Jump to the validation table

How Fireline harmonizes MODIS and VIIRS

Method version 1.0. How Fireline turns two generations of satellite fire detections into one monthly series from 2000 onward, how well that works in three test regions, and where it falls short.

1. The problem

NASA's Fire Information for Resource Management System (FIRMS) publishes hotspots: satellite pixels where a fire was burning as the satellite passed over. Two instruments produce them. MODIS flies on Terra (since 2000) and Aqua (since 2002). Its fire pixels are 1 km across at nadir and grow to roughly 2 by 5 km at the edge of a 2,340 km swath. VIIRS flies on Suomi NPP (since 2012), NOAA-20 and NOAA-21. Its fire product uses the 375 m I-band, and its pixels stay under about 0.8 km even at the edge of a wider 3,060 km swath.

Smaller pixels see smaller and cooler fires, and a fire that fills one MODIS pixel can show up as several VIIRS points. Over the same ground and month, VIIRS usually reports three to six times as many hotspots as MODIS. The satellites also look at different times of day. Terra crosses the equator around 10:30 and 22:30 local time; Aqua and the VIIRS satellites cross around 13:30 and 01:30. Fires follow a daily cycle, so each satellite samples a different part of it.

Plotted together, raw counts jump in 2012 because the instrument changed.

NASA plans to stop MODIS science data collection on Terra around February 2027 and on Aqua around September 2027, and FIRMS stopped delivering new Suomi NPP data on 2 November 2026. The 26-year record then continues only on NOAA-20 and NOAA-21, so Fireline expresses VIIRS in MODIS units.

2. Data

All detections come from FIRMS country files for Brazil, Canada and India, clipped to three regions: the Brazilian Pantanal, Western Canada (British Columbia and Alberta) and the Indian state of Punjab.

InstrumentSatelliteFIRMS products usedArchive starts
MODIS Collection 6.1, 1 kmTerra, AquaMCD14ML (standard), MCD14DL (near real time)Terra Nov 2000, Aqua Jul 2002
VIIRS Collection 2, 375 mSuomi NPPVNP14IMG (standard), VNP14IMGTDL_NRT20 Jan 2012; no new data after 2 Nov 2026
VIIRS Collection 2, 375 mNOAA-20VJ114IMG (standard), VJ114IMGTDL_NRT1 Apr 2018
VIIRS Collection 2, 375 mNOAA-21VJ214IMGTDL_NRT17 Jan 2024

Standard processing is the science-quality archive. FIRMS replaces near-real-time (NRT) data with it after a few months, so NRT only fills recent months. Exact date ranges per file are in data/raw/manifest.json.

Gaps in the current release. Standard data run through 31 December 2025 for all three satellites in use (Terra and Aqua, Suomi NPP, NOAA-20). NOAA-21 is not included yet: the FIRMS country archive does not carry it, so it needs an archive request. Data for 2026 wait on the NRT top-up, which needs a FIRMS MAP_KEY we do not have yet.

3. Cleaning

We keep only detections FIRMS classifies as vegetation fire (type = 0), which drops volcanoes, gas flares, other static heat sources and offshore detections. NRT files have no type field, so there we drop grid cells that appear in at least half of all months across five or more years of standard data. A cell that is hot that often is almost always a flare or an industrial site.

MODIS detections need confidence 30 or higher. VIIRS detections need nominal or high confidence.

Each detection is dated by local solar day (UTC plus longitude / 15 hours), so a 01:30 night pass lands on the local day it happened. When a day exists in both standard and NRT data, only the standard record is kept.

4. Common unit: fire cell-days

Counting points rewards small pixels. Fireline instead lays a hexagonal grid over the ground (H3 resolution 8, cells of about 0.74 km², close to one MODIS pixel at nadir). A fire cell-day is one cell with at least one detection from one satellite on one local day. Seven VIIRS points in one cell on one day count once, like a single MODIS point.

The MODIS series adds Terra and Aqua, which look at different times of day. The VIIRS series is the mean across VIIRS satellites flying that day, since they pass at about the same time; the launch of NOAA-21 does not add fire. Monthly totals are stored on a coarser grid (H3 resolution 6, about 36 km² per cell) with the number of days that had data.

5. Calibration

For each region and calendar month there is one scale factor:

k = total MODIS cell-days / total VIIRS cell-days, over 2012 to 2019.

