Skip to content
Open access

Per-Species Conformal Calibration for LiDAR Forest Inventory: Coverage Imbalance and Extrapolation Failure

2026 · IEEE Geoscience and Remote Sensing Letters · Vol 23, pp. 2507103-2507103 · 0 citations · 15 references

Abstract

Split conformal prediction provides distribution-free coverage guarantees for forest attribute estimates derived from airborne LiDAR—but only marginally. Using a 1003-plot, four-species subtropical inventory and nine ensemble/regression methods evaluated under spatially stratified GroupKFold, we show that marginal coverage close to nominal (0.888–0.898 at 90%) masks per-species imbalance of up to 0.17 in coverage: one species is systematically over-covered (0.996), while another is under-covered (0.808). Mondrian calibration, applied over prespecified species groups, reduces this dispersion by 23%–79%; exact conditional coverage for every covariate value is theoretically unattainable, but per-group calibration is both achievable and operationally necessary. Under leave-one-species-out (LOSO) extrapolation, the guarantee collapses entirely: nominal 90% coverage ranges 13.4%–97.3% (mean 58.6%) across 16 attribute <inline-formula> <tex-math notation="LaTeX">$\times $ </tex-math></inline-formula> species combinations, showing that marginal conformal validity does not survive taxonomic covariate shift. To explain why species drives both accuracy and calibration, we apply bootstrap variance partitioning: species information raises <inline-formula> <tex-math notation="LaTeX">$R^{2}$ </tex-math></inline-formula> by 0.035–0.270 (95% CI excludes zero on 3/4 attributes), while region contributes nothing distinguishable from zero—confirmed independently by a linear Bayesian model with species-level shrinkage matching or exceeding a three-level hierarchical ensemble (HEL) on every attribute (<inline-formula> <tex-math notation="LaTeX">$R^{2}$ </tex-math></inline-formula> up to 0.849). Finally, we identify the mechanism behind HEL’s LOSO failure—an algebraic identity reducing to its global Gradient Boosting expert, whose prediction variance exceeds Random Forest’s under this shift despite near-identical bias—and validate a costless fix.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.