论文信息

论文解读

AI 辅助整理
方法
消费级无人机RGB视频、运动恢复结构、单轨道与多高度盘旋航线对比。
发现
单轨道点云更集中空隙更少,加蛇形回环使可见空隙从17.57%降至5.07%。
机会点
预印本尚缺独立几何验证,可研究航线设计与点云质量对定量表型精度的影响。

论文摘要

用消费级无人机RGB视频与运动恢复结构重建田间玉米器官,比较单轨道与多高度盘旋及帧采样。单轨道点云更集中、空隙更少,增加蛇形回环使可见空隙从17.57%降至5.07%。属预印本,仍需独立几何验证方可用于定量表型。

全文

arXiv (2026)

Pointillist Maize: 3D Reconstruction of Field-Grown Maize Using
RGB UA V Imagery
Dewi Endah Kharismawati 1,2,3,4,* and Toni Kazic 1,2,3,5
1Department of Electrical Engineering and Computer Science,
University of Missouri, Columbia, MO, USA
2Missouri Maize Center, University of Missouri, Columbia, MO, USA
3Interdisciplinary Plant Group, University of Missouri, Columbia, MO, USA
4Department of Food, Agricultural and Biological Engineering,
The Ohio State University, Columbus, OH, USA
5MU Institute for Data Science and Informatics, University of Missouri, Columbia, MO, USA
*Corresponding author: kharismawati.2@osu.edu
Abstract
Three-dimensional reconstruction could extend field phenotyping from canopy summaries to
the morphology of individual plants and organs, but field-grown maize is difficult to reconstruct
because thin, flexible, mutually occluding structures move during image acquisition. We inves-
tigated whether consumer-grade UAV RGB video processed using Structure from Motion and
multi-view stereo could preserve recognizable maize structures, and how acquisition trajectories
and frame sampling affected reconstruction characteristics. COLMAP reconstructions preserved
recognizable leaves, stems, tassel branches, ears, visible silks, and brace roots in individual
field-grown plants. To compare image sampling strategies, we reconstructed the same plant using
two sets of 295 frames from the same UAV video: a single orbit sampled at 5 frames/second and
three consecutive orbits at different altitudes (ribbon skirt) sampled at 2 frames/second. After
ground removal, the reconstructions contained 520,601 and 375,318 points, respectively. The
single orbit produced greater point concentration within the plant structure and fewer visible
gaps, whereas the ribbon skirt produced fewer scattered points and slightly lower proportions
of isolated points under the tested neighborhood thresholds. In densely planted nursery rows,
visible reconstruction voids occupied 17.57% of the rendered scene using a single orbit at 1
frame/second, decreasing to 5.07% with additional serpentine passes and 2.68% when the original
orbit was sampled at 3 frames/second. However, additional observations also introduced greater
point dispersion associated with plant movement. A curlicue trajectory evaluated in a different
field during the 2024 growing season produced distinguishable individual plants without visible
duplicated structures, although substantial voids remained, primarily on the soil surface. These
findings demonstrate the feasibility of reconstructing recognizable maize organs from consumer
UAV imagery, highlight trade-offs among acquisition strategies and reconstruction completeness,
and provide a foundation for publicly accessible three-dimensional maize datasets, although
independent geometric validation remains necessary for quantitative phenotyping.
Keywords:3D reconstruction, Structure from Motion, point clouds, UAV, RGB video,COLMAP,
maize, field phenotyping
1
arXiv:2611.20502v1 [cs.CV] 16 Nov 2026

