This page describes how every figure on the site was produced, in the order the work was done..
Research flowchart
-
1. Data collectionNASA FIRMS hotspots: VIIRS 375 m and MODIS 1 km, 2023–2026CHIRPS rainfall, MOD16A2 evapotranspiration, ERA5-Land weatherLand cover: ESA WorldCover, Dynamic World, JRC TMFSentinel-2 L2A with Cloud Score+ cloud scoresS2DR4 1 m imagery (Gamma Earth)Airport weather reports (METAR WABB) and Himawari-9
-
2. Natural factorsMonthly water balance and the 46-year rainfall rankFire Weather Index from ERA5-Land, checked against cloud-free daysSix-class land cover map from three products
-
3. Hotspot processingQuality flags, no rows deleted; times in UTC and WITEvents by DBSCAN: 750 m, 30 hours, at least 2 hotspotsAirport smoke reports matched to events by time and direction
-
4. Burned-area mappingIndex choice: false alarms of 7 indices over sea and cloudNearest cloud-free image per pixel: up to 60 days before, 30 days afterdNBR+ above a local threshold per event gives the burned polygons
-
5. Validation142 boxes of 60 m judged on 1 m imageryStratified area estimate (Olofsson et al. 2014) with 95% intervalsExtra checks: classifiers, SAM outline model, age of the before image
-
6. ResultsInteractive map, charts, daily briefs, and the manuscript
1. Hotspot data
Hotspots come from NASA FIRMS: VIIRS on Suomi NPP, NOAA-20 and NOAA-21 at 375 m, and MODIS on Terra and Aqua at 1 km. The store holds 1,166 detections from 1 September 2023 to 23 September 2026 (WIT days). Times are stored in UTC, and a WIT (UTC+9) column is derived from them; every daily count uses the WIT day, because a night overpass would otherwise fall on the wrong date.
The confidence column is kept as text. VIIRS gives l, n or h and MODIS gives 0 to 100 in the same column, and turning it into a number would drop most VIIRS rows. Fire radiative power (FRP) is kept on every row. Quality control adds flag columns and never deletes rows, so offshore pixels and the three recurrent heat sources stay in the data with a flag.
Polar-orbiting satellites pass over Biak twice a day, around 00:31–01:58 WIT and 09:00–15:48 WIT. From 15:00 to 00:31 WIT none of them sees the island. Himawari-9 covers those hours with 2 km pixels, but its daytime flags appear as often on days without fires, so they cannot time the fires. Airport weather reports from Frans Kaisiepo (WABB) give an independent record of smoke from the ground.
2. Rainfall, water balance and fire weather
Monthly rainfall is the CHIRPS mean over land, with permanent water masked out. The full series from 1981 to September 2026 was used to rank the 2026 dry season against every earlier year. Evapotranspiration comes from MOD16A2, which starts in January 2021; rainfall minus evapotranspiration is the monthly water balance.
The Canadian Fire Weather Index system was computed from ERA5-Land at noon WIT for 30 land cells. Each index was scored by how well it ranks days with hotspots, one year at a time, against a day-of-year baseline. Because dry days are also cloud-free days, the test was repeated on days when the airport reported no broken or overcast cloud.
3. Land cover map
A six-class land cover map was built by reclassifying ESA WorldCover 2021, Dynamic World and JRC Tropical Moist Forest, with TMF splitting the single WorldCover tree class into closed forest and regrowth. No training data were used, so the map has no accuracy figure. It is used to compare each class's share of land with its share of hotspots and of mapped burned area.
4. Events
Hotspots were grouped into events with DBSCAN, using a 750 m radius, a 30-hour window and at least 2 hotspots. The 30-hour window follows the satellite timing: VIIRS passes between 12:34 and 14:21 WIT and Aqua as late as 15:48, so a plot that burns on two afternoons in a row is seen again 21 to 27 hours later. The grouping gives 488 events, 31 of them with at least 5 hotspots. An event is a group of hotspots, not a confirmed fire.
