Optimizing GEDI Simulator Configuration for European Temperate Forests

• In French-forest simulations, weighting laser returns by reflected intensity substantially reduces canopy-height error and removes the systematic seasonal bias that made winter (leaf-off) simulations less reliable than summer ones.Source: Plain Language Summary · p. 2
• Leaf-off and low-sensitivity shots can be used alongside standard data to expand calibration and inference datasets.The improved settings remove the seasonal bias that made winter simulations less reliable than summer simulations.Source: Plain Language Summary · p. 2
• About 9,500 synthetic GEDI measurements generated from French national airborne LiDAR data were compared with real GEDI observations.Source: Plain Language Summary · p. 2

• Accurate aboveground biomass density estimates matter for carbon accounting, where small errors can propagate into large regional or global uncertainties.Source: §1 · p. 2
• GEDI footprints remain sparse in space and time, constraining local calibration and validation of biomass models.Footprints are spaced about 60 m along-track and 600 m across-track, with limited revisits at inventory plots.Source: §1 · p. 3
• The Hancock et al. (2019) simulator is widely used to generate GEDI-like waveforms from LiDAR point clouds for biomass-model calibration.Source: §1 · p. 3
• Only a few studies directly compared simulated waveforms with GEDI observations, and prior simulator validation covered a limited set of biomes.A comprehensive assessment is needed because differences may propagate into biomass and carbon products.Source: §1 · p. 3
• The study independently evaluates the simulator across French forest types and phenological conditions using national airborne LiDAR data, including leaf-off observations.The authors describe it as the first assessment over European temperate forests.Source: §1 · pp. 3–4
• The study lowers the sensitivity threshold to 0.9 to test whether lower-sensitivity shots, often coverage beams, can be simulated as faithfully as high-sensitivity shots.The authors also test whether leaf-off data can complement leaf-on acquisitions for biomass calibration and inference.Source: §1 · p. 4

2.1 Study Area

• Metropolitan France provides the study area, spanning temperate broadleaf forests in the north, Mediterranean woodlands in the south, and varied topographic and climatic gradients.Source: §2.1 · p. 5

2.2 Datasets

• GEDI L2A profiles from algorithms a1–a6, including the designated default, provide observed RH metrics; L1B waveforms are collocated with simulations for comparison.Tracks were retained when at least 60% of footprints were high-quality shots; the analysis retained shots above sensitivity 0.9 rather than applying stricter 0.96–0.98 thresholds.Source: §2.2 · pp. 5–6
• The national ALS program supplies high-density point clouds, with a minimum target of 10 points per m2, as simulator input and a high-resolution reference.Source: §2.2 · p. 6
• GEDI shots were restricted to the acquisition dates of their corresponding ALS unit, and seasonal groups used the per-shot leaf-off flag.Forest type was classified as broadleaf, coniferous, or mixed to maintain a balanced dataset of about 3,000 shots per type.Source: §2.2 · p. 6

2.3 Indicators Derived From ALS Point Cloud

• ALS indicators include ground elevation from the center of gravity of ground-classified points and canopy cover from canopy versus total returns, adjusted for reflectance differences.The canopy-cover correction uses a vegetation-to-ground reflectance ratio of 0.57/0.40.Source: §2.3 · p. 6

• The effective ground slope is estimated from the width of the simulated, noise-free ground return.Source: §2.4 · p. 7
• The ground-return width calculation uses the ground-peak spread, system pulse width, and footprint width.The prose defines these quantities as σg, σp, and σf, respectively.Source: §2.4 · p. 7
• The resulting metric represents combined topographic slope and surface roughness within the footprint and is expressed in degrees.Source: §2.4 · p. 7

2.5.1 Theoretical Basis for Waveform Simulation

• The count method weights returns equally, frac assigns each return 1/nhits of pulse energy, and int weights returns by recorded ALS intensity; count is the baseline.Density normalization is enabled for all simulations and adjusts return contributions using local ALS pulse density.Source: §2.5.1 · pp. 7–8
• ALS returns are spatially weighted, assigned return-specific weights, binned vertically, and convolved with the instrument pulse to produce a noise-free GEDI-like waveform.This study uses the empirically measured GEDI pulse rather than the default Gaussian pulse.Source: §2.5.1 · p. 8

2.5.2 Simulation Tools and Processing Workflow

• The collocateWaves tool searches horizontal and vertical offsets to maximize track-level waveform correlation, using a 20 m × 20 m horizontal search window.Collocation retains tracks meeting GEDI quality and degradation-flag criteria for at least 60% of shots in the LiDAR area.Source: §2.5.2 · p. 9
• The gediMetric tool extracts GEDI-like RH percentiles, canopy cover, and foliage height diversity, and supports sensitivity analysis through synthetic noise and denoising thresholds.A threshold based on waveform-bin mean and standard deviation both removes low-amplitude bins and defines the vertical signal extent.Source: §2.5.2 · p. 9

