🌐 CRMA scenario website: https://crma-frontend-yiyrp6yumq-uc.a.run.app/scenario
🌐 Presentation Slide: https://drive.google.com/file/d/1rR5uhCan3snwHgv6A-5ch6IXfw8ey6wj/view?usp=drive_link
Participants relive a real East African drought or flood as it unfolded. Working only with the evidence available at the time, they grade each admin-1 region with a risk colour — the call a Disaster Operations Centre (DOC) has to make.
End outcome: every admin-1 marked with a CRMA colour grade, justified by the evidence weighed.
🟢 Monitor · 🟡 Evaluate · 🟠 Assess · 🔴 Actionable Risk
A date cursor is stepped through the event window:
- Flood — daily: last ~7 days observed rain + next ~7 days forecast (ECMWF ensemble), tested against a return-period threshold — the ~15-day lead → escalation → onset window.
- Drought — monthly (SPI-3): last ~6 months observed + the next season (MAM / JJA / OND / DJF) forecast at a ~4–6-month lead from the init month.
The monitoring calendar/timeline for the event's window is read (daily for flood, monthly for drought) to build the situational picture.
Monitoring calendar — monthly for drought (shown), daily for flood; each cell shaded by the share of admin-1s at Actionable Risk.
The evidence is classified and weighed — hard (what we measure) · soft (what we estimate) · virtual (what we imagine). A Bayesian Network combines it into a hidden risk grade (Minimal → Extreme, expert-rules judgment) and a CRMA decision.
Per-admin-1 BN: evidence nodes (Current SPI-3, Forecast Deficit, Spatial, Trend) → hidden risk-level grade → CRMA decision (here Monitor / green, P(High∪Extreme) = 0%, γ = 0.20).
Participants commit a CRMA colour grade for each admin-1. The recorded loss & damage is then revealed and compared with their call and with the model.
The end outcome — each admin-1 graded green → red on the available evidence (Risk Monitoring, Dec 2023).
- Scenario app & CRMA frontend — the
arco-ibfweb app used to run this exercise: https://github.com/icpac-igad/arco-ibf - Drought Bayesian-network analysis — SPI-3 monthly BN (evidence → risk grade → CRMA decision): https://github.com/nishadhka/bn-ibf/tree/jua-bnet/drought_ibf
- Flood Bayesian-network analysis — daily BN on the same topology: https://github.com/nishadhka/bn-ibf/tree/jua-bnet/flood_ibf
- EPS data-streaming method — GRIB-index + Kerchunk for ensemble forecasts: https://github.com/icpac-igad/grib-index-kerchunk
- Storylines & hazard modelling — event storylines and hazard model DevOps: https://github.com/icpac-igad/DevOps-hazard-modeling
Analysis-Ready, Cloud-Optimized (ARCO) datasets and streaming formats:
- Observations (ERA5 SPI, IMERG …): https://source.coop/e4drr-project/observations
- Forecasts (SEAS5 SPI-3, ECMWF …): https://source.coop/e4drr-project/forecasts
- ECMWF EPS — GRIB-index Parquet: https://huggingface.co/datasets/E4DRR/gik-ecmwf-par
- GEFS EPS — GRIB-index Parquet: https://huggingface.co/datasets/E4DRR/gik-gefs-par
- Flood hazard model output (RIM2D): https://huggingface.co/datasets/E4DRR/rim2d-simulations
- Drought hazard model output (wflow.jl): https://huggingface.co/datasets/E4DRR/wflow.jl-simulations
Per-event ARCO → Bayesian-Network provenance notebooks: how each event's evidence is streamed from the ARCO stores and turned into the BN risk grade. Open any one directly in Google Colab — no local setup:
- Kenya — Tana / ASAL drought 2020 —
- Burundi drought 2021 —
- Eritrea Highlands drought 2021 —
- Djibouti drought 2022 —
- Kenya — Nairobi flood 2026 —
This work is part of the E4DRR project — hazard modelling, impact estimation, and climate storylines building an event catalogue of drought and flood disasters in Eastern Africa: https://icpac-igad.github.io/e4drr/
Funded by the United Nations Complex Risk Analytics Fund (CRAF'd).
Data & services — built on open data from AWS Open Data, ECMWF, NOAA, the EC Joint Research Centre (JRC) — Global Drought Observatory, and ICPAC's East Africa Hazard Watch.
Open-source software — powered by free and open-source tools, including Icechunk, Xarray, and Kerchunk.