per return period on a fine terrain model, a surface-water layer, the building's own floor and storey, and a written rule that turns all of it into a score, a zone and a confidence label. Bytero implements that chain for a single address and returns the return periods, depths, controlling layer and defense screening alongside the score, in Bytero Shield; availability by market is on the coverage page. The public data behind each state's maps is inventoried in the data-source guides linked below.

Questions