Compound Flood Modelling (SFINCS + Delft-FIAT)

Ingested from Thomas’s “Compound flood modelling report” (Resilient Hydrotwin, WP4 — “Integration of Hydro Models with Digital Twin for Improved Decision-Making”), 2026-07. This is the methodology and data behind the Rotterdam SFINCS setup in WP4/delftsfincs-rotterdam/ — read it alongside Raster & Scenario Data Layout (how outputs are served) and Schema Reconnaissance (Phase 0) (how the platform’s database represents scenarios/interventions today).

Context¶

Rotterdam is simultaneously exposed to pluvial (rainfall), fluvial (river discharge), and coastal (storm surge) flooding. Traditional modelling assesses these separately; this work package models them as compound events — multiple hazards occurring together — using SFINCS. The pilot domain runs from the Tweede Maasvlakte (North Sea coast) to Rotterdam’s eastern boundary, covering port infrastructure, residential areas, and the North Sea–to–inland-river hydraulic system.

The SFINCS model¶

SFINCS (Super-Fast INundation of CoastS, Deltares) solves the Local Inertial Equations — a simplified form of the shallow water equations — combined with a subgrid scheme (Leijnse et al., 2021), giving 35–1700× speedup over full hydrodynamic models at ~5–14 cm accuracy on hurricane test cases (van Ormondt et al., 2025). Valid for subcritical flow (Froude < 0.5); optimized for inundation and compound flooding, not for supercritical/transcritical regimes.

Model inputs and their provenance¶

This is the authoritative source for the DEM/data provenance question raised in RHT-96 / RHT-98:

Model grid: 25 m calculation grid, 0.5 m subgrid, EPSG:28992 (Amersfoort / RD New), built via HydroMT-SFINCS. This matches model_subgrid_builder/sfincs.inp exactly (dx=dy=25, epsg=28992).

Forcing scenarios¶

Three hazard types, evaluated individually and in compound combinations:

• Pluvial — return periods 1/10, 1/100, 1/250 yr → 63 mm, 99 mm, 117 mm rainfall in 24h (STOWA 2024 statistics). Net effective load subtracts an assumed 25 mm of drainage/storage capacity: e.g. 63 mm event → (63 − 25) / 24 = 1.58 mm/h effective load fed to SFINCS.

• Fluvial — derived from 10 years of Rijkswaterstaat 10-minute discharge records at Lobith; the 1/250-yr vs 1/10-yr discharge differs by ~50% (Parmet et al., 2001), so Brienenoord/Puttershoek discharges are scaled ×1.5 for the extreme scenario.

• Coastal — storm surge boundary at Hoek van Holland, based on Storm Report SR88 (~1/10-yr return period; Bijlsma, 2007). The 1953 flood is ~1/300-yr.

Because 1/250-yr and 1/300-yr return periods showed inconsistencies across hazard components, a unified 1/300-year return period was assumed for precipitation, storm surge, and fluvial discharge together.

The 1/100-yr values (used throughout the delivered pipeline, e.g. 6_IRP_100_years) come from a two-point Gumbel fit anchored on the 1/10-yr and 1/300-yr values, with uncertainty propagated via 10,000-sample Monte Carlo across 5 independent sources (surge uncertainty at both anchor points, discharge uncertainty and scaling, GEV shape parameter). Resulting nominal 1/100-yr estimates: 3.64 m MSL storm surge, 4886 m³/s at Brienenoord, 3789 m³/s at Puttershoek — these are the exact values hard-coded in WP4/delftsfincs-rotterdam/6_superimpose_boundaryconditions.py (TARGET_WATER_LEVEL_PEAK_M = 3.64, etc.), confirming that script’s targets are this Gumbel/Monte Carlo analysis, not arbitrary numbers.

Barrier operation: the Maeslantkering is modelled as closed (fixed levels/discharges) consistent with its ~1/100-yr failure-to-close probability (recent analysis suggests this may be conservative — Deltares, 2023); the Hartelkering is assumed to close under comparable conditions.

