De Olho no Fogo — point-level harmonization of MODIS and VIIRS hot spots

NASA Space Apps Challenge 2026, "Harmonization of MODIS and VIIRS Hot Spots". Team Space Uai, Ouro Preto, Brazil.

The product is a point-level harmonization on the MODIS ruler: every FIRMS detection from MODIS (Terra and Aqua, 1 km) and VIIRS S-NPP (375 m) is placed on the MODIS 1 km sinusoidal grid for its local solar day, and occupied cells that touch each other on the same day are merged into one harmonized focus. Hot spots that, at MODIS scale, can be the same fire count once, whether they come from neighbouring pixels, from the two sensors or from different overpasses.

Coverage: the whole world for 2025, and Brazil from November 2000 to December 2025. All figures below were generated on 2026-11-14 and are published in /data/harm/stats.json.

Earlier versions of the project converted counts with factors instead of merging detections. They are superseded; section 6 keeps their validation figures for comparison.

1. Point-level harmonization

1.1 Why the two records disagree

A fire a few kilometres wide lights several VIIRS pixels inside one MODIS pixel, the same fire is seen again on the next overpass, and a small fire is seen by VIIRS only. Adding the two lists counts the same fire several times; switching from one list to the other creates a step. In 2025, FIRMS has 3,502,268 MODIS and 17,014,675 VIIRS S-NPP vegetation-fire detections worldwide.

1.2 Algorithm

Implemented in scripts/harmonizar_global.py (Python, NumPy only).

1. Input. FIRMS archive files, MODIS Collection 6.1 (Terra + Aqua) and VIIRS S-NPP 375 m: the 2025 "all countries" yearly zips for the world, and the yearly country files for Brazil, 2000–2025. All confidence levels are kept. NOAA-20 is not used: we only hold it for 2025 and it would add a step to the series.
2. Type. type = 0 (presumed vegetation fire) enters the harmonization. type = 1 (active volcano), 2 (other static land source) and 3 (offshore) go to a separate fixed-source layer (section 2).
3. Local solar day. Local time = UTC + longitude / 15 h (UTC + 4 · lon minutes, floored to the minute). A detection belongs to its local date, so the ~01:30 night overpass and the ~13:30 day overpass of the same place fall on the same day everywhere on Earth, and a UTC midnight never splits a fire.
4. Grid. MODIS 1 km sinusoidal grid: sphere radius R = 6,371,007.181 m, cell = 926.625 m. row = floor(R·φ / cell), col = floor(R·λ·cos φ / cell), φ and λ in radians.
5. Occupied cells. Each detection occupies the cell that contains its point. Footprint rule: a MODIS detection with scan > 1 or track > 1 km also occupies the cells whose centre falls inside its footprint (± scan/2 east–west, ± track/2 north–south; the rows directly above and below are tested). VIIRS detections occupy only their own cell.
6. Harmonized focus. Connected component, with 8-neighbour connectivity, of the cells occupied on one local day, using both sensors and every overpass of that day. Sheared-grid neighbour lookup: away from the central meridian the sinusoidal grid is sheared, so the cell "above" is not in the same column. For the rows above and below, the neighbour is searched at the column of the same longitude, c' = floor((c + 0.5) · cos φ' / cos φ), and at c' ± 1.
7. Label. Each focus is both (has MODIS and VIIRS detections), MODIS only or VIIRS only.

Each focus stores: FRP-weighted centroid, local minute of the first and last detection, which satellites saw it (Terra, Aqua, S-NPP), whether it had day and night overpasses, number of 1 km cells with a detection (ncel), number of detections per sensor, the largest FRP per sensor (MODIS and VIIRS FRP are not summed: pixels and overpass times differ) and the exact set of cells it occupies. Raw detections are kept alongside, unchanged.

The clustering is also run with each sensor alone. The MODIS-alone foci are the ruler that exists for every year since 2000 (section 1.4).

