Proof, method and validation
How do we know the repair is real?
Three steps, tested on years the model never saw, checked against a sensor that never changed, and honest about where it is weakest. Countries with enough data get their own fitted curve; smaller ones borrow a shared (pooled) curve, counted and marked below. These are regional results, not a global correction.
The method in 60 seconds
01
The problem
In 2012 the record switched from MODIS (1 km) to VIIRS (375 m), which sees many more small fires. Joined naively, fire counts jump overnight.
02
The fix
From fires both satellites saw, we learn per country how likely MODIS was to see a fire of a given strength, then count VIIRS fires the way MODIS would have.
03
The test
Fitted on 2012–2018, scored on 2019–2021, years the model never saw. The 2012 jump disappears in 146 of 171 countries; past disasters rank where they should.
04
The limits
Estimates, not MODIS observations. Typical monthly error ±31%, larger where fires are rare. A simple ratio nearly ties us on one score, and we show it.
Proof, one test, four methods
Four corrections, the same held-out test
Proof, the 2012 step
The sensor jump collapses onto the unchanged sensor
Each country's change in yearly fire counts from 2009–11 to 2012–14. Across: Aqua MODIS alone, a sensor that did not change. Up: the same step in our harmonized record (ember) and in the naive splice of MODIS onto VIIRS (blue). On the diagonal the record moves exactly as the unchanged sensor does. Hover or tap a dot.
Proof, known fire disasters
Does it find the fires everyone remembers?
Widely documented extreme fire seasons, listed before any result was computed. Each season's total is ranked against the same months in every other year; rank 1 is the largest. The raw splice is MODIS (Terra and Aqua) to 2011, then Suomi-NPP VIIRS from 2012.
01, Match
Each VIIRS detection gets a yes/no label: “seen by Aqua” if Aqua MODIS detected fire on the same UTC date in the same small grid cell (H3 resolution 7) or one of its 6 neighbours, about 3.5 km. Both satellites pass at about 13:30 local time in 2012–2021, so they look at nearly the same moment.
02, Learn
From 2012–2018 we learn, per country, the chance that MODIS sees a fire, given how much energy it radiates. Big fires: almost always. Small ones: rarely.
03, Rebuild
Every VIIRS detection is weighted by that chance. The total is the count a MODIS-class sensor would have made, comparable back to 2003.