Simulation window: 5 days spin-up (moderate discharge/tidal levels) + the main event + 7 days recession = 12 days total — close to, but not identical to, the 13-day window (tstart–tstop) in the delivered sfincs.inp.

The 13 compound scenarios¶

Scenarios 1–7 match the 7 folders already on disk under data/rat_data-500/rotterdam/ by name and order (1_base_scenario … 7_IRP_100_years_storagebasins) — this table is their authoritative definition. Scenarios 8–13 are defined here but not yet generated on disk; relevant to RHT-81 (generating the real runs the golden question set needs).

Output files¶

Per-timestep GeoTIFFs, {variable}_{YYYYMMDD}_{HHMMSS}.tif, Deflate compression, all in EPSG:28992:

This matches SCENARIO_PROCESSING.md’s description of stage 4 (4_create_subgrid_results.py) exactly.

Flood damage assessment (Delft-FIAT)¶

Depth–damage approach via HydroMT-FIAT: per building, max potential damage (replacement value) × a depth–damage curve fraction = damage at a given depth, aggregated over the exposure set.

• Vulnerability curves: JRC Global flood depth–damage functions (Huizinga et al., 2017), European curves, per building/occupancy type.

• Replacement values: JRC country-level construction costs (Netherlands), adjusted via preprocess_jrc_damage_values (construction cost vs. depreciated value, content share, undamageable part, material use).

• Units: EUR, price reference year 2010 — not USD (eur_to_us_dollar = False); a 2010 EUR–USD rate of 1.327 is available in code if needed.

• Exposure: building footprints/occupancy from OpenStreetMap (confirms the OSM sourcing already documented for the knowledge graph — see Knowledge Graph: Buildings & Flooding §“PDOK/BAG enrichment”).

• Full settings.toml config in the source report’s Appendix 3 (hazard file hazard/hmax.nc, elevation_reference = "datum", risk = false, i.e. this pipeline runs single-event damage, not probabilistic risk).

Measures — the authoritative vocabulary¶

This supersedes the “5 real types” understanding used earlier in this project. Django’s InterventionArea.InterventionType currently implements 5 of the following; the other 6 are real, documented measures this modelling framework supports but that don’t have InterventionArea rows yet. Per direction from the requester (2026-07-31): treat this table as the ceiling for RHT-78 — don’t add measure types beyond what’s documented here, and don’t expand InterventionType to the full 11 yet.

Bold = already implemented as an InterventionType choice in Django (water_square, compartmentalization, green_roofs, permeable_pavement, vegetation_parks).

Two corrections this resolves for the eval question set (eval/golden.yaml/eval/adversarial.yaml, which assumed GRN_ROOF, WADI, WATER_SQ, PERM_PAVE, PUMP_CAP, QUAY_RAISE, SEWER_UP, RETENTION):

• “Wadi” is not a separate measure from “water square” — Thomas’s table lists them as one row (“Water square or wadi”), sharing the same mechanism (level_change on the subgrid DEM). The eval set’s WADI and WATER_SQ codes should map to the same real type, not two.

• Pump capacity and sewer capacity are one combined measure (“Pump capacity / sewer capacity increase”), not two separate codes (PUMP_CAP, SEWER_UP) — both act through the same mechanism (net rainfall / precipfile).

• QUAY_RAISE and RETENTION are real (“Dike raising / quay wall heightening”, “Retention basin”) — just not yet implemented as InterventionType choices.

This also confirms the mechanism mapping used in RHT-102 (automating a SFINCS run per intervention): compartmentalization → weir, water square → subgrid DEM edit, green roofs → storage volume, permeable pavement/vegetation → Manning’s roughness — all independently verified here rather than inferred.

References¶

Full citation list in the source report’s References section — key ones: Leijnse et al. (2021, SFINCS method), van Ormondt et al. (2025, subgrid accuracy), Huizinga et al. (2017, JRC damage functions), STOWA (2024, rainfall statistics), Parmet et al. (2001, Lobith discharge scaling), Bijlsma (2007, SR88 storm report).