Storm Impact Harmonisation
Reconciling cyclone population exposure across CHD, GDACS, ADAM and PDC
There is no single authoritative figure for how many people a cyclone exposed. CHD, GDACS and WFP ADAM each produce national population exposure estimates, and they disagree — not marginally, but on both magnitude and on who was exposed at all.
ADAM is close to a duplicate of GDACS (log-correlation 0.94), so the three sources are effectively two. Where CHD and GDACS both find people, CHD reads about half of GDACS (median ratio ≈ 0.5), and the two agree on which country–storm units are exposed at roughly chance level (κ ≈ 0). The comparable record is also thin: GDACS yields a usable per-country exposure for only ~60 storms, owing to gaps in its 2016–2022 data.
Taking the per-cell maximum of the three — the “one number from three” approach — roughly triples the CHD figure (~3.2×). That number is substantially an ADAM/GDACS one: weighted by people, ADAM supplies 68% of the country-level MAX and 54% subnationally, while CHD contributes 4–10%.
Source comparison. ADAM’s redundancy with GDACS reduces a three-source problem to two. Within the comparable subset, disagreement is both in magnitude (median CHD/GDACS ≈ 0.5 on both-positive units) and in classification (κ ≈ 0 on whether a unit is exposed). Early exploration of the GDACS and ADAM historical exports revealed substantial temporal gaps (see Section 11.3) — GDACS rows exist for 2016–2021 but carry null exposure values, and ADAM begins only in 2023.
MAX aggregation. The per-cell maximum is usually close to its runner-up (median ≈ 1.08×), so it behaves less like an outlier-picker than expected; about 1 cell in 8 disagrees by ≥2×. Subnationally the MAX surface stitches all three methodologies together — across the admin units of a single storm, CHD, GDACS and ADAM each win somewhere.
Monitoring pipeline. GDACS exposure can be retrieved in real time rather than from historical exports, with per-issuance (“episode”) resources distinguishing current storm state from cumulative impact and from forecast evolution. GDACS events match to the NHC A-deck via atcf_id by comparing forecast valid_time coordinates — the TCM advisory updates only the observational position, leaving forecast steps identical. NHC wind-speed probability polygons are basin-wide rather than per-storm, tiled as non-overlapping concentric rings; spatial attribution resolves ~76% of ambiguous timestamps. The PDC Hazards API was evaluated as a candidate third source and is captured continuously but not integrated (ADR 0005).
0.1 Background
This book documents work by OCHA’s Centre for Humanitarian Data (CHD) Data Science team on two related streams. First, a pipeline estimating national population exposure to tropical cyclone wind speeds, using IBTrACS storm tracks combined with wind field models. Second, an effort to reconcile exposure estimates from multiple providers — GDACS, WFP ADAM, PDC and the in-house CHD model — into something a humanitarian decision-maker can act on.
The motivating problem is practical. Operational users want one number for “people exposed”, and cross-agency monitoring requires that different organisations reporting on the same storm not contradict each other. But the providers use different wind footprint geometries, different population rasters and different reporting thresholds, so their estimates are not interchangeable. Understanding the size and structure of that disagreement is a precondition for standardising on any single figure.
0.2 Research Questions
Source comparison. How far do CHD, GDACS and ADAM disagree on national cyclone population exposure — in magnitude, and in which units they report as exposed at all? Is the historical record complete enough to compare them?
Aggregation. If a single number is required, what does taking the maximum across sources actually produce, and which source ends up supplying it?
Monitoring. Can these sources be retrieved in real time, matched to a common storm identity, and combined into a monitoring pipeline?
0.3 Data Sources
| Dataset | Blob Path | Description |
|---|---|---|
| OCHA exposure | ds-storm-impact-harmonisation/processed/adm0_ibtracs_exp_all.parquet |
OCHA national pop. exposure (34/50/64 kt), 2001–2026 |
| ADAM exposure | ds-cyclone-exposure/adam_historical_national_exposure.csv |
WFP ADAM national pop. exposure (60/90/120 km/h) |
| GDACS exposure | ds-cyclone-exposure/gdacs_historical_national_exposure.csv |
GDACS national pop. exposure (34 kt / 64 kt) |
| CERF storms | ds-storm-impact-harmonisation/processed/cerf-storms-with-sids-2024-02-27.csv |
CERF-funded storm events with IBTrACS storm IDs (appendices only) |
Live sources (GDACS API, NHC GIS archive, PDC Hazards API) and the exposure database are documented in docs/.
0.4 Book Structure
- Monitoring Pipeline Exploration — real-time GDACS retrieval, GDACS per-issuance episodes, NHC wind-speed probability polygons, GDACS-to-NHC A-deck matching, and the PDC evaluation (initial, then revisited on the 2026 season).
- Exposure Method Comparison — CHD vs GDACS vs ADAM across magnitude and classification, and the MAX methodology for producing one number from three.
- Appendices — the data merge pipeline (storm ID matching, exposure joining), the ADAM/GDACS/CERF overlap analysis and its completeness blockers, and a superseded two-source comparison kept for reference.
0.5 Next Steps
- Multi-hazard factors: incorporate rainfall and storm surge exposure alongside wind, since many cyclone impacts come from flooding rather than wind damage alone.
- Resolve the historical gaps: the GDACS 2016–2021 null-exposure rows and ADAM’s 2023 start date are the binding constraint on any long-record comparison.
- Operationalise the harmonised estimate: the MAX surface exists subnationally, but choosing and publishing a single agreed figure is a cross-agency decision, not only a methodological one.
This book previously included a CERF Predictions section replicating and extending the CERF rapid-response regression model (3RM) — including whether cyclone exposure improves it, and whether INFORM indices can replace CIRV. That work now lives in OCHA-DAP/ds-cerf-allocation-patterns, together with its data loaders, models and refresh scripts. See ADR 0007.