Effect of GNSS positional errors on predictive models performance in LiDAR-based forest inventories

Airborne laser scanning (ALS, or airborne LiDAR) has been widely incorporated into forest inventory applications, as it can provide quantitative information on many forest attributes. Accurate alignment of field measurement plots with ALS data is essential for area-based forest inventory modeling. However, errors in horizontal position estimates from global navigation satellite system (GNSS) receivers can mislead field crews when locating plot centers. This misalignment may introduce co-registration errors between field measurement and LiDAR datasets, potentially degrading model quality. In this study, we evaluated the sensitivity of predictive models for basal area, tree volume, and aboveground biomass to plausible GNSS positional uncertainty in the Talladega National Forest, USA. Physical data from 254 field measurement plots were combined with 50 LiDAR-derived metrics concerning the vegetation contained in the geographic area of those plots to build regression models using an adaptive LASSO approach. To assess positional uncertainty, we applied Monte Carlo simulation to randomly shift plot centers up to 5 m and recalculating LiDAR metrics across 30 iterations. In total, 30 regression models were developed for each forest inventory attribute using LiDAR data extracted from the shifted plot locations (imputed datasets). Results from statistical tests revealed significant differences ( p < 0.05) between original and imputed models, with R 2 adj values declining for basal area, while some tree volume and aboveground biomass models improved slightly (up to +0.02). The limited magnitude of improvement was attributed to the relatively small horizontal displacement between the original and shifted plot locations (3.33 m on average), the heterogeneous biophysical structure of the national forest and the relatively large plot size (i.e., 404.7 m 2 ). Five LiDAR metrics, including p 2th , p zabove2 , i sd , i kurt , and i skew , consistently appeared in both original and imputed models. These metrics reflected mid-canopy structure, canopy density, and indirectly, tree height. Overall, our findings demonstrate that inherent GNSS positional errors can meaningfully affect predictive model accuracy and estimates, even when average positional variability is modest (e.g., 1.72 m in our case). We therefore recommend that future ALS-based forest inventory modeling efforts explicitly report GNSS positional accuracy alongside standard performance indicators (e.g., R 2 adj , RMSE, bias) to enhance model transparency and reliability.