Fig. 5. Calculated vegetation and litter voxel volumes using
reference ground surface and reconstructed ground surface.
Discussion
The outcome of this research illuminates some considerations specific to
the vegetation of the study area, chiefly the gracile structure of
medusahead and the layer of litter that it deposits. Broadly, we found
that the largest cause for uncertainty in our estimations resulted from
the fine fuel litter layer obscuring the ground surface. However, with
some additional processing, we were able to attain some informative
results.
Structure-from-Motion Sample Frame and Reconstruction
Reconstructing point clouds from the field photography is a feasible
sampling strategy for studies of this scale. However an effort must be
taken to automate as much as possible, since it can be labor- and
processing-intensive. The addition of coded targets at known, local
coordinates greatly reduces the need for manual steps in the process. We
observed that reconstruction success of the point clouds was mostly
dependent on the photo quality (e.g. not blurry), amount of overlap, and
that the targets were not obstructed by even a small amount of
vegetation. Environmental conditions also played a role in the success
of the point cloud reconstruction; occasionally, lighting conditions
created too many shadows or too much variability during the
photographing process. Due to the relatively delicate structure of the
site vegetation, small gusts of wind can be troublesome and cause a plot
to fail reconstruction quality thresholds, or fail to process
altogether.
Because the per-year and per-site regressions did not have significant
relationships (excepting 2020-2021 South Camp Kettle South, n = 36,
R2 = 0.62, 5 mm voxel size), we hypothesize that the
conditions for all fields were similar and that the reconstructed point
clouds were of similar quality and captured the variability of
vegetation types equally across the study environment. We observed no
significant relationship between vegetation types and volume
calculations. The fidelity of the reconstructed point clouds to the
vegetation structure was mostly influenced by whether the ground surface
was visible, and whether there was a dense layer of litter or stand of
vegetation.
Ground and Vegetation Classification and Separability in Point
Clouds
Vegetation in the Three Fingers Allotment can often contain a dense
layer of dead medusahead from previous growing seasons, or a thick cover
of living medusahead, or both. This means that passive remote sensing
techniques such as Structure-from-Motion may have difficulty measuring
both the internal volume and the ground surface. Such obstructions
likely decreased the correlation between calculated volume and measured
biomass. Iterative classification of the ground surface points increased
the observed relationship between volume and biomass for all
combinations of vegetation types and biomass. In addition to occluding
the ground surface, the layer of litter also confounded the ability to
accurately characterize the ground surface.
However, our comparison between our reconstructed ground surface with
the ‘true’ reference ground surface elucidates two points. First, the
decrease in predictive ability between the reference surface and our
reconstruction is only 2% (from R2 of 0.44 to
R2 of 0.42), which assists in constraining
uncertainties in volumetric calculations using our ground and vegetation
classification or similar process. Second, the range of litter biomass
in this regression ranged from 20% to 100% of the total vegetation
biomass with evenly distributed errors. This indicates that the
disagreement between calculated volumes is independent from the
proportion of litter in the total biomass. Thus, it may be possible to
adequately quantify uncertainty in volumetric measurements at
pasture-scales. It is possible that the observed relationship is related
to the compaction of the litter layer over time. However, a larger and
more varied sample size at the smaller plot scale would help further
understand this relationship.
Additionally, we explored using a convex hull volumetric calculation to
compare to the voxel volume calculations. While the convex hull method
may better handle the under-sampled volumes internal to a mat of litter
or dense vegetation, we found that although the results were similar to
the voxel models, the convex hull method generally performed worse
across the different classifications and comparison models.
Voxel Volume Analysis
Because the total volumetric calculations were statistically significant
for each voxel size, we ascertain that the volumetric calculations are
highly dependent on the resolution of the final point cloud. Volumes
from decreasing voxel sizes were increasingly similar to the measured
biomass, up to a point: the distribution of voxel volumes at 5 mm more
closely matched the biomass distribution than the 2 mm voxel sizes. This
may indicate that the scale of obscured volume is exacerbated as
decreasing voxel sizes are employed for the vegetation communities; the
smaller voxel sizes may be more reflective of the surface of the
vegetation. More simply, it may be that the density of such plant
communities is beyond the limits of the resolution of the imagery
(Wallace et al. 2017).
