NASA Space Apps Challenge 2026 · Harmonization of MODIS and VIIRS Hot Spots (CONAE) · Team GeoFuego.com.ar
Argentina's hotspot record is two records: MODIS since 2000 and VIIRS since 2012. VIIRS sees about 3 fire pixels where MODIS sees one, and the ratio changes with region, calendar month, satellite overpass time and the kind of heat source. Glued together, the record shows a false jump in 2012; kept apart, a fire manager cannot compare this season with 2008. This page says how the two were made one, how well it works on years the model never saw, and what it cannot know.
In 2012 the satellite changed. The record should not jump with it.
The app shows this panel for any selected area (raw MODIS, raw VIIRS, the splice and the harmonized series with its 80 % range).
1 · The unit: VIIRS-equivalent hotspots on a 0.25° grid
Every value is a count of satellite fire detections expressed in VIIRS-equivalent pixels (375 m), per 0.25° cell (≈ 25 × 25 km; 4 193 cells cover Argentina) and calendar month, November 2000 to today. From 2012 the values are observed by VIIRS S-NPP; before 2012 they are estimated from MODIS. Counts are not a number of fires and not burned area: a fire that burns for several days is counted on every satellite pass.
Why pixel counts and not cell-days. A cell-day unit (one count per cell with fire on a day, whatever the pixel size) removes the pixel-size effect by construction, but it saturates: once a cell has fire on a day, more fire inside it does not count, so the largest seasons are flattened. In Corrientes in January–February 2022, the biggest fire season of the record (45 451 VIIRS-equivalent hotspots in the year, 2.4× the median year), most of the province burned at once; a count of cells with fire cannot say by how much. Pixel-equivalent counts keep that intensity, and the challenge asks which years and months had the highest fire activity. The price is the conversion, which is calibrated and carries an interval.
2 · The factor: per cell × calendar month × satellite set, with shrinkage
- A VIIRS/MODIS factor per 0.25° cell × calendar month × satellite set (Terra + Aqua; Terra only), fitted as a ratio of sums on paired days of the 2012–2023 overlap. The orbit-drift years 2024–2025 (Terra and Aqua read 12–25 % low) are excluded from the fit and used as test years.
- Hierarchical shrinkage cell → province → fire-regime region (NEA, NOA, Centro, Patagonia) → country, constant K = 10. The cell level is estimated where the overlap has enough MODIS detections (28 % of cell-months); the other 72 % inherit the province or region value and are labelled so in the data (factor_source).
- No intercept: where MODIS saw nothing, the reconstruction is zero (see Limits). Day and night detections are pooled: the night VIIRS/MODIS ratio is 6.1–6.7 against 2.7–3.0 by day, but separate factors gave no out-of-sample gain, because the pooled cell × month factor already absorbs the local day/night mix.
- Terra-only era (November 2000 – June 2002, before Aqua): its own factor set; shown only for areas of 100 cells or more.
- 80 % prediction interval on every reconstructed month, calibrated on out-of-sample residuals, asymmetric, binned by predicted count and satellite set, per fire-regime region and per size class of the area (1 cell · 2–9 · 10–49 · 50–199 · 200+ cells). Widened in strong ENSO months and in cloudier-than-usual months. The lower bound never exceeds the value.
3 · Fixed heat sources: a separate class, never subtracted silently
Gas flares, refineries, kilns, sawmill boilers and landfills are hot every month and have a different VIIRS/MODIS ratio from vegetation fire. Three rules put them in their own column (the app shows it as “fixed heat sources”):
- Persistent: ≥ 9 months in 12 and ≥ 50 % at night.
- Intermittent flares: ≥ 4 months in 12, ≥ 70 % at night, median VIIRS FRP < 3 MW.
- An input defect, fixed: the FIRMS MODIS “type” flag changes with the archive version (type 2 is 0.4–1.8 % of detections through 2017 and 2.3–4.9 % from 2018, the same fixed sites). Cells with ≥ 3 MODIS type-2 detections in ≥ 2 years of 2018–2025 are fixed sources in all years for both sensors; their pre-2018 “vegetation” detections have 11 % same-day VIIRS co-location (76 % elsewhere), are 97 % daytime, peak in Dec–Feb: hot bare surfaces, not fire. Filtering on the flag alone would leave those sites inside the vegetation record before 2018 and inside the factor.
Result: 1 628 cells of 0.02°, 3.0 % of all hotspots; 56 % of Neuquén's, 28 % of Santa Cruz's and 18 % of Tierra del Fuego's 2012–2025 detections are in the class (flares). 97 % of VIIRS type-2 detections fall inside it.
