A Blind Benchmark of Satellite Forest Carbon and Conformal Crediting with Calibrated Risk
Description
Satellite platforms promise to replace field inventories in forest carbon crediting, but their estimates are rarely tested blind against the standard-formula answer. We separate the question into two isolated grids. The right grid computes sealed field ground truth to SNI 7724, IPCC, CDM, and Verra rules. The left grid sees only plot coordinates and dates and estimates carbon from time-matched Sentinel-1/2, Landsat, PALSAR, canopy height, and GEDI lidar. On 476 plots, blind error was 21% on 1 ha plots and 7–9% on project totals outside Indonesia, near 100% on small Indonesian plots. Counting trees on a 1 m canopy height map found 3% of the stems and 25% of the carbon (62% on the 2026 map). In 1,000 simulated projects the best map found a median 95% of the field carbon; satellite embeddings and a regional bias correction raised the share within ±25% from 80% to 88.5%, and on 35 unseen Bornean plots the frozen method found 85%. We propose conformal crediting, a lower bound calibrated on other projects with field truth. It met a 10% over-crediting target for five maps within 2.9 points, was invariant to map inflation, and credited 72–84% of the true stock without a field plot; the VM0047 deduction over-credited 23–29% of projects with field plots. No satellite estimate of annual growth beat assuming no change
Files
Institutions
- Binus UniversityJakarta, Jakarta