1.3 Results, world, 2025

Of the harmonized foci that contain MODIS, 51.1% also contain VIIRS. A MODIS-only focus is not an error: Terra's morning overpass has no VIIRS counterpart, and cloud or smoke can hide a fire from one overpass.

Focus size, in 1 km cells: 1 cell 4,770,364 (65.7%); 2 cells 1,353,283; 3–5 cells 853,599; 6–20 cells 260,214; 21–100 cells 22,298; more than 100 cells 1,065; largest 1,490 cells.

Countries with most harmonized foci (FIRMS country file of most detections in the focus; 216 countries and territories): Democratic Republic of the Congo 999,303; Angola 502,889; Zambia 370,394; Brazil 359,348; India 359,055.

Measured overpass hours (2025 detections by local solar hour): Terra peaks at 09 h and 21 h, Aqua at 14 h and 02 h, S-NPP at 13 h and 01 h. Terra and Aqua are away from their nominal 10:30 and 13:30 crossings because of orbit drift, which is also why the long-series factor is calibrated on 2012–2021 only.

1.4 Brazil 2000–2025: a comparable long series

The same clustering is run for every year with the Brazil country files (detections inside Brazil only).

VIIRS does not exist before 2012, so the ruler across years is the MODIS-alone foci, available from 2000. The full harmonized total (MODIS ∪ VIIRS) is:

• measured from February 2012 on;
• estimated before that: full = MODIS-alone foci × f(month), with f(m) = Σ full ÷ Σ MODIS-alone foci over 2012–2021, one factor per calendar month.

f(month), January to December: 3.70, 3.36, 2.65, 2.62, 2.88, 2.96, 3.03, 2.78, 2.84, 3.48, 3.64, 3.74 (range 2.62–3.74; yearly mean 3.13). The calibration stops in 2021, before the MODIS orbit drift.

Validation (leave one year out). One year of 2012–2021 is hidden, f(month) is fitted on the other nine, and the hidden year's full total is predicted from its MODIS-alone foci.

Outside the calibration period, with the final factor: 2022 +2.71%, 2023 +4.16%, 2024 +14.05%, 2025 +6.57%. The errors are all positive and reach 14% in 2024, which is consistent with the MODIS orbit drift changing the relation between the sensors. The factor is therefore used only where VIIRS does not exist; from February 2012 on the series is the measured union.

1.5 Deviations and limits

• Global coverage is 2025 only. The long series exists for Brazil only.
• Terra-only months. Aqua enters FIRMS in July 2002; until June 2002 only Terra exists. For those months full = Terra-alone foci × f_Terra(month) × f(month), with f_Terra(m) = Σ MODIS-alone foci ÷ Σ Terra-alone foci over 2003–2011 (2.24–3.92). This step has no validation of its own: order of magnitude only. 2000 has November and December only.
• January 2012. S-NPP data start on 20 January 2012, so January 2012 is excluded from the calibration and reported as estimated.
• Pre-2012 totals for Brazil are estimates, with the errors of section 1.4. The measured quantity for those years is the MODIS-alone foci.
• Hot spots are not burned area, and a focus is not a confirmed wildfire. A focus is a count at MODIS scale: two fires less than one cell apart on the same day are one focus.
• Type 0 only. Detections with a local date outside 2025 are dropped from the global set (2,402 type-0 detections), and those of 31 Dec 2024 UTC that belong to 1 Jan 2025 local time are not in the 2025 files.
• Antimeridian. A fire crossing ±180° longitude becomes two foci.
• MODIS footprint. Treated as a rectangle aligned east–west, an approximation of the real scan geometry.
• Border foci. In the global run a focus on a border joins detections from both sides; the Brazil series clusters detections inside Brazil only. Hence Brazil has 359,392 foci in its own series and 359,348 in the global country table.
• Saturation of counts in the files. Per-focus detection counts are stored in one byte and saturate at 255 (1,163 of 7.26 million foci); ncel saturates at 65,535 (largest real value 1,490); overview-grid counts saturate at 65,535 (largest real value 5,320). stats.json totals are computed before packing and are not affected.
• One VIIRS satellite. S-NPP only; NOAA-20 and NOAA-21 are not in the series.
• Cloud, smoke and overpass time hide fires from both sensors. No focus does not mean no fire.