1 Introduction
High-throughput phenotyping remains a major bottleneck in crop genetics and breeding, especially
in field experiments where large populations must be observed repeatedly and at biologically relevant
stages (Furbank and Tester, 2011; Araus and Cairns, 2014; Minervini et al., 2015; Hui et al., 2018).
Unoccupied aerial vehicles (UAVs) have expanded the spatial coverage, frequency, and flexibility
of field data collection and enabled applications including multispectral crop-disease assessment
(Liebisch et al., 2015; Hu et al., 2018; Wang et al., 2019; Kharismawati et al., 2026; Liu et al., 2020;
Zhu et al., 2020b; Velumani et al., 2021; Adak et al., 2023). Much of this work summarizes plots or
canopies through spectral response, vegetation indices, temperature, or height. These measurements
are valuable for monitoring crop status, but canopy summaries do not directly represent the
morphology and development of individual leaves, tassels, ears, brace roots, or damaged tissues.
Those structures are often where the effects of genotype, development, disease, and environmental
stress become visible.
Extracting reliable plant-level information from aerial imagery is itself a computer vision problem.
Crop fields contain repeated rows and visually similar plants, while moving foliage, illumination
changes, shadows, and soil backgrounds introduce additional appearance variability. In freely flown
maize video, limited telemetry and the repetition of plants and rows also complicate frame registration
and mosaic construction (Kharismawati and Kazic, 2025a; Aktar et al., 2020; Kharismawati et al.,
2020). A successful field pipeline therefore depends not only on the downstream model, but also on
what the acquisition makes observable, whether images can be registered consistently, and whether
the representation preserves the spatial relationships required by the phenotype of interest. This
dependence on representation is also evident in maize stand counting, where detections registered
directly from raw video frames agreed more closely with field counts than detections from mosaicked
imagery (Kharismawati and Kazic, 2025b).
Organ-level phenotyping particularly requires structural information. A two-dimensional image
projects different depths onto the same plane: organs from adjacent plants may overlap, a front leaf
may hide an ear or stem, and a visible fragment may be difficult to assign to its plant of origin.
A three-dimensional representation can preserve relative position and orientation and can support
organ segmentation and structural measurement (Li et al., 2022a,b; Yang et al., 2024). This potential
is especially relevant in maize. Its leaves are long, thin, curved, and flexible; tassels contain many fine
branches; ears and silks may be partly concealed; brace roots occur near a cluttered soil boundary;
and neighboring plants crowd each other as the canopy closes. The same morphology that makes a
three-dimensional representation useful also makes it difficult to construct.
Recent studies demonstrate what can be achieved when acquisition is sufficiently controlled.
Multi-view RGB reconstruction has supported measurements of plant height, leaf length, leaf width,
and relative leaf area in maize seedlings (Li et al., 2022a). Organ-level point-cloud segmentation has
enabled extraction of leaf inclination, leaf growth height, and related structural traits from individual
maize shoots (Li et al., 2022b). Larger annotated maize point-cloud datasets further demonstrate the
importance of preserving stem and leaf geometry for automated analysis across diverse architectures
(Yang et al., 2024). LiDAR, controlled multi-camera systems, and turntable imaging can likewise
provide detailed plant models (Luo et al., 2021; Yin et al., 2020; Lin and Habib, 2021; Gollob et al.,
2021; Zhu et al., 2020a). These results establish the value of plant-scale 3D data, but many of these
approaches rely on stationary or potted plants, structured imaging platforms, controlled backgrounds,
or carefully constrained camera motion. Such conditions reduce the challenges associated with plant
movement, occlusion, and variable illumination, but are difficult to reproduce at the scale of an
open agricultural field. Consequently, methods that perform well under controlled conditions may
not scale readily to field environments, where large plant populations and dense canopies make it
2