2.6 Workflow

• Simulated and algorithm-derived RH outputs are referenced to common ALS ground to separate ground-detection errors from waveform-modeling differences and identify the best simulation–algorithm pairing.The selected configuration is compared with the baseline across forest types, phenological conditions, and canopy cover.Source: §2.6 · p. 10

2.6.1 Datasets Summary

• Three return-weighting datasets—count, int, and frac—are compared with the default GEDI L2A configuration and each of the six algorithms; varScale remains at 3 after trials showed no systematic improvement.The explored varScale changes sought to better match the default GEDI L2A signal extent.Source: §2.6.1 · pp. 10–11
• Filtering retains temporally overlapping, fully covered shots with canopy cover between 0 and 1 and ALS-referenced simulated RH100 from 5 m to 60 m.The RH100 range excludes potential land-cover misclassifications and heights considered physically implausible for France.Source: §2.6.1 · p. 11
• Table 2 summarizes final shot-level pairs by forest and beam type, including pair counts, mean canopy cover and slope, and the proportions assigned to default algorithms.Dataset size may vary with the GEDI algorithm and its quality flags.Source: §2.6.1 · p. 11, Table 2 · p. 11

2.7.1 Signal Overall Comparison

• Track-level collocation selects 3D displacements that maximize mean Pearson correlation across valid footprints, with before-and-after correlations used to quantify alignment improvement.This optimization primarily assesses vertical-structure agreement rather than absolute energy values.Source: §2.7.1 · p. 11

2.7.2 GEDI L2A Algorithm Evaluation

• Each L2A algorithm is assessed for ground-elevation accuracy against ALS and for RH-profile agreement when combined with each return-weighting strategy.Ground-detection offsets can propagate into all RH metrics, while algorithm-specific denoising and thresholds affect the cumulative RH profile.Source: §2.7.2 · p. 13

2.7.3 RH Metrics Comparison

• Observed GEDI RH metrics are vertically translated to ALS ground before comparison, isolating canopy-energy-distribution differences from ground-detection artifacts.The full-profile accuracy summary is mean absolute bias averaged across RH1–RH99; RH0 and RH100 are excluded as noise- and threshold-sensitive.Source: §2.7.3 · p. 13

2.7.4 Stratification of the Analysis

• The MSD decomposition distinguishes squared bias, differences in standard deviations, and correlation loss; stratified analyses report bias directly in meters across forest type, phenology, and sensitivity groups.Direct bias is used because squared bias visually under-represents residual errors below 1 m.Source: §2.7.3 · p. 14, §2.7.4 · p. 14

3.1 Collocation Results

• Mean waveform correlation rose from 63.7% to 89.5% after 3D collocation, an average improvement of 25.8 percentage points.Vertical adjustments were centered near -48 m across tracks, reflecting the difference between the GEDI WGS84 ellipsoid and ALS IGN69 orthometric datum.Source: §3.1 · p. 15

3.2 Ground Detection Accuracy Across GEDI Algorithms

• a3 has a mean ground bias of -0.03 m (RMSE 4.57 m), compared with approximately -0.88 m (RMSE 4.15 m) for the default configuration.Across sensitivity groups, a3 biases are +0.59 m for low-sensitivity and -0.61 m for high-sensitivity shots; the default biases are -0.19 m and -1.11 m, respectively.Source: §3.2 · p. 15, Table 3 · p. 16

3.3 Comparison of Simulated RH Profiles Across GEDI Algorithms

• SB drops from 0.94 (count–default) to 0.09 (int–a3), while MSD falls from 10.22 to 8.34; int–a3 is selected for subsequent analyses.The int–a3 configuration also has the lowest RH bias across weighting strategies for all shots with sensitivity ≥ 0.90.Source: §3.3 · pp. 16–17, Table 4 · p. 16

3.4 Drivers of Differences Between Simulated and Real GEDI RH

• The analysis examines residual bias under the optimized configuration across forest types and phenological states, with high- and low-sensitivity shots separated.The phenological comparison focuses on broadleaf and mixed forests.Source: §3.4 · p. 18, §3.4.1 · p. 18, §3.4.2 · p. 20

3.4.1 Forest Type

• Under int–a3, mean absolute bias is 0.31 m for broadleaf, 0.24 m for coniferous, and 0.25 m for mixed forests.Comparison with count–a3 indicates that most of this improvement stems from intensity weighting; biases are larger near RH5 and RH98, where sensitivity groups have opposite signs.Source: §3.4.1 · p. 18

3.4.2 Phenological Effects

• For high-sensitivity shots, mean absolute bias shifts from 0.69 m leaf-on and 1.28 m leaf-off (count–default) to 0.44 m and 0.40 m (int–a3).For low-sensitivity shots, int–a3 yields 0.37 m leaf-on and 0.34 m leaf-off; a strong seasonal divergence remains at RH98 for this group.Source: §3.4.2 · p. 20, Table 5 · p. 22

