Technical project
Atlas: Flash-Flood Decision Support for the High Atlas
Forecast-driven flash-flood decision support for the Ounila valley (Morocco), built on a physically based rainfall–runoff–2D hydraulics–impact model chain and a static React + MapLibre web app.
- Python
- Hydrology (HEC-HMS)
- 2D Hydraulics
- Copernicus DEM
- ERA5 / IMERG
- React
- MapLibre
- GitHub Actions
Overview
Atlas answers one question for the people who have to act when a storm comes: rain is coming — what will happen, where, when, and what should we do first? A live 7-day weather forecast is turned into an impact-based alert, village-by-village effects, road isolation, an early-action plan with deadlines, and evacuation routes, all served from a static web app on top of a model chain grounded in observed data.
Context
The Ounila valley drains 731 km² of the High Atlas down to the UNESCO ksar of Aït Ben Haddou. Its earthen villages, fords and wadi-bed tracks are exposed to sudden convective floods such as the November 2014 southern Morocco floods. Atlas started as a 48-hour prototype at the Cassini Hackathon and was rebuilt from scratch as Atlas v2 with a validated, physically based model chain.
Objective
To turn a numerical weather forecast into actionable, place-specific flood information for local authorities: the probability of each flood level, when water reaches each village, which roads close and which villages are cut off, and what to do first.
Method
-
01
Terrain: catchment and 21 sub-basins delineated from Copernicus GLO-30, matching the DRPE Agouillal gauge area (740 km²); flow paths breached rather than filled so radar-DSM artefacts in gorges do not create artificial reservoirs
-
02
Rainfall: GEV (L-moments) on 76 years of hourly liquid ERA5-Land rain, storm shape from GPM IMERG, areal bias correction ×1.43 against the Ouarzazate gauge, with bootstrap confidence intervals
-
03
Hydrology: HEC-HMS-equivalent engine in Python (SCS-CN loss, Clark unit hydrograph, Muskingum-Cunge routing), exported as a HEC-HMS project for independent cross-checks
-
04
Hydraulics: own 2D local-inertia shallow-water solver (LISFLOOD-FP formulation) on 30 m cells over 101 km² of valley corridors, run for a 13-storm library from 20 to 160 mm / 24 h
-
05
Impacts: DEFRA FD2320 hazard on 16,653 Google building footprints, ARR 2019 vehicle limits on OSM roads and fords, and road-graph isolation to the RN9 recomputed every 30 minutes
-
06
Forecast: Open-Meteo ECMWF, GFS, ICON and a 51-member ensemble, adjusted for antecedent soil moisture and matched to the precomputed library, with identical logic in Python and TypeScript (parity test)
-
07
Delivery: static React + MapLibre app on GitHub Pages; a GitHub Actions workflow refreshes the forecast every 6 hours
Data
- Copernicus GLO-30 DEM
- ERA5-Land hourly rainfall (1950–present), GPM IMERG V07, GHCN Ouarzazate gauge
- ESA WorldCover and OpenLandMap soils
- Google Open Buildings (via VIDA), OpenStreetMap roads, fords and facilities
- GHS-POP 2020 and WorldPop 2020 population
- Open-Meteo deterministic and ensemble forecasts
Visuals
Analysis
- Modelled storm runoff at the Agouillal gauge is 28 mm/yr, inside the observed 19–36 mm/yr, and design peaks at all 21 control points fall within the High Atlas regional envelope (Ourika, Rheraya)
- The 2D solver reaches the analytical Manning normal depth within 5% and every library run closes its mass balance to 10⁻⁴ or better
- 2D peaks are 8–25% below the HMS peaks: the wide gravel beds store and attenuate floods, most strongly for small events
- Two independent population products agree within 5% (15,571 residents in the catchment)
Results
- The dominant impact is isolation, not flooded houses: from a 5–10-year flood upward, 19 villages lose 4WD access to the RN9 because tracks and the RP1506 run in or across the wadi
- A 100-year storm reaches 230 buildings (64 in extreme hazard) and 249 residents, with collapse possible for 46 earthen houses
- Flood-proofing the 10 most critical road links cuts isolation in a 10-year flood from 402 to 144 village-hours, with 10 of 19 villages keeping access; check dams barely change flood peaks
- The T100 storm generates 128 early actions (36 critical) from an editable IFRC-style playbook, the first due 5.5 h after storm start
- Past storms such as the November 2014 floods can be replayed through the same logic for exercises and training
Reflection
- Atlas is a research prototype for planning and training, not a certified operational warning service
- No flood-peak record is public for this river, so peaks are bounded by regional evidence and the outlet water balance rather than calibrated event by event
- The 30 m terrain smooths gorges narrower than one cell, and forecasts inherit NWP errors in convective rain
- Precomputing the 2D flood library is what lets a static site deliver flood maps for any forecast in milliseconds; a HEC-RAS 2D package is prepared for an independent cross-check