The project publishes two burned-area numbers and they are not alternatives. One is what the detection-driven product drew across the season. The other is what hand interpretation of 1 m imagery says actually burned, inside three event areas, with an interval. The second is not a correction of the first, because it does not cover the same ground.
1. The two area figures
| Figure | Value | Covers |
|---|
The product's mapped extent falls outside the interval, below it. That is the finding: inside the ground we labelled, the product maps less than burned. The gap is omission, and it is not small.
2. How the stratified estimate works
The estimate is stratified because the two kinds of ground behave nothing alike. Where the product claims burning it is nearly always right. Where it claims nothing, inside an event area, it is wrong about half the time. Pooling those two would hide exactly the thing worth knowing, so the strata are never pooled.
| Stratum | Area | Boxes judged | Read as burned | Rate (95%) |
|---|
3. Why NBR+? Seven-index sea and cloud false-alarm benchmark
Biak is a narrow coral-karst island where nearly every event area borders shallow reef flats, surf zones, and persistent tropical cumulus. We benchmarked seven spectral burn indices on the primary Sentinel-2 pair (2026-07-19 pre → 2026-08-23 post) across 1,482,657 open-water pixels and 1,178,295 cloud/shadow pixels, and against land near and far from any hotspot.
4. Learned spectral classifiers vs per-event local threshold
On unmapped ground inside event areas (the footprint_unmapped sample group), the local dNBR+ threshold finds only 0.27 of the boxes judged burned. To test whether that ground can be found from the same data, three simple classifiers (decision tree of depth 3, HistGradientBoosting, random forest) were trained on 24 columns of native 20 m Sentinel-2 data from before and after each fire, and tested leaving one of the three labelled events out at a time (142 boxes, 1,230 pixels).
5. Old before-images, too little comparison land & 1 m SAM outline-model test
Because Biak is often cloudy, the cloud-free look before a fire was a median 12.2 days old, searching up to 60 days back, against 2.9 days after it. In 14 of the 23 assessed events an earlier hotspot inside the event area predates the before-image, so an older scar may already be in it and the area may be understated. Separately, 8 events have fewer than 2,500 cloud-free pixels of comparison land (land within 10 km of the event and more than 3 km from any hotspot); they are reported as not assessed (not zero).
6. What these numbers do not support
- The validated figure does not scale to the season. Multiplying the season's mapped extent by the ratio between these two numbers would be an extrapolation from three event areas to twenty-five events, and nobody has tested whether it holds.
- Events the product could not assess are excluded, not zero. Where no usable nearby comparison land existed, the event carries no area. That is a missing measurement.
- The sampling unit is a 60 m box, not a pixel. A box is a cluster of pixels treated as one judgement, which is why the interval is wider than a pixel count would suggest.
- One reviewer produced the labels. Agreement between reviewers was never measured, because no second reviewer was available. That is a permanent limitation of this sample, not a gap waiting to be filled.
- No ignition cause or attribution. Satellite thermal anomalies and post-fire spectral change cannot determine how a fire started or who holds tenure over the land.
- Persistent heat sources are not wildfires. Three recurrent sites (176 detections between 2023 and 2026, all in Distrik Yendidori) stay in the data with a flag and are drawn in grey on the map, not deleted.
All figures on this site are findings of this study, computed from the data sources listed under References. Where a figure was corrected during the study, only the corrected value is shown.