There is no intercept, so a month with no VIIRS fire gets no harmonized fire. Each month gets its own k because the ratio depends on the kind of fire: Punjab's small crop-residue fires in autumn look different to MODIS than Canada's large summer forest fires. The fit stops at 2019, while Terra and Aqua were still in their designed orbits.

Months with little fire give a noisy k, so we pull them toward the region-wide k:

k_hat = (n × k_month + n0 × k_region) / (n + n0)

Here n is the number of MODIS cell-days behind that month's ratio and n0 = 50. A month with 50 MODIS cell-days gets half its own ratio and half the regional one; a month with 5,000 keeps almost all of its own.

For uncertainty we resample the eight fit years with replacement 1,000 times (fixed seed), refit k each time, and keep the 5th and 95th percentiles as k_lo and k_hi, a 90% band. We resample whole years because months in the same year share weather and fire season.

6. Splice and the post-2027 continuation

PeriodHarmonized valueSource label
Nov 2000 to Jun 2002Terra × the (Terra + Aqua) / Terra ratio for that calendar month, fitted on Jul 2002 to Dec 2007terra_scaled
Jul 2002 to Dec 2019Observed MODIS (Terra + Aqua)modis
Jan 2020 onwardk × VIIRS, with a band from k_lo × VIIRS to k_hi × VIIRSviirs_scaled

Before Aqua launched, Terra months are scaled up to the two-satellite level. Switching to VIIRS in 2020 avoids using MODIS while its orbits drift. When MODIS stops in 2027 nothing changes, because the series after 2020 already comes from VIIRS; NOAA-20 and NOAA-21 carry it forward.

Months built from NRT data, and the current partial month, are flagged provisional. The app sums stored cells inside any area a user draws and compares each month with the median of that calendar month over 2003 to 2022.

7. Validation

Each region is tested separately:

MetricExpectedPantanalWestern CanadaPunjab
Raw bias, VIIRS / MODISabout 3x to 6x2.18x2.98x4.44x
Harmonized bias, harmonized / MODISclose to 1x1.06x1.18x1.14x
Leave-one-year-out rclose to 10.980.980.97
Leave-one-year-out NRMSEclose to 00.230.390.29
Forward r, 2020 to 2025close to 1 (drift lowers it)0.990.990.98
Seam ratio at Jan 2020close to 1x2.18x0.25x1.31x
Change-point p at Jan 2020above 0.05 (no step)< 0.0010.740.01
Peak-month match, 2012 to 2019close to 100%88%100%100%

pipeline/render_method.py fills every value in this table from validation.json; "n/a" means the metric could not be computed. Per-region scatter plots (harmonized against MODIS) and seam charts are listed under figures in the same file.

The seam test does not work as a check on the method. At the switch the ratios are 2.18x (Pantanal), 1.31x (Punjab) and 0.25x (Western Canada). At the placebo dates, where the series is MODIS on both sides or VIIRS on both sides, the ratios run from 0.50x to 11.21x, and 3 of 9 placebo change-point tests give p below 0.05. A 12-month before/after ratio mostly measures real year-to-year swings in fire. The Pantanal's record 2020 season falls in the first year after the switch, and its change point (2019-01) marks the start of the 2019 to 2020 run of severe seasons, not the sensor change. So the seam test cannot tell a sensor artifact from a real change. The case that the splice is sound rests on the overlap bias (raw 2.18x to 4.44x, harmonized 1.06x to 1.18x) and on the leave-one-year-out and forward tests.

Independent check: Sentinel-3 SLSTR

ESA's Sentinel-3 satellites carry SLSTR, which has its own fire algorithm and a morning overpass (about 10:00 local time). Its monthly fire detections over the Pantanal from September 2020 to December 2025 (64 months) are compared with Fireline's harmonized series. Spearman rank correlation: 0.87. The two series peak in the same month in 80% of years. Details and the chart are in sentinel3.md.

8. Limitations

Different overpass times. MODIS combines morning and afternoon views; VIIRS sees early afternoon and the middle of the night. k absorbs the average difference over 2012 to 2019. If fires shift to a different time of day, for example farmers burning later in the afternoon, that k no longer fits.

No VIIRS morning orbit after Terra. Once Terra stops, nothing in the record looks at mid-morning. Fire only a morning pass would catch is inferred through k, not observed.