difficult to capture the structural details of individual plants.
Field-grown plants present several challenges for conventional three-dimensional reconstruction
methods. Structure from Motion (SfM) estimates camera poses from correspondences among images,
while multi-view stereo densifies the recovered scene (Schönberger and Frahm, 2016; Schönberger
et al., 2016). Both are most straightforward when the scene is rigid, distinctive visual features persist
across views, and each surface is observed repeatedly (Schönberger and Frahm, 2016; Schönberger
et al., 2016). Maize has slender organs and extensive self-occlusion, which can produce incomplete
leaves and missing tips in multi-view reconstructions (Wu et al., 2022; Li et al., 2022a). Wind
can move individual organs while images are recorded, and variable illumination can change their
appearance between views (Wu et al., 2022; Li et al., 2022a). At range scale, overlapping leaves and
increasingly dense canopies increase occlusion and make individual plants more difficult to separate
(Zermas et al., 2020; Sweet et al., 2022). Consequently, a failure may reflect missing observations,
nonrigid movement, ambiguous correspondence, insufficient image resolution, or a combination of
these factors rather than simply an inadequate number of frames.
UAV reconstruction has been successful for field structure and plant height estimation, but it
has often emphasized canopy-scale surfaces rather than individual organs (Zermas et al., 2020).
Acquisition geometry is critical because a reconstruction cannot recover a surface that was never
observed. In field maize canopies, oblique UAV photography produced more complete vertical profiles
than nadir photography and recovered substantially more points below the upper canopy (Che et al.,
2020). A crop-surface-model study similarly found that oblique views improved maize reconstruction
relative to nadir views and that reduced spatial resolution lowered plant height accuracy (Niu
et al., 2024). Controlled multi-view platforms impose overlap requirements within and between
camera levels and report incomplete leaves and tips when coverage is insufficient; wind and variable
illumination remain important limitations when such systems move outside controlled environments
(Wu et al., 2022). Together, this literature indicates that camera placement, overlap, distance, and
view angle determine which structural evidence is available for reconstruction.
This creates a field acquisition trade-off that is not captured by frame count alone. Sampling
many adjacent frames from a video can increase temporal redundancy and local point density, but
closely spaced frames may contribute nearly the same evidence and cannot expose a persistently
occluded surface (Gibbs et al., 2020). Changing elevation or camera angle can reveal surfaces that
are hidden in nadir or single-level views (Che et al., 2020; Niu et al., 2024; Wu et al., 2022). Longer
acquisitions also increase exposure to wind-driven motion of flexible plant organs, which can reduce
consistency among views (Wu et al., 2022; Kharismawati and Kazic, 2026). Flying farther away
increases coverage but decreases spatial resolution and the number of pixels representing a leaf
edge, tassel branch, or silk (Niu et al., 2024). Thus, temporal sampling, viewpoint diversity, target
resolution, acquisition duration, and scene motion interact. A fair comparison must hold at least the
source sequence and image budget constant rather than equating more frames or more reconstructed
points with better evidence.
Given the prior work, it is not obvious that plants in field environments can be reconstructed
sufficiently for organ phenotyping. We therefore undertook a qualitative proof-of-concept study to
determine the feasibility of plant reconstructions for different planting situations in the field, moving
from isolated plants to the standard nursery planting of rows and ranges. Here, we show that plants
in field environments can be reconstructed in three dimensions for organ phenotyping from video shot
from consumer-grade UAVs. We explore different planting schemes and trajectories and examine the
trade-offs between these, reconstruction quality and resolution, and acquisition times. Established
software such asCOLMAPmakes SfM and multi-view stereo accessible without requiring a new
reconstruction algorithm (Schönberger and Frahm, 2022–present, 2016). The unresolved question is
what a conventional and reproducible workflow can recover from UAV video of mature maize in the
3

