Daily fire detections per study region, scored against the region's own ±15-day seasonal baseline using a median and scaled median-absolute-deviation. Includes context features: night-time share, cell concentration, FRP per detection, satellite count.

Source

NASA FIRMS — MODIS and VIIRS active fire products

Grain

one row per region per day, including zero-detection days

Time column

observation_date

Detections aggregated into H3 resolution-7 hexagons (~5.2 km²). This is the map layer: sparse by design, because an absent cell means no detection rather than zero activity in a cell that was observed.

Source

NASA FIRMS

Grain

one row per H3 cell per day, only where detections exist

Time column

period_start

Daily hemispheric sea-ice extent scored against the published NSIDC 1981–2010 per-day-of-year normal and its own standard deviation. The strongest baseline in the pipeline: 30 years, published, needing no estimation from our rows.

Source

NSIDC Sea Ice Index v4.0

Grain

one row per hemisphere per day

Time column

period_start

Monthly mean sea-surface temperature anomaly, sampled at 1° and bucketed into H3 resolution-5 cells (~253 km²). Marine heatwaves are a shared precursor for Mediterranean fire risk and Norwegian glacier melt, which is why these cells belong on the same map as the fire cells.

Source

NOAA OISST v2.1 (via NOAA PSL OPeNDAP)

Grain

one row per H3 cell per month

Time column

period_start

Monthly mean NDVI per region with the same month a year earlier alongside, scored against that region's own distribution of year-over-year changes. A vegetation loss shows up as a negative z-score, so the anomaly test is on the lower tail.

Source

Copernicus Sentinel-2 via Google Earth Engine

Grain

one row per region per month

Time column

month_start

Blocked by: Sentinel-2 via Google Earth Engine. The collector and models are implemented; the backfill needs a Google Earth Engine service account.

Monthly Sentinel-1 VV backscatter per glacier region, scored against its own year-over-year change distribution. No published climatology exists for backscatter, so this baseline is weaker than the sea-ice arm's and `baseline_source` says so rather than implying equivalence.

Source

Copernicus Sentinel-1 via Google Earth Engine

Grain

one row per region per month

Time column

period_start

Blocked by: Sentinel-1 via Google Earth Engine. The collector and models are implemented; the backfill needs a Google Earth Engine service account.

Year-over-year NDVI or SAR change per H3 resolution-5 cell (~253 km²). Deliberately coarse: bucketing Sentinel at resolution 7 would mean 200k+ polygons per region in a single Earth Engine call, which no free quota absorbs.

Source

Copernicus Sentinel-1/-2 via Google Earth Engine

Grain

one row per H3 cell per composite window

Time column

period_start

Blocked by: Sentinel-1/-2 via Google Earth Engine. The collector and models are implemented; the backfill needs a Google Earth Engine service account.

Isolation Forest percentile score per region per day, ranked against the model's own training distribution so a percentile means the same thing now as it did at training time. Written by the batch scoring job, never by the transform.

Source

TerraSentinel ML layer

Grain

one row per region per day per model version

Time column

observation_date

One row per scheduled workflow execution: status, row counts, duration, and the GitHub run it came from. This is what the dashboard's freshness indicators read, so pipeline health is answered from data rather than from an uptime ping.

Source

TerraSentinel workflows

Grain

one row per workflow run

Time column

started_at

Upstream sources with their cadence, attribution text and credential presence. The dashboard footer renders from this, so a licence attribution cannot drift out of date.

Source

TerraSentinel configuration

Grain

one row per source

Time column

—