4.1.1 Ground Detection: An Overlooked Source of Bias

• a3 has an overall ground residual of -0.03 m versus -0.88 m for the default, but the default performs better within the low-sensitivity class.For low-sensitivity shots, default bias and RMSE are -0.19 m and 2.74 m, versus +0.59 m and 4.69 m for a3; its more restrictive classification yields a less challenging subset.Source: §4.1.1 · p. 23

4.1.2 Return-Weighting Strategies

• Intensity weighting better represents the backscatter profile: it gives the ground return and dominant canopy layers weights reflecting their radiometric properties.Unlike count weighting, which treats returns equally, intensity weighting better matches the observed GEDI ground and canopy signals and reduces residual biases.Source: §4.1.2 · pp. 23–24
• 0.40 m (leaf-off) and 0.44 m (leaf-on) mean absolute bias under int–a3, versus 1.28 m and 0.69 m under count–default, nearly remove the seasonal gap.The comparison is for high-sensitivity shots.Source: §4.1.2 · p. 24
• Leaf-off, low-sensitivity RH98 remains an exception, with all weighting methods showing approximately 2 m positive bias; the authors advise caution or exclusion downstream.Weak returns from bare upper branches can fall below GEDI’s 3σ signal threshold.Source: §4.1.2 · pp. 24–25

4.1.3 Low and High Sensitivity Shots

• Under int–a3, low- and high-sensitivity shots have comparable full-profile mean SB (0.24 vs 0.22), though low-sensitivity shots show higher MSD (9.86 vs 6.9 m2) from greater dispersion.RH20–RH90 is the most robust low-sensitivity percentile range; RH5 and RH98 remain problematic, particularly leaf-off. Coverage-beam performance across topographic gradients was not explicitly investigated.Source: §4.1.3 · p. 25

4.1.4 Generalisability and Recommendations for Simulator Users

• The authors recommend int–a3 as a practical default for European temperate forests using modern high-density ALS, while advising independent validation for other biomes or older sensors.The recommendation is supported by comparable forest classes and increasingly common modern ALS sensor configurations; boreal forests and dense Mediterranean woodlands may require independent validation.Source: §4.1.4 · p. 26

5 Conclusion

• The study compares approximately 9,500 collocated simulated and real GEDI waveform pairs across French forests and finds that optimal return weighting varies with ALS sensor characteristics.Source: §5 · p. 26
• 0.44 m in leaf-on and 0.40 m in leaf-off mean absolute bias under intensity weighting with a3, down from 0.69 m and 1.28 m under the baseline, respectively.The a3 ground-detection bias is -0.03 m versus -0.88 m for the default.Source: §5 · pp. 26–27
• Leaf-off acquisitions and low-sensitivity shots can expand calibration and inference datasets, except for RH98 under leaf-off conditions.The conclusion also notes that lower signal-to-noise affects low-sensitivity precision rather than accuracy.Source: §5 · p. 27

5.1 Perspectives

• The next step is to test whether leaf-off ALS calibration and leaf-off or low-sensitivity GEDI inference improve biomass estimates and spatial coverage relative to standard approaches.Source: §5.1 · p. 27

Open Research Section

• GEDI L1B and L2A V2 products were downloaded from NASA Earthdata Search.Source: Open Research Section · p. 27
• French national LiDAR HD point clouds are distributed by IGN via the Geoplateforme, and Corine Land Cover 2018 is available from Copernicus.Source: Open Research Section · p. 27
• The GEDI simulator code is available online, and the study’s analysis scripts will be made available upon manuscript acceptance.Source: Open Research Section · p. 27

Appendix A Real and Simulated Waveforms

• Figure A1 presents simulated and real GEDI waveforms for a broadleaf canopy in leaf-on conditions.Source: Figure A1 · p. 31
• Figure A2 presents simulated and real GEDI waveforms for a broadleaf canopy in leaf-off conditions.Source: Figure A2 · p. 32
• Figure A3 shows a coniferous canopy, with columns corresponding to different weighting methods and overlaid default and simulated RH profiles derived from ALS ground.Source: Appendix A · p. 31, Figure A3 · p. 33

Appendix B Low and High Sensitivity Shots Analysis

• Tables B1–B2 report mean MSD decomposition (SB, SDSD, LCS, MSD) separately for low- and high-sensitivity shots.The table notes specify sensitivity ranges of 0.90–0.96 and above 0.96, respectively.Source: Table B1 · p. 34, Appendix B · pp. 34–35, Table B2 · p. 35
• For sensitivity 0.90–0.96, sample sizes range from 98 (a5) to 4,999 (a1), varying with algorithm-specific quality and sensitivity filters.For sensitivity above 0.96, sample sizes range from 4,042 (a1 and a4) to 9,517 (a5).Source: Appendix B · pp. 34–35
• Bias profiles compare low- and high-sensitivity shots across forest types for count–a3 and frac–a3 configurations, using ALS-derived ground elevation.Wide light bars represent low-sensitivity shots; narrow dark bars and red dots summarize high-sensitivity and overall bias, respectively.Source: Figure C1 · p. 36