field and what limits emerge as acquisition scales up from one to many plants. This is important for
researchers designing data acquisition strategies, whether with their existing UAVs or any prospective
purchase of specialized, more complex instruments.
This study makes three contributions. First, it demonstrates the feasibility of imaging unstruc-
tured fields and the quality of the three-dimensional reconstructions of maize plants and their organs,
including leaves, tassels, ears, silks, and brace roots. Second, it compares two image sampling
strategies, examining how different combinations of temporal sampling and acquisition viewpoints
affect how the points are distributed in the three-dimensional point cloud relative to the plants’
structures. Third, it investigates how the plants’ structural details change with flight trajectories
that capture larger and more complex scenes. Together, these findings provide practical guidance for
acquiring and processing UAV imagery for 3D reconstruction and support the longer term goal of
developing publicly accessible 3D datasets of maize grown under field conditions.
2 Materials and Methods
2.1 Field Site and Plant Material
Maize (Zea mays) was grown in Columbia, Missouri, USA, during the 2022 and 2024 field season.
The material included three inbreds, an elite line, and approximately 100 near-isogenic disease lesion
mimic mutant lines. The field contained two relevant layouts. In a tiling section, individual plants
were placed at vertices of nearly equilateral triangles with sides of approximately 4.5 m. Five plants
each of an inbred, an elite line, and one lesion mimic mutant (Mo20W/les∗-N1395C) were planted in
this section. The separation permitted imaging around an individual plant with limited interference
from neighboring plants. In the genetic nursery, rows were 0.9 m apart and plants were spaced 24.5
cm apart within rows. A range is a set of nursery rows bounded by unplanted alleys. The same field
layout was maintained during the 2024 growing season, with the addition of 14-inch pots installed
around individual plants to improve watering efficiency.
2.2 UAV Imaging and Flight Configurations
Video was acquired manually with a 2021 DJI Mavic 2 Pro carrying its RGB camera. The source
videos were recorded at 30 frames/second and3840× 2160pixels. Flights were designed to capture
diverse viewpoints of plant surfaces rather than follow predefined automated flight paths.
2.2.1 Individual Plant Acquisition
Figure 3 shows two different orbital trajectories for acquiring individual plants. In the top panel of
Figure 3(a) we show a single-orbit trajectory, in which the UAV completed one full circuit around
the plant at one flight level, at an approximate radius of 1–2 m, and different camera angles. The
bottom panel of Figure 3(a) shows a multi-level trajectory that we define here as aribbon skirt:
three completed, vertically stacked orbits around the same plant, with the flight altitude and camera
angle changed between orbits. The UAV completed each full orbit at a fixed level before increasing
its altitude and beginning the next orbit. During the first ribbon skirt orbit, the UAV flew at
approximately half the plant height with a parallel camera pose, defined here as a dead-on view
with the camera optical axis approximately parallel to the ground. During the second orbit, the
UAV flew at approximately the height of the plant and the camera was angled slightly downward to
provide an oblique view. During the third orbit, the UAV flew approximately 2 m above the top
of the plant and the camera was angled more steeply downward to provide a sharply oblique view.
4

Thus, the UAV ascended only between completed orbits; the ribbon skirt was not a continuously
ascending spiral. The UAV was flown manually, and the 1–2 m orbit radius, flight altitudes, and
viewing angles are approximate rather than surveyed values.
2.2.2 Range Level Acquisition
Ranges were imaged at greater distance using three distinct trajectory types: an orbit around the
range with an oblique camera pose, serpentine passes approximately perpendicular to the rows with
oblique or nadir views, and a curlicue path. Curlicue denotes a continuous looping pattern with a
smaller radius at approximately constant altitude. It was used for a range containing multiple rows
and is distinct from the multi-level individual plant acquisition. Source descriptions give approximate
altitude ranges of 5–20 m for range orbits and 3–10 m for serpentine passes. The principal range
example was acquired at approximately 5–7 m. These values describe the broader collection rather
than a controlled single-factor comparison. The 2024 curlicue experiment was conducted in a different
field during a subsequent growing season, while maintaining the same planting layout as the 2022
nursery range experiments.
2.3 Video Frame Extraction and Experimental Configurations
Frames were extracted from video at the sampling rates listed in Table 1. Comparisons of different
flights are descriptive because plant material, flight altitude, camera geometry, scene complexity,
and environmental motion can vary together.
The controlled equal-frame comparison used one source video of the same individual maize plant.
Configuration A comprised 295 frames from one orbital level and viewing configuration, sampled at
5 frames/second. Configuration B comprised 295 frames spanning three consecutive complete orbits
at different elevations and angles, sampled at 2 frames/second. Thus total frame count and source
video were held constant, while the distribution of viewpoints differed. Frame sampling rate and
viewpoint diversity are therefore separate experimental factors.
Scene T rajectory camera pose frames/second frames purpose
plant one orbital level single level/view 5 295 equal-frame comparison
plant three complete orbits parallel to oblique 2 295 equal-frame comparison
plant two orbits parallel 1 177 wild-type example
plant two orbits parallel 1 129 mutant example
plant two orbits parallel and oblique 2 219 pose and masking examples
tiling one orbit oblique 1 150 separated plant group
range one orbit oblique 1 126 range example
range orbit plus serpentine oblique and nadir 1 208 trajectory comparison
range one orbit oblique 3 426 sampling comparison
range (2024) curlicue oblique∼0.5 287 curlicue reconstruction
Table 1: Verified video subsets represented in the manuscript sources. The two 295-frame configura-
tions came from the same source video. Other rows are not controlled single-factor comparisons.
Frames/second is the rate at which frames were extracted from the 30 frames/second source video.
2.4 3D Reconstruction and Postprocessing
Extracted frames were processed withCOLMAPusing one NVIDIA RTX 2080 Ti GPU in a
computer with two RTX 2080 Ti GPUs and 128 GB RAM.COLMAPdetected and matched image
features, estimated camera poses with SfM, calculated depth with multi-view stereo, and fused the
5

