All public data, and the whole Canadian pipeline runs with no API keys at all. The NRCan feeds below are published through the CWFIS Datamart.

Source	What it gives	Role
CWFIF national GeoServerNRCan · WFS	Every reported fire with its full revision history: size, control status, cause, response type. The bitemporal stamps on these records are what make an honest forecast possible.	Fire state and the label
CWFIS satellite hotspotsNRCan · daily + season archives	Satellite detections already carrying Canadian fire-behaviour outputs: head fire intensity, rate of spread, fuel type and consumption. 9.1 million detections across 2023–26.	The strongest features
CWFIS station weather & FWINRCan · decadal archive + daily	Noon-local readings from ~2,250 stations: temperature, humidity, wind, rain and all six Fire Weather Index System codes — taken whether or not anything nearby is burning, which is the whole point.	The ignition model's only exogenous input
Natural Earth	Provincial and territorial boundaries, simplified from 705 KB to 23 KB so the page carries its own basemap. Also the country filter for the ignition grid.	This map, and the study area
CIFFC situation reportvia api.ciffc.net/v1/sitrep	National and per-agency preparedness levels. The human judgement about how stretched crews and aircraft are, which no machine feed produces.	Ingested. Measured effect: none
Open-Meteo ERA5 archive	Hourly reanalysis weather back to 1940: temperature, humidity, wind, gusts, precipitation. No key required.	Already covered by hotspot FWI
NASA FIRMSMODIS & VIIRS	Near-real-time active fire detections worldwide, within about three hours of satellite overpass.	Same satellites as CWFIS, less detail

Why three of these are not feeding the model

CIFFC preparedness is a result, not a gap. It is ingested, backfilled to 2019, and joined onto 98.5% of fires with a point-in-time-safe as-of join on publication time. Measured on escalation with a paired bootstrap, dropping the whole block gains 0.011 PR-AUC, with an interval straddling zero: no detectable benefit.

The defence used to be that it was simply the wrong shape — one value per agency per day, so every fire burning in one province that day shares it, and day-of-year plus agency already encode most of that seasonal load pattern. An agency-and-day covariate ought to suit a model whose unit of prediction is a cell and a day. So it was re-tested there, on the ignition model, where the sitrep even carries each agency's own forecast of tomorrow's lightning- and human-caused ignition load — very nearly this model's target, written down by a human. It did nothing there either. Reporting it as helpful would be easier and false.

Open-Meteo would mostly repeat what the hotspots already carry. Every satellite detection arrives with the Fire Weather Index computed at that pixel, and FWI is precisely a compression of the drought, wind and humidity history that ERA5 would supply. So the marginal gain is unproven, while the cost is one request per grid cell per date window, thousands per season against a free public service. The client is built and a flag turns it on; the experiment to show it earns that cost has not been run, and until it has, leaving it off is the honest default rather than the lazy one. The station-weather result on the ignition model — a real, non-circular fire-weather feed worth about a thousandth of a PR-AUC once the model already knows where and when — suggests what that experiment will find.

NASA FIRMS is a strict subset here. It publishes raw detections from the same satellites CWFIS uses, VIIRS and MODIS, but NRCan ships them already enriched with Canadian Fire Behaviour Prediction outputs: head fire intensity, rate of spread, fuel type, fuel consumption. Using FIRMS instead would mean re-deriving fire science that has already been done properly. It earns its place for two things this project does not currently need: latency (about three hours after overpass, against next morning) and coverage outside Canada. It is also the only source that requires a key.

Does it work

How it works

Where the data comes from