Skip to content

FIRMS — introduction#

NASA FIRMS (Fire Information for Resource Management System) distributes active-fire / thermal-anomaly detections from the MODIS (Terra/Aqua, C6.1), VIIRS (Suomi-NPP / NOAA-20 / NOAA-21), geostationary GOES (ABI), and Landsat (8/9) instruments. Each detection is a single fire pixel — a point with a location, an acquisition time, a brightness temperature, a detection confidence, and a fire-radiative-power (FRP) estimate. FIRMS serves both a near-real-time stream (detections within hours of overpass) and a standard-quality archive going back to the start of each mission.

This page orients the earthlens FIRMS backend. For the hands-on download walkthrough see Usage; for credentials see Authentication; the rendered API is the Reference page.

What earthlens returns#

A FIRMS detection is an inherently vector datum — a geolocated point, not a gridded cell — so the backend returns a pyramids FeatureCollection (a geopandas.GeoDataFrame, CRS EPSG:4326), one row per fire pixel:

Column Meaning
latitude / longitude detection centroid (WGS84)
acq_datetime acquisition timestamp (tz-aware UTC)
sensor the requested FIRMS sensor code
satellite reporting platform (Terra, Aqua, N, …)
confidence raw confidence as a string — "85" (MODIS) or l/n/h (VIIRS); use confidence_pct for numeric work
confidence_pct normalised 0-100 confidence (categorical tokens → 25/60/90; GOES passes its provider value through)
brightness_k brightness temperature in kelvin
frp fire radiative power (MW)
daynight D (day) or N (night) overpass
geometry Point(longitude, latitude)

This makes FIRMS a vector backend (FIRMS.OUTPUT_KIND == "vector"). Because a detection table is not a gridded array, the EarthLens facade rejects an aggregate= argument for this backend — the raster time-window reducer has no meaning on point detections. Post-process the returned GeoDataFrame directly instead (count per day, sum FRP, rasterise with pyramids, …).

Why it matters here#

The raster backends (ERA5 via CDS, MODIS/VIIRS imagery and the burned-area products via Google Earth Engine) give you gridded fields with full spatial coverage. FIRMS is the complementary event feed: sparse but direct per-pixel fire detections, refreshed within hours. A common workflow pairs FIRMS detections with a GEE burned-area product (MCD64A1, FireCCI) or a CDS meteorology field over the same bbox and window to study fire onset, spread, and weather drivers.

Things to know up front#

  • Detections, not perimeters. FIRMS reports where a sensor saw a thermal anomaly at overpass time — not the final fire perimeter or the total area burned. (FIRMS' data_availability index also lists BA_MODIS / BA_VIIRS, but those are burned-area raster products, not area-CSV active-fire sources — they return nothing through this backend and are deliberately not catalogued. For gridded burned-area rasters (MCD64A1, FireCCI) use the Google Earth Engine backend.)
  • A free MAP_KEY is required. Every request carries a MAP_KEY as a URL path segment. Request a free key at https://firms.modaps.eosdis.nasa.gov/api/map_key/; see Authentication.
  • NRT vs archive. *_NRT sensors cover only roughly the last two months; older data lives in the *_SP standard-quality archive sensors. A request for an old window against an *_NRT sensor returns an empty result, so the backend logs a warning naming the *_SP variant — it does not silently swap the sensor.
  • Confidence differs by family. MODIS reports a numeric 0-100 confidence; VIIRS reports a categorical l/n/h; LANDSAT reports l/m/h; GOES reports a provider-defined numeric value. The backend keeps the raw value and adds a uniform confidence_pct (categorical tokens map to 25/60/90) so one min_confidence= threshold works across MODIS / VIIRS / LANDSAT. GOES confidence is not a 0-100 percent, so min_confidence is skipped for GOES (with a warning) rather than silently dropping its detections.
  • A 5-day-per-request cap and a transaction quota. FIRMS serves at most 5 days per request and one sensor per request, and allows ~5000 transactions per rolling 10 minutes. The backend chunks longer windows transparently and throttles against the quota with a back-off — both are invisible to the caller. See Usage.