The problem
0
years in one archive
0
satellites, 2 sensors
0
districts and provinces
0
fire detections processed
+0.0%
harmonization error (held-out years)
raw MODIS: −75%
0.00
ROC-AUC of the risk model
The fire calendar
Each ring is a year, each slice a week. Season onset, peak weeks, the end of the season and unusual years are all found automatically.
Season
late June – late February
Peak period
late June – mid September
Unusual years
| Year | Total |
|---|---|
| 2001 | 7775 |
| 2002 | 7125 |
| 2003 | 5850 |
| 2004 | 8142 |
| 2005 | 4581 |
| 2006 | 6038 |
| 2007 | 4139 |
| 2008 | 4290 |
| 2009 | 3969 |
| 2010 | 6858 |
| 2011 | 4908 |
| 2012 | 4240 |
| 2013 | 4761 |
| 2014 | 3456 |
| 2015 | 8290 |
| 2016 | 5415 |
| 2017 | 5862 |
| 2018 | 4651 |
| 2019 | 7824 |
| 2020 | 4344 |
| 2021 | 4533 |
| 2022 | 4052 |
| 2023 | 2047 |
| 2024 | 1866 |
| 2025 | 4050 |
| 2026 | 2790 |
Two audiences
How it works
NASA FIRMS (MODIS C6.1, VIIRS S-NPP / NOAA-20 / NOAA-21), NASA POWER weather, NDVI, Open-Meteo forecasts.
Quantile mapping per region and season, conditional bias correction, and block-bootstrap uncertainty.
Season onset, peak weeks, critical windows and unusual years, all on one consistent record.
LightGBM: a calibrated probability for the next 7 days, a 13-week season outlook, and the reasons behind each.
Why now
The overlap period, when both sensors fly at once, is the only basis for calibration. Once it ends, building one consistent archive gets much harder.