Fig. 3: Point cloud processing workflow.
Evaluating Volumetric Measurement Methods
Once the data were classified into ground and vegetation, we investigated several methods to compare volume with above-ground biomass. We assessed the data based on a per site (n = 6) and as combined sites per year, and by combining years. Our interest in assessing on a per-site basis was to consider possible effects of changing photography conditions (one site photographed per day), or influences of the site location such as general vegetation type. Similar logic was applied to the per year versus both years combined datasets.
We examined how we accounted for voxel volume by using vegetation-classified points as the volumetric measure (vegetation voxel volume), and by using the total plot voxel volume with no classification of vegetation and ground. We performed the latter to test whether litter or other obstructions had a detrimental effect on the ability to detect the ground surface. Ultimately, in this model, litter was considered as part of the biomass volume.
We also computed a vegetation voxel volume above the surface, by fitting a surface to the ground and computing the vegetation volume above the surface. Additionally, we examined the relationships between volumes and vegetation type (e.g. perennial and annual grasses, forbs). Linear models and logarithmic models were both investigated for the relationships between voxel volume and biomass.
Finally, we also assessed the scale at which volumetric data is best reconstructed for sagebrush-steppe vegetation (e.g., fine annual grasses) and to inform the possibility for using volumes reconstructed from UAS imagery. This included voxelizing the point clouds at several resolutions ranging from 2 mm to 100 mm, comparing volumes of all reconstructed points and volumes of vegetation-classified points only.

Results

  1. Volume and Biomass Allometry Modeling and Sensitivity Analysis
Table 1 depicts the models we tested that had a significant (> 0.30 R2) relationship between voxel volume and field biomass. We found the best correlation (R2 = 0.48) with a logarithmic regression between 5 mm voxel volumes of vegetation-classified points with above ground biomass for all sites across years (Table 1). We generally found that logarithmic models performed better than linear models when all plots were grouped together across years.
A linear regression of plots (n = 36) from both years from South Camp Kettle (South site) attained an R2 of 0.62 between classified vegetation points at 5 mm voxel sizes with the biomass of only annual and perennial grasses (i.e. no litter biomass was included). However, generally there was no pattern of linear or logarithmic modeling performing better for any of the comparisons when modeling by site or by year.
Generally, our vegetation voxel volume calculated from classified vegetation points (from Fig. 3, workflow) performed better than using a fitted surface to identify vegetation points and calculate a volume. Using a total plot voxel volume along with only annual and perennial grasses had a relatively high R2. No relationship with forb biomass was found due to their relative rarity in the study area.