Method
How two satellites
become one history.
Pilot region: Central Brazil (Cerrado), bounding box 60°W–44°W, 20°S–8°S. Data: NASA FIRMS, MODIS (Terra and Aqua) from Nov 2000 and VIIRS S-NPP from Jan 2012.
1. Common filters
Only detections of type 0 (presumed vegetation fire) are kept. MODIS needs confidence of at least 30%; VIIRS drops the "low" class. This removes most volcanoes, industry and doubtful pixels before any comparison.
2. Monthly counts
Counts are summed per calendar month and per sensor. Monthly totals smooth out the different overpass times of each satellite.
3. Calibration on the overlap
From Feb 2012 both sensors watch the same ground. We fit log(1 + VIIRS) = a + b · log(1 + MODIS). The model is trained on 2012–2019 and tested on 2020 onward, which the fit never saw. Before mid-2002 only Terra existed, so a second fit uses Terra alone.
4. The harmonized series
Where VIIRS observed the month, we use it. Before 2012 the value is estimated from MODIS and expressed in VIIRS-equivalent detections. Estimated months are drawn faded in the calendar.
5. Anomaly
Each month is divided by the median of the same calendar month across the whole series, so a high August is compared with other Augusts, not with a quiet February.
Validation
- Hold-out R² (2020+)
- —
- Hold-out error
- —
- Overlap months
- —
Limits, stated plainly
This is a regional statistical calibration, not a pixel-level match. It tells you how much fire a month had compared with normal; it does not locate each fire better. Estimates before 2012 carry more uncertainty than the hold-out error suggests. A hot spot is a thermal detection, not a confirmed wildfire. Only one region has been calibrated so far.