LandTrendr burn detection — 2021 Bootleg Fire (Kennedy et al. 2010)
Detecting trends in forest disturbance and recovery using yearly Landsat time series: 1. LandTrendr — Temporal Segmentation Algorithms · 10.1016/j.rse.2010.07.008
Per-pixel LandTrendr trajectory segmentation of annual Landsat NBR over the 2021 Bootleg Fire, scored against the MTBS perimeter. The displayed band is peak-to-trough disturbance depth, which clears every floor; the net-change band is carried in the validation record as a failing control.
Recipe source: demos/cng-nyc-2026/hero-lock/recipe/folia.yaml — view folia.yaml
- Input data
- Landsat C2 L2 annual NBR, 84 scenes via Microsoft Planetary Computer
- Validation reference
- MTBS Bootleg-2021 perimeter (mtbs_perimeter_data.zip), rasterised to the run grid
- Period
- 2013-2026
- Region
- Bootleg Fire scar, Lake and Klamath counties, Oregon, USA
- Bounding box
- -121.55, 42.60, -121.20, 42.85
Reference: MTBS Bootleg-2021 perimeter (mtbs_perimeter_data.zip), rasterised to the run grid
| metric | observed | threshold | measured by |
|---|---|---|---|
| peak_to_trough disturbed recall producers accuracy | 77.7% | ≥ 0.7 | agent · 2026-08-19 |
| peak_to_trough disturbed precision users accuracy | 94.7% | ≥ 0.8 | agent · 2026-08-19 |
| peak_to_trough kappa | 68.4% | ≥ 0.5 | agent · 2026-08-19 |
| onset year within 1 of 2021 overall accuracy | 87.2% | ≥ 0.7 | agent · 2026-08-19 |
| magnitude disturbed recall producers accuracy | 52.9% | ≥ 0.7 | agent · 2026-08-19 |
| magnitude disturbed precision users accuracy | 96.3% | ≥ 0.8 | agent · 2026-08-19 |
| magnitude kappa | 45.0% | ≥ 0.5 | agent · 2026-08-19 |
| band choice note | peak_to_trough = fitted[argmin] - fitted[argmax] is non-positive by construction and measures disturbance DEPTH; magnitude is signed net change over the whole period and is positive on 128,895 recovering pixels of the same grid. They are different questions, not two attempts at one. The recipe styles and thresholds band 4 for that reason. | ≥ 0 | agent · 2026-08-19 |
| registration note | A half-pixel (15 m) registration offset between the 967x1000 engine grid and the 614x976 truth grid is present in every recorded row and is not corrected here. 15,413 pixels are finite in magnitude but NaN in peak_to_trough and were counted as MISSES for band 4 rather than masked out, so the recall above is the pessimistic read. | ≥ 0 | agent · 2026-08-19 |
- PaperDetecting trends in forest disturbance and recovery using yearly Landsat time series: 1. LandTrendr — Temporal Segmentation Algorithms (10.1016/j.rse.2010.07.008)
- Specdemos/cng-nyc-2026/hero-lock/recipe/folia.yaml
- DataLandsat C2 L2 annual NBR, 84 scenes via Microsoft Planetary Computer
- Resultpeak_to_trough disturbed recall producers accuracy 77.7%
- Validationvs MTBS Bootleg-2021 perimeter (mtbs_perimeter_data.zip), rasterised to the run grid
The declarative recipe is the reproduction: paper linked to spec linked to data linked to result linked to validation. Full mechanical provenance ships with the Evidence Bundle.
No forks yet. Forking (change AOI / time / params and re-run) opens in the workbench — a community fork graph is planned (hub Phase 4).
Validation history
- validated 2026-08-19 peak_to_trough disturbed recall producers accuracy 77.7%
A scheduled validation agent re-runs each reproduction and flips the badge on drift (hub Phase 1).