Methodology & Documentation
Public reference for how MHEWS calculates, thresholds, and sources its data — no sign-in required. Aimed at Viewers and the public who want to understand what's behind a map layer, indicator, or alert before acting on it.
Contents
Overview
MHEWS is a decision-support platform, not an official alert-issuing authority (see the disclaimer in the footer of every page). This section documents how the platform turns raw monitoring/forecast data into the hazard, risk, and impact information shown on Dashboard, Forecasts, Impact-Based Forecasting, Early Warning, and Reports & Bulletins — so a reader can trace any number on screen back to its source and method, without needing an account.
Data sources
The same source names and cadences shown on the Dashboard's status pills and source list — documented here once, not re-described per screen.
| Dataset | Cadence | Used for | |
|---|---|---|---|
| CHIRPS rainfall (v3.0) | Daily | Drought/SPI indicator input | View methodology → |
| GloFAS flood levels | Hourly | Flood indicator + 3-day forecast | View methodology → |
| NDVI (HDX) | Dekadal composite | Drought/vegetation-stress indicator | View methodology → |
| INFORM risk index | Periodic (~monthly) | Risk layer, by hazard sub-index | View methodology → |
| RSMC cyclone outlook | 6-hourly | Cyclone track/wind-gust forecast | View methodology → |
CHIRPS rainfall (v3.0)
Climate Hazards center InfraRed Precipitation with Station data — a quasi-global, gridded daily rainfall estimate blending satellite imagery with in-situ station observations. MHEWS uses it as the primary Drought-indicator input, aggregated into the SPI (Standardized Precipitation Index) shown on Forecasts and Dashboard.
GloFAS flood levels
Global Flood Awareness System hydrological forecasts, refreshed hourly. MHEWS reads current river-level indicator values and the 3-day flood forecast from this feed for gauges within Madagascar's monitored basins (e.g. the Ikopa gauge referenced on Dashboard/Reports).
NDVI (HDX)
Normalized Difference Vegetation Index derived from MODIS satellite imagery, composited every 16 days. Used as a secondary Drought/vegetation-stress signal alongside CHIRPS rainfall; a "delayed" status pill reflects the underlying satellite composite's own refresh cadence, not a platform outage.
INFORM risk index
A composite risk baseline (hazard exposure × vulnerability × lack of coping capacity), following the INFORM methodology, sub-categorized by hazard on the Risk layer group. Updated periodically rather than in real time — see the "Estimate derived from Risk baseline — not a predictive model" caption shown wherever Impact figures appear on-screen.
RSMC cyclone outlook
Regional Specialized Meteorological Centre cyclone track and intensity outlook for the Southwest Indian Ocean basin, refreshed roughly every 6 hours. Feeds the Cyclone hazard tab's track-and-cone forecast field (wind-gust vector, not raster).
Indicators
MHEWS always separates two kinds of value for the same hazard: an Indicator (a current, observed reading — e.g. today's river level from a gauge) and a Forecast (what a model expects at a future lead time — e.g. GloFAS's 3-day projection). These are never merged into one number: Dashboard shows both per layer, Forecasts lets you compare Forecast sources on one chart against configured thresholds, and Reports & Bulletins' Situation section keeps a "Current (Indicator)" field separate from an "Expected (Forecast)" field for the same reason — so a reader always knows whether they're looking at what's happening now or what's predicted.
Triggers & thresholds
Each registered hazard/layer can carry threshold values with an operator (e.g. ≥/≤), a numeric value in whatever unit that layer measures (river depth in metres, wind gust in km/h, an index score), and a severity tier. Severity always uses the same fixed three-tier vocabulary everywhere on the platform — map legend, alerts, and report/bulletin text all read from one shared terminology source, never re-typed per screen:
Watch — early signal, low confidence, monitor closely
Warning — threshold confirmed, prepare to act
Severe — highest configured threshold breached, act now
Threshold changes go through a review/verification step by a second authorized user before taking effect (a visible "Pending verification" state on the Thresholds admin screen) — a threshold is never live the instant one person edits it. Each threshold also carries a sector-specific ResponseProtocol (who does what at each severity, tagged by sector — Agriculture/Water/Health/General) that the same protocol can drive a matching sector-targeted bulletin, alert distribution list, or Assistant framing.