Terra and Aqua orbit drift. Terra made its last inclination maneuver in February 2020 and left its constellation in October 2022; its crossing time has drifted earlier since. Aqua stopped holding its 13:30 crossing in 2021 and began leaving the A-Train in January 2022; it drifts later. This is why the fit ends in 2019 and why the forward test is not a clean check.

VIIRS satellites are treated as interchangeable. k was fitted on Suomi NPP alone until April 2018, then on Suomi NPP plus NOAA-20, and from November 2026 it is applied to NOAA-20 and NOAA-21 only. Any systematic difference between the instruments passes into the series.

NRT months are provisional. NRT data have no type field, can have geolocation errors and are reprocessed later. Those months will change when standard data replace them.

Small areas are noisy. k is a regional average per calendar month. A small drawn area may burn differently from its region, and a few cell-days either way swing its numbers. The bands cover uncertainty in k only, not clouds, smoke or missed overpasses.

Static sources in NRT. The persistent-cell mask removes most flares and industrial heat, but a new plant, or a field that burns most months, can be misclassified until standard data arrive.

Thin months can stay poorly constrained after shrinkage. In Punjab, August (monsoon) had 101 MODIS and 43.5 VIIRS cell-days over 2012 to 2019. That gives k = 1.64 with a 90% band of 0.79 to 3.39, against 0.15 to 0.55 in every other month. The effect on annual totals is negligible, because August is about 0.1% of Punjab's fire activity.

Missing data. NOAA-21 is not in the series yet, and 2026 is not covered until the NRT top-up runs (see section 2). Until then the record ends in December 2025.

9. References

Chen, Y., van der Werf, G. R., Mu, M., van Wees, D., Hall, J. V., Giglio, L., Vernooij, R., Morton, D. C., Xu, L., Liu, T., Scholten, R. C., Coffield, S. R., Follette-Cook, M. B., Ott, L., and Randerson, J. T. (2026). Tracking recent extremes and interannual variability of global fire emissions using a near-real-time extension to the Global Fire Emissions Database. Earth System Science Data, 18, 7535-7555. https://doi.org/10.5194/essd-18-7535-2026

Giglio, L., Schroeder, W., Hall, J. V., and Justice, C. O. (2021). MODIS Collection 6 and Collection 6.1 Active Fire Product User's Guide, Version 1.0. NASA. https://www.earthdata.nasa.gov/s3fs-public/2023-09/MODIS_C6_C6.1_Fire_User_Guide_1.0.pdf

Li, F., Zhang, X., Kondragunta, S., and Csiszar, I. (2018). Comparison of fire radiative power estimates from VIIRS and MODIS observations. Journal of Geophysical Research: Atmospheres, 123, 4545-4563. https://doi.org/10.1029/2017JD027823

Pettitt, A. N. (1979). A non-parametric approach to the change-point problem. Journal of the Royal Statistical Society, Series C (Applied Statistics), 28(2), 126-135. https://doi.org/10.2307/2346729

Schroeder, W., Giglio, L., and Hall, J. (2025). Collection 2 Visible Infrared Imaging Radiometer Suite (VIIRS) 375-m Active Fire Product User's Guide, Version 1.2. NASA. https://www.earthdata.nasa.gov/s3fs-public/2025-06/VIIRS_C2_AF-375m_User_Guide_1.2.pdf

Schroeder, W., Oliva, P., Giglio, L., and Csiszar, I. A. (2014). The New VIIRS 375 m active fire detection data product: Algorithm description and initial assessment. Remote Sensing of Environment, 143, 85-96. https://doi.org/10.1016/j.rse.2013.12.008

Vadrevu, K., and Lasko, K. (2018). Intercomparison of MODIS AQUA and VIIRS I-band fires and emissions in an agricultural landscape: Implications for air pollution research. Remote Sensing, 10(7), 978. https://doi.org/10.3390/rs10070978

Wooster, M. J., Xu, W., and Nightingale, T. (2012). Sentinel-3 SLSTR active fire detection and FRP product: Pre-launch algorithm development and performance evaluation using MODIS and ASTER datasets. Remote Sensing of Environment, 120, 236-254. https://doi.org/10.1016/j.rse.2011.09.033

Datasets

Fire data were provided by LANCE FIRMS, operated by NASA ESDIS with funding from NASA Headquarters.

Validation by region

Loading validation metrics...