Forest area

Calibrated metrics by year. n ≈ 4 949 holdout points/year. Probabilistic ranking (PR-AUC) at the two production thresholds — 0.40 broad / 0.60 strict — plus continuous Soft Brier and RMSE.

Tree height

Regression metrics vs GEDI L2A RH95 holdout — quality_flag = 1, power beams only, sensitivity ≥ 0.95. R² compared against Potapov 2021 global CHM (R² ≈ 0.62).

Carbon stock (CO₂e)

CO₂e raster vs GEDI L4A AGBD-derived CO₂e holdout. R² is circular by design — we validate against the same GEDI L4A product used in training. 2026 footprints are scarce (outside the GEDI Phase 2 window), so R² for that year degrades to sample-size noise.

Tree change

Reported under the strict methodology only (crown cover ≥ 0.60, aligned with FAO FRA 2020 and the IPCC GPG 2003 closed-canopy class). The broad threshold (0.40) is published as a state product but is not used for change detection — at that threshold the canopy density distribution sits near the cutoff and any reasonable hysteresis absorbs most legitimate signal.

Production tree_change emits area accounting only — stable / loss / gain hectares per pair, plus loss%/year and gain%/year as fractions of the valid pixel pool. Pixel-level Hansen accuracy metrics (FoM, PA, UA, kappa) are intentionally not produced here: Hansen detects stand-replacement disturbance whereas this product detects threshold-crossing canopy-cover change — physically different events. A Hansen cross-check is available as an opt-in run for ad-hoc analysis.