Where the threshold values come from
Threshold values are configuration, set per deployment — but each starts from a published methodology, named here so a reader can check the platform against it:
| Hazard | Basis |
|---|---|
| Drought | Copernicus Combined Drought Indicator, three-stage scheme: SPI-3 ≤ −1, then soil moisture anomaly ≤ −1, then FAPAR anomaly ≤ −1. The accumulation period is part of the threshold — SPI-3 and SPI-12 are different measures and are never quoted as just “SPI”. |
| Flood | Return periods (2 / 5 / 20-year) derived from a discharge reanalysis, raised at ≥30% ensemble exceedance probability. |
| Cyclone | South-west Indian Ocean scale, based on 10-minute mean sustained wind, as set by the responsible regional centre. Severity is carried in from that authority, not recomputed by the platform. |
Two trigger tiers
Readiness opens an internal record and begins preparation. Activation makes external release possible. Not every hazard has both: where no forecast exists, no readiness tier is possible and the hazard carries an activation trigger only.
Two further rules stop a threshold producing noise. A confirmation rule requires a set number of consecutive evaluation cycles to agree before escalation — drought escalates only after two consecutive monthly cycles. A recovery class decides whether a brief return past the value holds the record open or resolves it, so a status does not oscillate. And a classification issued at long lead is capped one level down, because forecast uncertainty at long lead would otherwise produce false-alarm churn.
Onset class decides the surface
A sudden-onset hazard produces an alert. A slow-onset hazard is evaluated on a schedule and produces a bulletin. They never share a notification surface — a persistent drought must not bury a next-week cyclone. This follows from the hazard's onset class, which is recorded on the registry.
Risk, exposure and impact — what the numbers mean
A standing risk baseline describes an area's long-run risk. An impact estimate describes what a specific event may do. They are computed differently and must not be mixed — which is why the platform states, on the output, which sourcing mode and which estimate method produced a number.
Sourcing modes
| Component | A · Supplied | B · Computed | C · Absent |
|---|---|---|---|
| Exposure | Uploaded national population, settlement, infrastructure or crop layers, or a global product | Intersect the hazard footprint with those layers at event time | Very unlikely. Only hazard-based warning would be possible |
| Vulnerability | INFORM Vulnerability + Lack-of-Coping-Capacity dimensions, or an uploaded national assessment | Weighted composite from indicator layers at configured weights | Fall back to exposure-based warning: report exposed counts, claim nothing about consequence |
| Risk baseline | INFORM subnational risk maps | Hazard history × exposure × vulnerability × capacity, configured model | Omit the risk context panel. Warnings still function |
| Impact estimate | Rarely supplied — occasionally a partner assessment for a specific event | Computed at event time from the event's exposure and hazard layers | Report exposure only. This is honest, and often correct |
The INFORM decomposition rule
The full INFORM composite is used for the standing baseline only. At event time the platform uses only the Vulnerability and Lack-of-Coping-Capacity dimensions. Nothing from INFORM is ever reused as an event-time hazard or exposure input.
The reason is arithmetic, not preference: the INFORM composite already contains a hazard dimension and an exposure dimension. At event time the platform computes its own hazard footprint from live data and intersects its own exposure layers. Carrying INFORM's hazard and exposure through as well would count the same thing twice and inflate every downstream figure.