1.6 Second ruler: 5 km cell-days (Brazil, 2012–2025)

A second, coarser indicator built by the team runs beside the foci. It answers a different question: on how many 5 km squares, per day, did any sensor see fire?

• Type 0 only, no confidence filter, FRP not used (same inputs as section 1.2: the FIRMS Brazil files, 36 CSVs, 34,540,714 detections).
• Coordinates projected to EPSG:6933 (equal-area cylindrical); cell = floor(x / 5000 m), floor(y / 5000 m).
• Day = acquisition date in UTC. A cell with any detection on a day is one cell-day, counted once whatever the number of sensors or detections.
• Three products: MODIS, VIIRS S-NPP and VIIRS NOAA-20 (JPSS-1, from April 2018). Each keeps its own series; the integrated indicator is the union.
• Agreement: cell-days seen by MODIS only, by VIIRS only, by both, and the Jaccard index J = both ÷ union (0.13–0.21 per year).

Result: 9,618,064 integrated cell-days. It is a presence indicator, not a calibrated count of fires: it does not correct sensor sensitivity, overpass time, cloud or the number of observation opportunities.

The two rulers move together: yearly totals correlate at r = 0.998 over 2012–2017 and r = 0.997 over 2018–2025. Between the two periods the ratio of cell-days to foci steps from 1.08–1.11 to 1.53–1.60, because NOAA-20 enters the cell-day union in 2018 and is not in the foci. Compare years across 2018 with the per-product columns, or with the foci, not with the integrated cell-day column.

scripts/celulas_dia_5km.py rebuilds the monthly table from the FIRMS files; for the years where the same sensors are on disk (2012–2017 and 2025) all 84 months match the team's CSV exactly. Files: public/data/harm/celulas-dia-5km/calendario_{diario,mensal,anual}_nacional.csv.

2. Fixed heat sources

FIRMS flags detections that are not vegetation fires: type = 1 active volcano, 2 other static land source, 3 offshore. These never enter the foci. They are grouped into sites (neighbouring 1 km cells form one site) and shown as a separate layer.

Suspect sites are 1 km cells of type = 0 detections that look like a fixed source FIRMS did not flag (gas flaring, steelworks and similar): detections on at least 24 distinct local days spread over at least 8 months of one year, in a cell that neither coincides with nor touches an already flagged fixed source (372 candidate cells discarded for that reason in the world set, 119 in Brazil). For Brazil the rule is applied year by year. Suspect sites are not removed from the foci or from the raw points; the layer only warns.

Files: public/data/harm/fixas-2025.json and fixas-brasil.json.

3. Burn scars (Brazil)

scripts/cicatrizes_brasil.py reads the monthly burned-area shapefiles of INPE's Programa Queimadas, product AQ1km, Collection 2 (MODIS, 1 km, provisional), for the whole of Brazil, September 2002 to October 2026. The polygons are rasterised on the product's own 0.01° grid (4000 × 4000 cells, no resampling) and clipped to Brazil with the IBGE state mesh; pixels of neighbouring countries that come in the INPE files are discarded. One file per year (25 files, 23.4 MB in total) stores, for each burned cell, the months in which it burned.

Examples of yearly totals, cells that burned at least once in the year: 2024, 536,522 cells (648,031 km² with the real cell area); 2025, 414,750 cells (502,752 km²).

Caveats: the product is provisional and Collection 2 was reprocessed in January 2026; a 1 km pixel counts as fully burned, so areas are well above 30 m maps such as MapBiomas Fogo; INPE's official totals count 1 km² per pixel, while the real 0.01° cell is 1.03 to 1.24 km² in Brazil, so areas computed here are about 20% above the official ones; 2002 and 2026 are partial years; Brazil only. Contract: docs/dados-cicatrizes.md.