depth estimates into dense point clouds, which were exported as PLY files (Schönberger and Frahm,
2022–present, 2016).
Some image sets were masked before reconstruction. One approach transformed RGB frames to
the device-independentLabcolor space, where L represents lightness, a* represents the green to
red color axis, and b* represents the blue to yellow color axis. Pixels were selected based on their
values in these color channels. A combined mask retained pixels satisfying
θ= [(a∗ <0)∧(b ∗ >0)]∨[(a ∗ >0)∧(b ∗ <0)],(1)
where θ represents the resulting mask, < and > indicate values below and above the specified
thresholds, respectively,∧ means that both conditions must be satisfied, and∨ means that either
group of conditions may be satisfied. The first group selects pixels with green and yellow color
components, while the second selects pixels with red and blue color components. The latter
was intended to retain some tassel regions but could also include sky pixels. A more restrictive
experimental mask used
θ= (L∗ <35)∧(a ∗ <−20)∧(b ∗ >10).(2)
This mask retained only pixels that simultaneously had a lightness value below 35, ana∗ value
below −20(toward green), and ab∗ value above 10 (toward yellow). These threshold values were
determined empirically through repeated visual inspection and adjustment of the resulting masks.
The selected values provided the best observed retention of plant structures, including leaves, tassels,
and ears, among the threshold combinations tested. RGB frames were also masked using excess
green,
ExG = 2G−R−B,(3)
where R, G, and B represent the red, green, and blue color channel values, respectively. Excess green
emphasizes vegetation by increasing the contribution of green relative to red and blue. The resulting
values were normalized to the range[0, 1], followed by Otsu thresholding to automatically separate
pixels into two groups based on their excess green values. Thresholds were selected by inspecting
whether masks retained leaves, tassels, and ears while excluding soil and sky. These methods were
exploratory and were not trained segmentation models.
Point clouds were inspected and postprocessed in MeshLab (Cignoni et al., 2008). Cropping
denotes restricting a reconstruction to a region of interest. Manual ground cleaning denotes selection
and removal of soil and ground points. For the equal-frame comparison, approximately the same
manual ground-removal procedure was applied to both clouds. Above-ground points around leaves,
tassels, and other structures were not manually removed. The reported 520,601 and 375,318 points are
therefore counts after soil and ground removal, not counts from the original unprocessedCOLMAP
clouds and not accuracy metrics.
Surface reconstruction was explored with Poisson and ball pivoting methods. The most useful
visual results in these experiments came from MeshLab’s automatic ball-radius estimate with a
clustering radius equal to 10% of the ball radius; one manually edited example used 5% (Bernardini
et al., 1999). These meshes were visualizations of the point clouds, not geometric validation. Manual
above-ground cleaning was performed only for the separate surface-reconstruction demonstration
and must not be confused with the equal-frame comparison.
2.5 Evaluation
Reconstructions were evaluated through visual inspection of recognizable morphological structures,
apparent structural continuity, scattered points, missing surfaces, duplicated structures, and misregis-
6