5. Choosing the burned-area index
Without a local burn-scar reference, seven indices (NBR, NBRSWIR, NDSWIR, MIRBI, BAIS2, NBR+ and ATBI) were compared by how often they flag places where burning is impossible: open sea, cloud and cloud shadow, and land far from any hotspot. Only NBR+ kept false alarms close to zero over both sea and cloud. ATBI flagged 43.5% of water pixels.
Index formulas
Sentinel-2 bands used: B2 blue, B3 green, B4 red, B6 and B7 red edge, B8A near infrared, B11 and B12 short-wave infrared. Reflectance is divided by 10,000 before any formula. Burned ground reflects less near infrared and more short-wave infrared than green vegetation, and the formulas below use that difference.
- NBR
(B12 − B8A) / (B12 + B8A)- Compares short-wave infrared with near infrared. Burning removes green leaves (less near infrared) and leaves dry, dark ground (more short-wave infrared), so the value rises. It is written here so that a higher value means burned; many papers use the opposite sign.
- NBR+
(B12 − B8A − B3 − B2) / (B12 + B8A + B3 + B2)- NBR with the blue and green bands subtracted. Water and cloud are bright in blue and green, so their value is pushed down and they are not mistaken for burned ground. This is the index behind the burned-area map.
- dNBR+
NBR+ after the fire − NBR+ before the fire- The change between the image before and the image after the fire. A large positive change means the ground became burned-looking. A pixel counts as burned when its dNBR+ is higher than 99% of the values on comparison land nearby.
- NDVI
(NIR − Red) / (NIR + Red), MODIS b02, b01- Greenness. Healthy plants reflect much near infrared and little red light, so a higher value means greener vegetation.
- NDMI
(NIR − SWIR1) / (NIR + SWIR1), MODIS b02, b06- Water in leaves. Wet leaves absorb short-wave infrared, so drier vegetation gives a lower value.
- NDSWIR, NBRSWIR, MIRBI, BAIS2, ATBI
NDSWIR = (B11 − B8A) / (B11 + B8A); NBRSWIR = (B12 − B11 − 0.02) / (B12 + B11 + 0.1); MIRBI = 10·B12 − 9.8·B11 + 2; BAIS2 = (1 − √(B6·B7·B8A / B4)) · ((B12 − B8A) / √(B12 + B8A) + 1); ATBI = NBR · (B12 / B8A)- Compared only, to choose the index, and not used for the map. Three of them flagged almost the whole sea as burned, and ATBI flagged 43.5% of water pixels.
6. Burned-area mapping
For each event with at least 5 hotspots, the event area is the union of 375 m circles around its hotspots. For every pixel, the pipeline takes the latest cloud-free Sentinel-2 image up to 60 days before the event and the earliest up to 30 days after, where cloud-free means Cloud Score+ cs_cdf of 0.6 or more. These are search windows; no average image is made. All index work is at 20 m.
A pixel is mapped as burned when its dNBR+ is higher than 99% of dNBR+ values on nearby land (within 10 km) that lies more than 3 km from any hotspot. If that land has fewer than 2,500 cloud-free pixels (100 ha), no threshold is set and the event is reported as not assessed, never as zero.
7. Validation
For the three event areas with 1 m S2DR4 imagery (E0301, E0313, E0368), 142 boxes of 60 m were drawn in three groups: inside the mapped polygons, on unmapped ground inside the event area, and at random. One reviewer judged each box as burned, unburned or unsure; unsure boxes are left out. Burned area was then estimated per group following Olofsson et al. (2014), with Wilson 95% intervals. The groups are never merged.
The same labels were also used to test three classifiers (decision tree, gradient boosting, random forest) against the threshold, leaving one event out at a time, and an automatic outline model (SAM) was tried and did not beat the simple baseline. S2DR4 stores its bands in the order B2, B3, B4, B8, B5, B6, B7, B11, B12, B8A without labels; reading them as a normal Sentinel-2 stack puts the wrong bands into every index.
8. What is published
Hotspot locations and burned polygons are published, because they are observations. Nothing on this site says who caused a fire: a 375 m pixel cannot point to a person, a landowner or a village. Burned polygons carry only their event number, with no link to land ownership. Administrative boundaries (from BIG) follow SNI 6502.2:2010, which states that such lines are not an official boundary reference.