Considerations for Future UAS-Derived and Non-Destructive
Biomass Data
Our study helps inform UAS-derived spatial and structural data in two
ways. Primarily, the unique characteristics of the vegetation and
ecological communities in this study area (perennial and annual
grass-dominated, of varying densities and heights, and with significant
layers of litter cover) strongly favor higher-resolution UAS data. While
it is not uncommon for UAS sensors to routinely output models at 5 cm
pixel sizes, our analysis suggests that additional mission planning and
sensor considerations to achieve closer to 1 cm pixel sizes may be
important to enable volumetric estimates of biomass in such
environments, as found by Cunliffe et al. (2016). Additionally, the
complications of our analysis given the common occlusion of the ground
surface indicate that volumetric estimations on such a scale (e.g.
centimeter to tens of centimeter litter layers, typically shorter than 1
m vegetation heights) are sensitive to the spatial resolution of the
remote sensing product and that using an elevation model (e.g. DSM) may
be insufficient to reconstruct a ground surface.
Given the uncertainty of identifying ground surface, whether from
classified points or a reconstructed surface, our proposed method would
benefit from refinement before being considered a replacement for
destructive harvesting. Such refinements might include systematic
comparisons with other commonly-used biomass sampling techniques for
grasslands such as rising plate meter devices. Separating ground from
vegetative biomass, or litter from ground and standing biomass may
benefit greatly from using multispectral imagery. Active remote sensing
techniques such as lidar, whether in conjunction with imaging or alone,
may also more reliably return ground points through litter layers and
dense vegetation, as well as providing other data such as intensity that
may enable greater discernment between ground, litter, and other
vegetation.
Other avenues to explore with our proposed method may explore other
calculations and measurements from the dense and high-resolution point
clouds, such as vegetation height or other allometry (Cunliffe et al.
2022, Schulze-Brüninghoff et al. 2021). This may include relationships
between biomass and structure, or as a more direct replacement of other
field-measured data such as canopy cover. With other passive sensor
types such as multispectral or hyperspectral imaging, or active sensors
such as lidar, other metrics such as moisture level or classifications
may be possible. Fusion of such data types has shown more success
together in grasslands in predicting biomass as well as other properties
such as moisture content (Schulze-Brüninghoff et al. 2021).
Conclusion
In this paper we explored using Structure-from-Motion at ultra-fine
scales in an environment dominated by exotic annual grasses and
associated litter. Understanding the amount of biomass and litter in
these rangeland environments is important for managing forage and fuels,
and understanding treatments and changes. UAS offer the possibility to
scale up traditional field methods to larger areas, yet challenges still
remain. This research aims to develop field methods and processes to
understand the limits of using Structure-from-Motion in semi-arid
ecosystems to estimate biomass. We believe that our research adds to
this process by establishing a methodology to further collect and
process small plots of biomass into classified point clouds in a
largely-automated fashion, so that additional work may be conducted and
expanded. Our results indicate that the ability to detect the ground may
be the limiting factor in attaining satisfactory volume-to-biomass
relationships in similar environments, and reiterating that fine spatial
resolutions are needed for fine-scale vegetation.
Author’s Contributions
Josh Enterkine: Methodology, Software, Validation,
Investigation, Resources, Data Curation, Writing - Review & Editing,
Supervision, Visualization. Ahmad Hojatimalekshah: Methodology,
Software, Validation, Formal Analysis, Data Curation, Writing - review
& editing, Visualization. Monica Vermillion: Methodology,
Software, Investigation, Writing – Original Draft. Thomas Van
Der Weide: Software. Sergio A. Arispe: Conceptualization,
Investigation, Supervision, Project administration, Funding acquisition.William J. Price: Methodology, Investigation, Resources, Data
Curation, Supervision. April Hulet: Conceptualization,
Methodology, Investigation, Resources, Funding acquisition.Nancy F Glenn: Conceptualization, Methodology, Writing –
Review & Editing, Supervision, Project administration, Funding
acquisition.
Running Title