Impact estimate methods
| Method | What it does | What it may claim |
|---|---|---|
| Exposure only | Counts people and assets inside the event footprint | Who and what is exposed. Nothing about consequence |
| Historical analogue | Scales from the nearest comparable past event in the event catalogue | A range anchored to something that actually happened, with the analogue named |
| Susceptibility-weighted | Applies vulnerability and coping-capacity weights to the exposed counts | A weighted estimate, only where vulnerability data exists for the area |
The method is labelled on every impact output. All three are decision-support estimates, not predictive or causal claims.
Missing data
Missing data is never shown as 0 unless the source
explicitly defines zero as the value. This applies to forecasts, risk scores and impact
estimates alike — an absent value is shown as an explicit gap, and the platform
reports the narrower thing it can honestly say instead.
Sources: INFORM subnational risk methodology (JRC) · UNDRR-ISC hazard classification · UNDRR Sendai Framework Terminology.
Vulnerability methods
Every vulnerability dimension is produced by exactly one of three methods, a controlled vocabulary in the same spirit as the fixed Watch/Warning/Severe severity scale. Configured on Admin → Hazard & Data Registry's Vulnerability registry; the Vulnerability Dashboard's analysis workspace composes a layer from whatever is registered there.
| Method | What it is | Good fit for |
|---|---|---|
| Matrix | A small expert-defined lookup table: a handful of ordinal factors (Low/Medium/High) cross-referenced to a rating. | Qualitative dimensions with no continuous measurement — e.g. access to early warning information, institutional capacity. |
| Curve | A continuous indicator run through a fragility/susceptibility function (0–1), then bucketed into Watch/Warning/Severe by configured thresholds — the same threshold mechanism already used for hazard indicators. | Dimensions with a real measured value — poverty rate, substandard housing share, settlement density. |
| SMCE | Several normalized indicators combined at configured weights into one composite score (spatial multi-criteria evaluation) — the same idea INFORM's own composite uses. | Dimensions with several partial, related indicators and no single measure that captures them alone — e.g. environmental degradation. |
A dimension with no method configured at all is omitted rather than shown as a default or neutral rating — the same Mode C / Absent convention as the risk baseline above.
Worked example — Matrix
Access to early warning information: two ordinal factors, comms infrastructure and literacy, cross-referenced to a rating in a small expert-defined table configured once per indicator (Admin → Hazard & Data Registry → Vulnerability registry → Method configuration).
| Comms infra. ↓ / Literacy → | Low | Medium | High |
|---|---|---|---|
| Low | High | High | Elevated |
| Medium | High | Elevated | Moderate |
| High | Elevated | Moderate | Low |
The table itself is the configuration — an expert sets the cells once per indicator, not per area.
Worked example — Curve
A curve is registered for exactly one hazard × asset combination — a residential building responds to flood depth differently than it does to wind, and differently again from a school, so each combination is its own registered curve rather than one shared shape. Shown here: Flood × Residential buildings. Hazard intensity (river depth, metres) runs through the curve to a 0–1 vulnerability fraction, then buckets by configured thresholds — the same mechanism as an SPI or river-level threshold elsewhere in the platform. The final figure a reader sees (expected affected buildings, population, etc.) is exposure × vulnerability: the curve's output fraction multiplied by the exposed count for that area.
River depth 3m → vulnerability 0.62 → High. A different hazard or a different asset type at the same depth reads a different curve entirely — see Admin → Hazard & Data Registry to register one.
Worked example — SMCE (weighted composite)
Environmental degradation: several partial indicators, none of which alone represents the dimension, normalized to 0–1 and combined at configured weights — weights are an expert-judgement input, stated explicitly rather than hidden inside a black-box score.
| Indicator | Normalized (0–1) | Weight | Contribution |
|---|---|---|---|
| Land-cover loss rate (5yr) | 0.62 | 45% | 0.28 |
| Slope instability index | 0.40 | 30% | 0.12 |
| Distance to watershed (inverse) | 0.55 | 25% | 0.14 |
| Composite score | 0.54 | ||
Composite 0.54 → Moderate. Weights are configured per dimension, the same expert-input concept as INFORM's own indicator weights — not recomputed by the platform from raw data.
