MODIS · Terra 10:30, Aqua 13:30
1 km pixels, since 2000.
Two satellites, morning and afternoon. A fire must heat enough of a square kilometre to be flagged, so small fires and short burns are often missed.
Method
How we make 25 years of fire records comparable.
MODIS and VIIRS see the same planet differently. Before we can say a month is unusual, we remove three differences between the sensors, then check that the join at 2012 has disappeared, with tests taken from the published work on these two sensors.
The numbers on this page come from our mock data set, which simulates known fires so we can test the method. They are replaced by results on the real NASA FIRMS archive from November 13.
VIIRS · Suomi NPP, NOAA-20, NOAA-21 · 13:30
375 m pixels, since 2012.
Seven times smaller pixels see far more of the same fires. The record has to bridge the two without inventing a jump in 2012.
The fixes
Three differences, three fixes.
Pixel size
Problem: MODIS pixels are 1 km wide, VIIRS pixels 375 m. One fire can light up one MODIS pixel but several VIIRS pixels.
Fix: count both on the same 0.1° grid (about 10 km), as detections and as fire days (days a cell burned at all), and keep the metric that validates better.
Sensitivity
Problem: VIIRS detects smaller and cooler fires, so it reports several times more hotspots over the same ground.
Fix: split VIIRS at the fire power MODIS needs to see a fire (its detection floor grows with pixel size). Fires above it convert from MODIS by a ratio learned on 2012–2022, per 10° tile and month. Fires below it, which MODIS can't see, come from each cell's own record in the VIIRS years, scaled by how active that year was. So quiet places don't turn into a patchwork of zeros before 2012.
Overpass time
Problem: Terra passes at 10:30, Aqua and VIIRS around 13:30. Fires are stronger in the afternoon, so the time of day changes the count.
Fix: calibrate Aqua against VIIRS (same pass) on 2012–2022, before Aqua's orbit drift, and Terra against Aqua on 2003–2011, day and night separately. NOAA-20, then NOAA-21, fill S-NPP gaps after their own calibration, and carry the record since S-NPP stopped on November 2, 2026.
From orbit to one record
The pipeline.
- NASA FIRMSMODIS C6.1 from 2000, VIIRS 375 m from 2012
- CleanLow confidence, volcanoes, gas flares and other static sources removed
- Grid0.1° cells × months: detections and fire days, per satellite and pass
- SplitVIIRS fires above and below MODIS's detection floor
- CalibrateFour models fitted on the overlap and scored year by year; the best one is used
- One recordCalendar, statistics, CSV / GeoJSON / Cloud-Optimized GeoTIFF
Validation
Does it work? We check it.
0.976R² predicting VIIRS from MODIS for years left out of the fit (leave-one-year-out, 1,341 10° tile × month totals, day + night)
2.87× → 1.03×Hotspots after 2012 vs before: raw join vs our record (1.00× = no break)
p 0.0003 → 0.85Pettitt change-point test: a clear break in the raw join, none in our record
0.85× to 1.08×The same after-vs-before ratio at fake break years inside one sensor era (2005, 2007, 2017, 2020): our 2012 ratio is within that natural range
Raw MODIS then VIIRS, glued togetherHarmonized (our record)Mock data until Nov 13
2012: VIIRS arrives. Raw counts leap; same fires, sharper eyes.
Overlap years: MODIS vs VIIRS
Models compared
| Model | R² tile | R² 1° | Bias |
|---|---|---|---|
| Ratio | 0.974 | 0.955 | 1.12 |
| Log-linear | 0.963 | 0.93 | 1.54 |
| Quantile mapping | 0.968 | 0.944 | 1.9 |
| Split at the MODIS floor (used) | 0.976 | 0.968 | 0.8 |
We use the M3 model: it scored best when predicting years it hadn't seen. 11 of 97 strata had too little overlap and used a larger group.
Against the known fires (mock only)
| Region | Raw jump | Ours |
|---|---|---|
| Nile Delta | 11.7× | 1.08× |
| Río Negro | 2.24× | 1.01× |
| Amazon | 2.67× | 1.02× |
| Central Africa | 3.18× | 1.01× |
| Northern Australia | 2.45× | 0.99× |
| California | 1.72× | 0.95× |
The mock data knows the true fires. Hotspots per true fire day, after 2012 vs before: the raw join jumps, our record stays near 1.
Behind the calendar
The formulas.
Conversion
VIIRS-equivalent = k × MODIS + small × r. k = Σ VIIRS above the floor ÷ Σ MODIS, per tile, month and pass. small is the cell's mean VIIRS count below the floor for that month in the overlap years; r = (tile's MODIS count + 5) ÷ (its usual count + 5). The floor is 4.44 × A + 0.52 MW by day and 4.43 × A + 1.01 MW by night, A the MODIS pixel area (Sofiev and Kouznetsov 2026). Ratio, log-linear and quantile-mapping models are scored the same way.
Percentile rank
The share of baseline years with a lower count for the same month. The baseline is 2001 to last year, without the year being judged.
z-score
(count − mean) ÷ standard deviation of the baseline. z ≥ 2 is "unusual", 1 to 2 "above normal", ≤ −1 "below normal". The trend is Sen's slope with a Mann-Kendall test.
Limits
What the data can and can't tell you.
A hotspot is not a fire
It is a pixel that looked hot from orbit. One fire can make many hotspots; a gas flare can make one too.
No hotspot ≠ no fire
Clouds, thick smoke, small fires, or burning between satellite passes can all hide a fire.
Counts, not area burned
The calendar shows how often an area was detected burning, not how many hectares burned.
Small areas are coarse
The record lives on a 0.1° grid (about 10 km). Areas under four cells get a warning and the raw points.
Data sources
- NASA FIRMS: MODIS Collection 6.1 active fires (Terra, Aqua)
- NASA FIRMS: VIIRS 375 m active fires (Suomi NPP, NOAA-20, NOAA-21)
- NASA GIBS: Blue Marble Next Generation and VIIRS Black Marble imagery for the map and globe
- NASA 3D Resources: the Aqua and Suomi NPP satellite models
- Higgsfield: the generated films and portraits on the landing page (marked as generated)
- geoBoundaries: administrative boundaries · Nominatim / OpenStreetMap: place search
What the method builds on
- Schroeder et al. 2014, Remote Sensing of Environment 143: the VIIRS 375 m product and how it compares with MODIS. doi
- Li et al. 2018, JGR Atmospheres 123: MODIS and VIIRS fire counts and FRP across the swath, and matching near-simultaneous passes. doi
- Giglio et al. 2006, JGR Biogeosciences 111: MODIS fire counts corrected for overpasses, and Terra vs Aqua. doi
- Chen et al. 2023 (GFED5) and 2026 (GFED5NRT), Earth System Science Data: scaling between MODIS and VIIRS by region, land cover and month. doi · doi
- Sofiev and Kouznetsov 2026, EGUsphere preprint: MODIS and VIIRS FRP detection limits as a function of pixel area. doi
- Giglio, May 2025, LANCE user group: Terra and Aqua orbit drift and fire products. slides