Two LiDAR platforms, one Bayesian change-of-support framework
The paper asks whether newer, higher-density drone LiDAR actually improves model-based stand-level forest inventory relative to the airborne LiDAR that forest agencies already fly. It compares public ALS data with newly acquired ULS data inside a single Bayesian change-of-support (COS) modeling framework, and separately tests two modeling choices: adding distributional LiDAR metrics beyond mean canopy height, and adding spatial random effects.
The study area is a mixed-species forest in Mecklenburg–West Pomerania, north-eastern Germany, divided into 311 stands covering 553.5 ha, of which 249 stands (513.1 ha) are forest. Stands are small and irregular: sizes range from 0.06 ha to 17.77 ha, with a mean of 2.06 ha and a median of 1.51 ha. Species are dominated by European beech and Scots pine, with Douglas fir, pedunculate oak, black alder and Norway spruce in varying proportions.
Field data came from a systematic 183 m × 183 m grid with randomized offset and rotation, giving 165 potential plots; 14 were inaccessible, leaving n = 151 measured plots. Measurements were taken 17–23 March 2026 under leaf-off conditions with a handheld personal laser scanner on 20 m radius circular plots, and 6,722 trees with DBH ≥ 10 cm were extracted. The ALS campaign was flown 16–17 March 2023 at about 1,000 m altitude, while ULS flights ran 2–4 March 2026 at about 96 m. That roughly three-year gap matters for how the comparison should be read.
Change-of-support fixes a scale mismatch, not a sensor problem
Model-based inventory has a structural problem: field plots are small supports, while LiDAR predictors are generated on fine raster grids, and the two do not share the same spatial support. Standard regression predictions cannot simply be pushed down to arbitrarily small raster cells, because the relationship estimated at plot scale need not hold at finer resolution.
The Bayesian COS framework used here predicts at a fine spatial resolution first and then aggregates those predictions to target polygons such as stands. The paper evaluates this at 3 m and 35 m predictor resolutions. All 3 m COS variants performed considerably better than their 35 m counterparts, which the authors interpret as evidence that operating at 3 m mitigates spatial averaging errors — a benefit that matters most for small, irregularly shaped management units like the ones in this study area.
This is the paper's answer to a question readers often conflate with sensor choice. Finer prediction resolution and a different LiDAR platform are separate levers, and the study tests them separately rather than bundling them into one comparison.
ULS models cut cross-validated error by 13.8%
The headline comparison comes from 10-fold cross-validation. With full predictor sets, the ULS 3 m spatial model reached RMSPE = 68.5 m3/ha and the ULS 3 m non-spatial model 68.7 m3/ha, against 79.5 m3/ha and 81.1 m3/ha for the corresponding ALS 3 m models. The abstract reports the headline pair as 68.5 m3/ha versus 79.5 m3/ha, a 13.8% reduction in prediction error.
The advantage is not uniform across model specifications. When only mean CHM height was used as a linear predictor, ULS and ALS performed similarly: RMSPE = 76.4 m3/ha for the ULS non-spatial model versus 79.7 m3/ha for the ALS non-spatial model at 3 m resolution. The large ULS advantage appears once distributional predictors are added, which is the paper's central interaction.
Other error metrics largely agree on the ordering. MAE, CRPS and mean prediction interval width follow a broadly similar pattern to RMSPE, and the ULS 3 m full-metrics non-spatial model produced the narrowest intervals at 266.7 m3/ha, compared with roughly 480 m3/ha for the intercept-only baseline and 373.2 to 392.7 m3/ha for ALS 35 m models. Bias was low and consistently negative across all models, indicating a slight tendency toward underprediction; the smallest absolute bias was −0.3 m3/ha in the ULS 3 m full-metrics spatial model, and the largest was −7.7 m3/ha in the ALS 35 m full-metrics non-spatial model.
Distributional metrics pay off for ULS, not for ALS at 3 m
The most decision-relevant result is how the two platforms respond to richer predictors. Adding distributional LiDAR metrics reduced RMSPE by 10.1% for both the non-spatial and spatial ULS variants at 3 m resolution, and by 7.1% and 7.6% respectively at 35 m resolution, relative to the corresponding CHM-only models.
For ALS the picture is different. Adding distributional predictors improved performance only at 35 m resolution, with smaller RMSPE reductions of 3.0% and 3.2% for the non-spatial and spatial variants. At 3 m resolution, adding distributional metrics to ALS actually increased error for both model variants — for example, non-spatial RMSPE rose from 79.7 m3/ha to 81.1 m3/ha.
The authors connect this to point density. The delivered ALS point cloud had a mean density of 31.4 points/m2 in forested areas, while the ULS cloud retained a mean density of 284.1 points/m2 after being randomly subsampled to 25% of its original density. The interpretation offered is that ULS's superior point density enabled effective use of fine-scale distributional metrics, which ALS data could not support at the finer resolution. This is the authors' explanation rather than a directly isolated causal test.
Spatial random effects add little once predictors are rich
Spatial model variants generally performed similarly to, or slightly better than, their non-spatial counterparts, but the gains were small and came at greater computational cost. The largest improvement was in the ALS 3 m models, where including a spatial process reduced RMSPE by 2.0% for the full predictor set and by 2.8% for the CHM-only model. Gains from the spatial process were more pronounced for ALS than for ULS.
