From Spectral Fidelity to LAI Identifiability: A PROSAIL–LESS–GEDI Assessment of Forest Overstory LAI Retrieval

One-dimensional canopy radiative transfer models are widely used for leaf area index (LAI) retrieval, but their horizontally homogeneous representation can inadequately describe crown clumping, canopy gaps, woody elements, and shadowing in structurally complex forests. Explicit three-dimensional (3D) models offer greater structural realism, but their computational cost limits direct application at regional-to-continental scales. This study evaluates whether a residual correction distilled from the LargE-Scale remote sensing data and image simulation framework (LESS) 3.x into a PROSPECT + SAIL (PROSAIL) emulator improves forest overstory LAI identifiability when the inversion already has Global Ecosystem Dynamics Investigation (GEDI) relative-height metrics, canopy cover, and vertical-shape descriptors. The high-fidelity simulation library contains 3980 LESS scenes with explicit crowns and trunks (excluding understory and branches). Recomputing PROSAIL at the same Sentinel-2 band-center wavelengths yields a near-infrared 1D–3D reflectance residual of about −0.17, confirming a substantial structural discrepancy not attributable to spectral-response-function mismatch. On a structure-blocked split (2699 training and 1281 held-out validation scenes), with GEDI plant area index (PAI) excluded, the 1D inversion (E2) and multi-fidelity inversion (E5) use identical three-member ensembles and achieve LAI root-mean-square errors (RMSEs) of 1.564 and 1.563, respectively (R2 = 0.52). Block-bootstrap equivalence analysis yields a median ΔRMSE of −0.0009 with a 95% confidence interval of [−0.0025, −0.0001], entirely within the pre-specified ±0.05 LAI equivalence margin. The 3D residual reduces inverse spectral reconstruction RMSE from 0.038 to 0.034, but does not produce a practically meaningful LAI gain. An auxiliary joint LAI–clumping experiment is not identifiable (clumping R2 = −2.08), supporting treatment of clumping as a diagnostic rather than a retrieved product. Independent validation against Ground-Based Observations for Validation (GBOV) RM-7 Warren overstory LAI (187 observations, 16 sites) gives R2 = 0.30, r = 0.63, RMSE = 1.36, and bias = +0.16. These results show that improved 3D forward-model fidelity does not necessarily improve parameter identifiability when independent structural constraints are already present.