Validation Rigor Determines Apparent Predictive Skill of UAV-LiDAR Carbon Models: A Cautionary Case Study in a Heterogeneous Tropical Savanna

Machine-learning models trained on remote sensing data are widely used to map ecosystem carbon stocks, but their validation often ignores spatial autocorrelation, inflating apparent predictive skill. We tested whether this concern, established for large-scale satellite-based biomass mapping, also holds for UAV-LiDAR carbon estimation at the plot scale, comparing continuous structural metrics, segmentation-derived crown metrics, and Sentinel-2 spectral variables for predicting aboveground biomass (AGB) and soil organic carbon (SOC) across 29 inventory plots in a Cerradão-dominated Pantanal–Cerrado ecotone, Brazil. Random Forest models were evaluated under fourteen predictor scenarios and two allometric formulations. A single spatial partition yielded seemingly reasonable accuracy for AGB (best R2 = 0.27), whereas SOC showed weak apparent skill, even under a single partition (best rRMSE = 33.7%, R2 = −0.14); repeating validation across 20 independent spatial partitions, combined with leakage-free nested variable selection, showed predictive skill collapsing to near-zero or negative R2 for all 98 tested scenario–response combinations (mean R2: −0.44 to −0.003), a pattern statistically indistinguishable from a null model built on permuted response data. Variable selection stability differed from full-dataset importance rankings: raw spectral bands and canopy vertical-complexity metrics were the most consistently retained predictors across partitions for AGB and SOC, respectively, even though a derived spectral index and a lower-canopy structural metric dominated single-model importance for each response. Cross-pool coupling tests showed no meaningful gain from using AGB to predict SOC or vice versa. Empirical variogram analysis further indicated that the spatial blocking scheme adequately separated the two response variables, although residual spatial dependence could not be excluded for a subset of predictors. Validation strategy, rather than sensor combination or vegetation representation, was the primary determinant of apparent predictive skill, cautioning against overinterpreting single-partition accuracy in small, heterogeneous tropical landscapes.