Methodology

How SENTINEL works.

This page exists to be the honest answer when someone asks "where did that number come from?" If it can't be answered here, the number shouldn't be on the site.

1. Live agency feeds

SENTINEL polls the public emergency management feeds for each state and territory on a regular cadence:

• NSW RFS — Major Incidents JSON feed, polled every 5 minutes
• VIC EMV — Public OSOM GeoJSON, polled every 5 minutes
• QLD QFD — Bushfire alert S3 endpoint, polled every 15 minutes
• WA DFES — Emergency Warnings API, polled every 5 minutes
• NSW Live Traffic — Road incident hazards, polled every 60 seconds during incidents
• USGS — Australian region earthquakes (M≥4), polled every 5 minutes
• AEMO — National Electricity Market dispatch summary, polled every 5 minutes
• Open-Meteo (BoM ACCESS-G mirror) — Weather forecasts for capital cities and key regional points

We don't poll agency APIs once per user — we poll once centrally and serve all users from cache. This is friendlier to the agencies and faster for everyone.

2. Historical event data

The historical events catalogue is curated manually. Every event has:

• Name, start/end dates, affected states
• An extent polygon sourced from the relevant agency (RFS fire perimeters, SES flood extents, BoM warning polygons)
• A plain-language summary of what happened
• Key statistics (deaths, properties destroyed, area burned, peak gauge level) — each with a citation
• A list of source documents used to compile the entry
• A published date and a most-recent-review date

If we can't find authoritative data for a statistic, the statistic doesn't appear. We will not invent numbers to fill space.

3. Risk overlays (modelled)

SENTINEL does not generate flood models, fire spread models, or storm surge models. We display modelled risk produced by authoritative sources and tell you what each model is and isn't.

For flood extent, we draw from:

• NSW Flood Data Portal (council and state flood studies)
• Geoscience Australia inundation products (national scale)
• Local council flood studies where published

For bushfire-prone area mapping:

• NSW RFS Bush Fire Prone Land mapping (statutory)
• Equivalent state datasets where available

Each overlay carries the source name, publication date, and a list of known limitations. A 1% AEP flood extent from a 2018 model is not the same as a 1% AEP flood extent from a 2024 model. We surface the vintage.

4. Address geocoding

We use OpenStreetMap's Nominatim service to convert addresses to coordinates. This is accurate in metro areas, generally accurate in regional areas, and sometimes coarse in remote areas. Each geocoded result includes a confidence rating — high (house-level), medium (street-level), low (suburb or coarser). When confidence is low, we say so.

Geocoded results are cached in our database for 90 days so the same address doesn't get re-geocoded on every lookup. This is also Nominatim's published usage policy expectation.

5. Comparison engine

When SENTINEL compares two events, it computes:

• Spatial overlap — the Jaccard index of the two extent polygons (intersection / union)
• Quantitative comparison — like-for-like statistics where available (peak gauge, max wind speed, area burned)
• Consequence comparison — population in extent, schools/hospitals affected, road segments inside extent

The comparison says this is what changed and this is what's similar. It does not say this means X will happen. Forecasting is the job of agency models. We're showing data, not predicting outcomes.

6. What we don't do

• We don't issue warnings
• We don't recommend actions
• We don't quantify personal property risk in dollar terms
• We don't predict where fires/floods will go next
• We don't replace agency advice — we direct users to it

7. Known limitations

See the limitations page for a current list of known gaps in our data and method. We update it as we find new issues.

8. Telling us we're wrong

If you find a number that's wrong, an event that's missing, or a methodology issue, please email us. We log every report and respond. Corrections to historical event entries are made visibly with an edit date.