4 · After Suomi NPP: the series continues on NOAA-20 and NOAA-21
Suomi NPP, the reference satellite of the harmonized record, stopped delivering data on 2 November 2026 (NASA Earthdata: “Suomi NPP is no longer delivering new data”), and MODIS on Terra and Aqua is scheduled to end in 2027. The next sensor seam is not a future problem: it is this month. The record crosses it with a bridge S-NPP ↔ NOAA-20 / NOAA-21 fitted on paired days of the 2023–2025 standard-processing archives: g = 0.979 (NOAA-20) and 0.951 (NOAA-21). Checked against a different pipeline (FIRMS near-real-time production data): 0.982 for NOAA-20; provincial spread 0.93–1.16 (NOAA-20) and 0.89–0.99 (NOAA-21) where the weight is above 0.9.
In the app, the current year is a preliminary near-real-time layer: per day, the mean over the satellites operating that day, converted to S-NPP-equivalent through the bridge; every month carries the count of days with 3, 2, 1 or 0 satellites, and a month with fewer than three is marked as a lower bound. The same method ties the next sensor to VIIRS when it comes. The app's 2012 panel includes a yearly table with raw counts per satellite, 2023–2025, where the three VIIRS satellites can be compared directly.
5 · Clouds and satellite outages: measured and flagged, not corrected
- Cloud: ERA5 cloud fraction by hour of day (at each sensor's overpass) and precipitation, 0.25°, 2000–2025. Cloud does not bias the factor (differential 13:30 − 10:30 cloud anomaly vs out-of-sample residual: r = 0.02); it does lower counts for both sensors (conditional on rain, +0.10 cloud anomaly ≈ −26 % detections). So counts are not corrected: every cell-month carries cloud_anom and cloud_underdetection, the calendar marks ▲ cloudier than usual at overpass time, and the upper bound is widened in reconstructed months.
- Outages: days when one sensor reported nothing while the other did are detected and filled across sensors (including the S-NPP pre-operational days of January 2012), and the month is marked ◆ partly modelled. All 7 outages of ≥ 8 days and 13 of the 24 runs of ≥ 3 days match official NASA mission records (2 more partially). Effect below 1 % of counts per year.
6 · Validation: every number out of sample
| Test | Result |
|---|---|
| Leave-one-year-out 2012–2023, fire-season error, Terra + Aqua | country 3.2 % · province 9.8 % · single cell 30 % (one national factor: 8.9 / 15.0 / 34 %) |
| Split windows (extrapolation in time): fit 2012–17 → test 2018–23, and the reverse | country predicted / observed 0.965 and 0.992; the ratio drifts ≈ 3 % between halves (≈ 9 % before the type-flag fix) |
| Ablation ladder, province error | scalar 15.0 → country × month 13.1 → + region 12.8 → + province 12.8 → + cell 9.8 % |
| Forward test on years never used for the factors nor for the interval | 2024: predicted / observed 1.044 · 2025: 1.024 (country) |
| Backward test inside the MODIS era, 2003–2011: Terra-only vs Terra + Aqua reconstructions | country 0.94–1.04 per year |
| 80 % interval, coverage on 2024 · 2025 (years never used) | single cells 80 % / 77 % · random areas of 2–9 / 10–49 / 50–199 cells 81 / 80 / 77 % (2024) and 77 / 74 / 73 % (2025) · provinces 76 % / 73 % · country 10 and 11 of 12 months |
| Continuity across 2012 vs MODIS burned area (MCD64A1): burned ha per hotspot after / before | 0.98 harmonized · 3.12 naive splice (not independent: MCD64A1 uses MODIS active fires) |
| S-NPP ↔ NOAA bridge vs production near-real-time data (different pipeline) | 0.982 vs 0.979 (NOAA-20) |
Tested and not adopted (negative results, kept)
- ENSO as a correction of the factor: the dependence is real (r = −0.30 monthly, −0.35 annual) but small; conditioning did not improve out-of-sample skill → carried as extra interval width in strong-phase months.
- Separate day and night factors: physics confirmed (night ratio 6.1–6.7 vs day 2.7–3.0), no net skill gain (country 3.3 → 2.9 %, province 10.2 → 10.6 %, cell 29.7 → 30.8 %) → not adopted.
- An additive term for months where MODIS saw nothing: fixes single cells (Terra-only cell error 54 → 42 %) but degrades every aggregate and both out-of-time tests → not adopted; the structural zero of small pre-2012 areas stays declared.
The method and the app were reviewed in five rounds by SI systems (Claude, Anthropic) in isolated sessions; every proposal was measured before being adopted or rejected. The reviews and the audit with every number are in the repository (docs/).