Multi-hazard methods
Two hazards affecting the same area are not always independent. The platform distinguishes Composite assessment (each selected hazard's risk stands alone and is blended into one district-wide picture) from Cascading / compound assessment (one hazard's occurrence changes the likelihood or severity of another) — a distinction drawn directly from the hazard interaction literature, not invented for this platform.
Hazard interaction types
| Type | Definition | Example |
|---|---|---|
| Triggering | The primary hazard directly causes the secondary one. | Cyclone rainfall triggers riverine flooding. |
| Compounding | Hazards co-occur (or occur in close succession) and their combined effect exceeds either alone. | Storm surge arriving during an already-elevated river level. |
| Amplifying | The primary hazard raises susceptibility to the secondary one without directly causing it. | Prior drought increases wildfire susceptibility; deforestation increases landslide susceptibility during subsequent heavy rain. |
Typology after Gill & Malamud (2014) and Kappes et al. (2012) — both review the ways natural hazards interact and propose this kind of classification as the basis for multi-hazard risk assessment, rather than treating every hazard pair as independent by default.
Cascading / compound parameters
| Parameter | What it captures | Basis |
|---|---|---|
| Triggering probability | P(secondary occurs | primary occurred) — how reliably the primary hazard actually produces the secondary one at this severity. | Gill & Malamud (2014) hazard-interaction matrix |
| Time lag | Hours between the two hazards' onsets. A short lag leaves little time to respond or recover before the next hazard lands. | De Ruiter et al. (2020), on consecutive disasters |
| Spatial footprint overlap | Whether the two hazards' affected areas actually coincide — a compounding effect requires spatial co-occurrence, not just temporal proximity. | Zscheischler et al. (2020), spatially compounding events |
| Coping-capacity depletion | How much a community's capacity to cope with the secondary hazard is reduced because it is still responding to the primary one — a vulnerability-side effect, distinct from the hazard-side parameters above. | De Ruiter et al. (2020) |
Combined risk is escalated one tier above the higher of the two independent hazard ratings only when triggering probability, time lag, and spatial overlap all indicate a genuine short-window, co-located cascade — never from a single parameter alone, and never silently: the workspace states which conditions held.
Sources: Gill, J.C. & Malamud, B.D. (2014). Reviewing and visualizing the interactions of natural hazards. Reviews of Geophysics. · Kappes, M.S., Keiler, M., von Elverfeldt, K., Glade, T. (2012). Challenges of analyzing multi-hazard risk: a review. Natural Hazards. · De Ruiter, M.C. et al. (2020). Why We Can No Longer Ignore Consecutive Disasters. Earth's Future. · Zscheischler, J. et al. (2020). A typology of compound weather and climate events. Nature Reviews Earth & Environment.
Layer catalogue
Every layer the platform can draw, where it comes from, and how often it refreshes. This is the same catalogue the platform runs on, not a re-typed copy.
How to read the dashboard
- The layer catalog is grouped into Hazard, Risk (INFORM, by hazard), Impact (population/buildings/infrastructure/productive assets, by sector), and Forecast (one main layer per hazard). Layers overlay simultaneously, each with its own independent opacity.
- An (i) info button on every layer row shows its source name, last-updated timestamp, and connection status (OK/Delayed/Failed) — the same small component reused everywhere a data layer is listed.
- A red "!" badge marks any admin area currently intersected by an active alert, at whatever admin level (Region/District/Commune) is currently displayed.
- Clicking a point on the map shows the exact value at that pixel and a rollup at the lowest available admin level — if a layer doesn't resolve that finely, you'll see an explicit "not available at this level" message rather than a blank or wrong number.
- The time control under the map applies to all layers by default, with a toggle to switch to per-layer time control, since sources refresh on very different cadences (e.g. cyclone outlook every 6 hours vs. NDVI every 16 days).