tration of soil or plants. Point counts were used to describe retained samples after specified processing
steps, while local point density and isolation were quantified for the two configurations compared
using an equal number of frames. The evaluation focused on visual and spatial characteristics of the
reconstructions because independent physical 3D references and validated organ measurements were
not available.
For the equal-frame comparison, local point density and isolation were also quantified directly
from the two post-ground-removal PLY files using Python 3.9, NumPy 2.0.1, and SciPy 1.13.1.
To define a common spatial scale,s, we calculated the pooled median Euclidean nearest-neighbor
distance across both point clouds using a k-d tree (s = 0.001279coordinate units). Then for each
pointi, the number of neighboring points within a fixed radiusrwas calculated as
Ni(r) =
{j̸=i:∥x j −x i∥2 ≤r} ,(4)
where the outer vertical bars denote the number of points satisfying the distance condition,xi andx j
represent the three-dimensional coordinates of pointsi and j, respectively, andr∈ {2s, 3s, 5s, 10s}.
The median number of neighboring points and the proportions of points having zero, at most two, or
at most five neighbors were calculated to characterize point concentration and isolation. To determine
whether the results depended on differences in sampling density, the nearest neighbor distance for
each point was also divided by the median nearest neighbor distance within its respective cloud.
Similarly, the mean distance to the eight nearest neighboring points was divided by the corresponding
median within each cloud, and the proportions of points exceeding predefined multiples of these
normalized distances were reported. These measurements characterize the spatial distribution of
points, distinguishing densely concentrated points within the reconstructed plant structure from
sparse or isolated points in peripheral regions of the reconstructions. However, one or more isolated
points are not necessarily a reconstruction artifact, as they may represent a legitimate leaf edge,
tip, tassel branch, or an incompletely observed surface, or an organ moving during acquisition. No
statistical significance test was performed at the individual point level because the two point clouds
represent single experimental units, and their constituent points are not independent biological or
acquisition replicates.
For nursery range reconstructions, we quantified visible reconstruction voids using snapshots of
the rendered point clouds shown in Figure 5. Void regions were identified by their yellow color in
the rendered images. For the three reconstructions from the 2022 field experiment, snapshots were
selected from the same nursery location and viewing region to ensure a consistent comparison. For
the curlicue reconstruction acquired in a different field during the 2024 growing season, a snapshot
of a comparable nursery region was selected. The visible void area percentage was calculated as
the number of yellow pixels corresponding to reconstruction voids divided by the total number of
pixels in each snapshot, multiplied by 100. This measurement represents the projected extent of
visible reconstruction gaps rather than the three-dimensional area or volume of missing geometry.
Because the curlicue experiment was conducted in a different field and growing season, its void
percentage was interpreted descriptively rather than as a direct performance comparison with the
2022 reconstructions.
3 Results
3.1 Feasibility of Individual Plant Reconstruction
The first question was whether consumer UAV RGB video could preserve recognizable maize
morphology in an isolated field-grown plant. For the W23 plant, two orbital trajectories were
7

(a) Representative RGB
frame.
(b) Estimated camera poses and
sparse point cloud.
(c) Dense point cloud.
Figure 1: Three-dimensional reconstruction of an individual W23 maize plant using two orbital
trajectories with the same parallel camera orientation. (a) Representative RGB frame showing
the field-grown plant. (b) Estimated camera poses and sparse point cloud from the two orbital
acquisitions. (c) Dense point cloud showing recognizable plant architecture, including the stem,
curved leaves, brace roots, a senescent lower leaf, and the tassel.
acquired with the camera oriented approximately parallel to the ground (Figure 1a,b). The estimated
camera poses surrounded the plant, providing overlapping observations from multiple directions.
The resulting dense point cloud preserved the overall plant architecture, including the stem, curved
leaves with different orientations, brace roots, a senescent lower leaf, and the tassel (Figure 1c).
Individual leaves remained visually distinguishable, including those that crossed or partially occluded
one another in the original images. The initial dense reconstruction contained 17,009,512 points, of
which 1,990,141 remained after cropping to the plant region and 613,244 remained after removing
most of the surrounding soil. The reconstruction also contained scattered points, particularly around
leaf tips and the tassel, where movement during acquisition affected the consistency of observations.
Points near the plant boundaries sometimes retained colors associated with the surrounding soil or
sky. Despite these artifacts, the reconstructed point cloud preserved recognizable vegetative and
reproductive structures of the W23 plant.
A second plant was reconstructed using two orbital trajectories with different camera orientations.
The first orbit used a camera orientation approximately parallel to the ground, while the second
used an oblique orientation to capture additional viewpoints of the upper plant (Figure 2a,b). The
parallel orbit was acquired at approximately 0.6 m above the ground, while the oblique orbit was
acquired at approximately 3 m with a camera angle of 20–30 degrees. COLMAP processed 219
video frames and generated 25,010,235 points, with 1,288,682 points remaining after removing
the surrounding background and soil. The resulting dense point cloud preserved the overall plant
architecture and finer reproductive structures, including individual tassel branches, an ear, and
8