7 · What it cannot know (declared limits)
- Nothing independent validates 2000–2011. The factors are fitted on 2012–2023 and every out-of-time test stays inside the VIIRS era. The reconstruction is correct to the degree that the MODIS/VIIRS relation of 2012–2023 held before 2012; the measured drift inside the VIIRS era (≈ 3 % between halves) is the best available bound on that assumption, not a test of it.
- Small areas still show the sensor change. MODIS at 1 km misses fire that VIIRS at 375 m sees: on paired days 2012–2023, the share of VIIRS-active months in which MODIS saw nothing (and the model predicts zero) is 45 % for one cell, 26 % for 25 cells, 6 % for a province, 0 % for the country. For areas under about 25 cells a zero before 2012 means “no MODIS detection”, not “no fire”. The app applies size tiers for this reason.
- The Terra-only era (November 2000 – June 2002) is shown only for areas of 100 cells or more: one satellite, province error 16.8 % (9.8 % with Terra + Aqua), and Terra alone saw nothing in 67 % of the VIIRS-active cell-months.
- A cloudy month still reads low (obscuration is flagged, not corrected). The interval of Tierra del Fuego cannot be verified (3 province-months with enough fire in 2012–2023) and the app shows it from 2012 only; the factors of Tierra del Fuego and of the City of Buenos Aires are assigned from their region, not estimated. The fixed-source class is time-invariant (vegetation fire inside its cells is removed: 0.3 % nationally, 6.4 % in Neuquén). The MODIS archive mixes file versions (6.2, 6.03, 61.03), all Collection 6 / 6.1 processing, treated as one series. Argentina only, by design.
Version history
Version 1.1.0 (16 November 2026), coverage correction. The NASA FIRMS country archive for Argentina does not include the Islas Malvinas, part of the province of Tierra del Fuego, Antártida e Islas del Atlántico Sur (Ley 23.775; Instituto Geográfico Nacional); FIRMS distributes them in a separate archive. v1.1.0 adds them (43 grid cells already assigned to the province by the IGN polygon; detections in 38) with the frozen v1.0.0 province factor (own weight 0, the rule for any cell without own paired data). No factor, band or validation figure is refitted; every value outside the islands is identical to v1.0.0; national totals change by at most 0.3 %. Tierra del Fuego now counts more cells, so its interval uses the next size class. 28 detections fall on islets the simplified IGN polygon does not cover, the same rule as on the mainland coast. FIRMS publishes no archive for the Islas Georgias del Sur and Sandwich del Sur (only volcanic anomalies would be expected there); the Antártida Argentina lies outside the grid and has no vegetation fires. Version submitted for judging: v1.0.0 (git tag v1.0.0, DOI 10.5281/zenodo.24220107).
8 · Reproduce it
Every number in the app and on this page is computed by the open scripts of the repository from NASA FIRMS data. Three levels, from no network to everything:
- From the versioned intermediates, no network, no credentials, about one minute: factors → leave-one-year-out → interval bins → harmonized product, byte for byte (sha256 of the outputs in the README). Validation from the same inputs: npm run validate.
- From the raw hotspot archives (≈ 600 MB of FIRMS CSVs + 60 MB of ERA5, published in the Zenodo record): npm run build, about 75 seconds, regenerates everything above.
- From the sources: npm run fetch needs a free FIRMS MAP_KEY and a free Copernicus CDS key.
node scripts/04_factors.mjs && node scripts/05_loyo.mjs && node scripts/05c_aoi_error.mjs \ && node scripts/05g_size_bins.mjs && node scripts/06_harmonize.mjs
Data
- NASA FIRMS active fire country archives: MODIS Collection 6 / 6.1 (2000–2025), VIIRS S-NPP (2012–2025); FIRMS Area API standard processing: S-NPP, NOAA-20, NOAA-21 (2023–2025, bridge) and near-real-time (current year).
- ERA5 (Copernicus CDS): monthly cloud fraction by hour of day and precipitation, 0.25°, 2000–2025 · NOAA CPC Oceanic Niño Index · GWIS / JRC burned-area tables (MCD64A1) · IGN Argentina province polygons.
- Products: harmonized_provinces.csv (24 provinces + Argentina, month by month, with H, H_lo / H_hi, sensor, gap_fill, cloud flags, static column, factor_source), harmonized_cells.csv (the same per 0.25° cell, 291 k rows), factors.csv, bridge.csv. Code MIT, data CC BY 4.0.
Repository (code, docs, audit) Zenodo record (data products) Back to the fire calendar