The Bayesian R2 comparison shows why this can be misleading. In non-spatial models, ULS had considerably higher R2 than ALS at the same resolution, and 3 m models had higher R2 than 35 m models. Those differences were not apparent in the spatial models, where Bayesian R2 was similar across data sources and resolutions. The authors note that R2 in spatial models includes variation explained by the spatial random field, so it should not be read as covariate-explained variance alone — which is why cross-validation provides the more direct basis for comparing predictive performance.
Variance partitioning also differed. At the same spatial resolution, the proportion of fine-scale residual variance attributable to the spatial component was slightly larger for ALS models. The fine-scale nugget variance was several orders of magnitude larger in the 3 m COS models than in the 35 m models, while the spatial variance component was similar but slightly larger at 35 m. Posterior median effective range was consistently larger for ALS parameter sets, exceeding the 183 m sample grid spacing for ALS models, but with wide and substantially overlapping credible intervals.
ULS also narrows stand-level predictive uncertainty
Beyond cross-validated error, the paper reports stand-level predictive uncertainty for latent stand-mean GSV. Using 3 m spatial models with full predictor sets, posterior standard deviation ranged from 14.8 to 76.6 m3/ha for the ALS model, compared with 11.7 to 48.3 m3/ha for the ULS model. Neither model showed prominent large-scale spatial patterns in posterior standard deviation, but overall uncertainty was noticeably higher for ALS.
Uncertainty also tracked stand size. Posterior SD tended to decrease with larger stand sizes, though the trend was moderate and appeared mainly as a reduction in maximum uncertainty values. Small stands showed both the highest and most variable posterior SDs, while larger stands produced lower, more stable uncertainty bounds. The authors attribute this to spatial aggregation reducing variance in stand-level means, while noting that structural conditions and stand volume also strongly influence stand-level precision.
A separate diagnostic helps explain where the two platforms diverge. Stand-wise differences in predicted GSV between ULS and ALS were positively correlated with differences in mean CHM height (r = 0.64). Stands with markedly negative CHM differences — typically corresponding to thinnings or canopy removals between sensor flights — yielded lower GSV predictions under ULS than under ALS. This supports the authors' suggestion that acquisition-related canopy height differences, not only platform characteristics, partly drive the divergence.
Model-based means sit inside the design-based benchmark
At population level, design-based estimators of GSV ranged from 304.7 to 330.6 m3/ha. The Horvitz-Thompson estimator gave 330.6 m3/ha with a standard error of 10.9 and a relative standard error of 3.3%. The regression estimator gave 304.7 m3/ha with a standard error of 6.3, a 7.8% lower mean than the SYS estimator and a 42.2% reduction in standard error relative to the SYS baseline.
None of the three post-stratification schemes — by previous stand inventory dominant leaf type, by CORINE dominant leaf type, or by previous stand inventory age class — achieved a notable reduction in sampling error, indicating that neither age-class nor canopy-type strata captured major deviations from the global population mean.
The model-based population means were 314.5 m3/ha for ALS and 309.6 m3/ha for ULS. Taking the regression estimator as the primary design-based benchmark, both model-based estimates fell well within its 95% confidence interval. The Bayesian credible intervals were narrower than the design-based confidence intervals, but the paper explicitly cautions that this difference should be viewed descriptively given the distinct statistical principles underlying the two paradigms, and should be based on population-support prediction rather than aggregation of stand-wise marginal draws.
Temporal mismatch limits the platform comparison
The study's own framing identifies the key limitation: the ALS campaign was flown in March 2023 and the ULS campaign in March 2026, while field measurements were taken in March 2026. The authors state that the ULS advantage may partly reflect closer temporal alignment with field data and finer-scale predictor information, and they call for controlled comparisons with temporally matched ALS and ULS data to distinguish platform effects from temporal mismatch.
That caveat directly qualifies the headline 13.8% figure. The comparison is between an operational ALS dataset and a purpose-flown ULS dataset acquired three years later, not between two platforms flown under matched conditions. The r = 0.64 correlation between stand-wise GSV differences and CHM height differences is consistent with this concern, since stands that lost canopy between flights showed systematically lower ULS-based GSV predictions.
Operational feasibility is the other boundary. The paper notes that ULS is promising where timely acquisition and high-resolution canopy characterization are operationally feasible, and that managers must weigh the performance gains against the additional costs of dedicated UAV surveys compared with existing regional ALS coverage. The field inventory data cannot be shared publicly due to legal and privacy restrictions, though the analytical code and R scripts are available from the corresponding author on reasonable request.
Sources used in this answer
Evaluating LiDAR Data Sources, Predictor Resolution, and Spatial Random Effects in Bayesian Change-of-Support Models for Forest Inventory
ULS-based Bayesian change-of-support models reduced cross-validated GSV prediction error by 13.8% versus ALS (68.5 vs 79.5 m3/ha), with distributional metrics helping ULS far more than ALS and spatial random effects adding only minor gains at greater computational cost.