4. Data files and formats

Everything the globe reads is in public/data/harm/ (about 315 MB, about 8,460 files). The binary layouts, with JavaScript decoders, are specified in docs/dados-harmonizados.md (foci, raw points, overview grid, fixed sources, statistics) and docs/dados-cicatrizes.md (burn scars).

5. Reproduce

python scripts/baixar_firms.py            # FIRMS yearly country files for Brazil -> dados/firms/brasil/
python scripts/harmonizar_global.py       # about 4 min; writes public/data/harm/ (cache in cache/harm/)
python scripts/cicatrizes_brasil.py       # downloads what is missing from INPE; writes public/data/harm/cicatrizes/
python scripts/celulas_dia_5km.py         # checks the 5 km cell-day table against the team's CSV

harmonizar_global.py needs NumPy and the three inputs in dados/: modis_2025_all_countries.zip and viirs-snpp_2025_all_countries.zip (FIRMS yearly "all countries" archives) and dados/firms/brasil/*.csv.gz. Options: --so-brasil, --so-global, --refazer (ignore the cache), --proc N. The raw FIRMS files are not in the repository (size).

6. Superseded approaches (factor conversion)

Before the point-level method, the project converted counts with factors. These approaches are superseded and no longer served by the site (their pages and routes were retired on 2026-11-14 and moved to arquivo/retirado-2026-11-14/). The first was a regional prototype for one Brazilian state: monthly MODIS counts converted to the VIIRS scale by one factor per calendar month, which against INPE's single-satellite reference series reduced the 2012 sensor-change step from 4.39× to 1.14× (yearly correlation 0.98). Two factor-based reconciliations for the whole of Brazil followed; their validation figures are kept for comparison.

• Active 25 km cells, 2025 (scripts/harmonizar_celulas.py): MODIS-equivalent = active VIIRS cells × factor(IBGE macro-region, month). Split-half validation, 20 random splits: median monthly error 1.6% for Brazil (90% of months below 6.0%), 4.8% per region; a single factor gives 12.4%.
• Brazil 2000–2025 reconciliation (team package, rebuilt independently by scripts/conciliar_brasil.py): satellite value × factor(region, month) on the MODIS Terra+Aqua scale; VIIRS factor fitted on 2012–2021, Terra factor on 2003–2011. Leave-one-year-out error, 25 km cells: 2.5% yearly, 4.6% monthly, 6.6% per region (team package: 2.0%, 4.4%, 6.7%); fire counts: 1.3%, 4.4%, 6.8%. Drift of converted VIIRS against measured MODIS in 2022–2025: −1%, −4%, −9%, +1%.

Why they were replaced: a factor converts totals of a region and month and says nothing about any single fire. The point-level method keeps every detection, shows on the map which detections were merged and which sensor saw each focus, and needs a factor only where one sensor does not exist (Brazil before 2012). For that one use, its leave-one-year-out error (3.35% yearly, 6.38% monthly, on foci) is of the same order as the factor methods above (which measure a different quantity, so the figures are not strictly comparable).

Sources

NASA FIRMS (https://www.earthdata.nasa.gov/data/tools/firms, the resource listed by the challenge; archive downloads at https://firms.modaps.eosdis.nasa.gov), INPE Programa Queimadas (https://data.inpe.br/queimadas), IBGE (territorial meshes). Full list with licences and citations at /referencias.

Improvement tests (Brazil, 14 Nov 2026)

Run by scripts/melhorias_brasil.py on the FIRMS yearly files for Brazil, with the same grouping code as the product. Results in public/data/harm/melhorias-brasil.json. None of these change the published series yet; they measure what each change would do.

With Google Earth Engine and the FIRMS API (15 Nov 2026)

Run by scripts/ee_brasil.py (project registered for non-commercial use), scripts/melhorias_ee.py and scripts/brasil_2026.py. Results in public/data/harm/melhorias-ee.json and brasil-2026.json.

Still not done: global burn scars from GWIS (GlobFire perimeters); the GWIS server certificate had expired and the download was blocked.