visible silks (Figure 2c–e). The oblique viewpoints provided additional observations of the upper
plant and tassel, which were less accessible from the lower parallel camera orientation.
Plant movement during orbital acquisition introduced additional structures that did not cor-
respond to distinct physical organs. Leaves changed position between observations in response
to ambient wind and air currents generated by the UAV. These changes violated the static-scene
assumption underlying the reconstruction and produced duplicated or displaced leaf structures,
referred to here as ghost organs. One example is visible in Figure 2c (white arrow), where a moving
leaf was reconstructed in two different positions. Movement was particularly apparent in younger
upper leaves and leaf tips, which are more flexible than the larger and more rigid plant structures.
The resulting ghost organs introduced ambiguity into the reconstructed leaf geometry even though
the overall plant architecture remained recognizable.
To obtain a surface representation of the reconstructed plant, the ball pivoting algorithm in
MeshLab was applied to the dense point cloud. The algorithm connected neighboring points with
triangular faces, converting the point-based representation into a surface mesh. The resulting
surfaces preserved recognizable morphological features, including leaf midribs, ear husk leaves, visible
silks, and tassel branches (Figure 2f,g). Some fine structures, particularly the silks, became easier
to distinguish in the surface representation than in the original point cloud. However, regions
with incomplete or inconsistent point coverage contained holes and irregular surface connections,
particularly around thin and complex organs. The ball pivoting reconstruction therefore provided
a surface representation of the maize plant, but did not recover missing geometry or eliminate
motion-related artifacts already present in the underlying point cloud.
3.2 Equal Frame Budget, Different Viewpoint Distributions
The equal-frame experiment examined how different viewpoint distributions and temporal sampling
rates affected the reconstruction of an individual maize plant while keeping the total number of input
frames constant. Both reconstructions used 295 frames extracted from the same source video. The
single-orbit configuration sampled frames at 5 frames/second from one orbital trajectory, with the
camera positioned at approximately the same height as the plant and oriented slightly downward. In
contrast, the ribbon skirt configuration sampled frames at 2 frames/second across three consecutive
orbits at different altitudes and camera orientations (Figure 3a). After comparable manual removal
of ground and soil points, the single-orbit reconstruction contained 520,601 points, while the ribbon
skirt reconstruction contained 375,318 points. No scattered points above the ground were selectively
removed from either reconstruction.
The two configurations produced visibly different point-cloud structures (Figure 3b,c). The single-
orbit reconstruction contained more points within the plant structure and showed fewer visible gaps
in the leaves and other reconstructed organs. However, it also contained more stray points distributed
outside the recognizable plant structure, particularly around the leaf boundaries. In contrast, the
ribbon skirt reconstruction showed fewer stray points and a more clearly defined distribution of
points around the plant. Nevertheless, several regions of the plant contained visible gaps, indicating
incomplete reconstruction of some organ surfaces. Thus, the single-orbit configuration produced
a visually more complete representation of the plant structure, but with greater point dispersion,
whereas the ribbon skirt produced a cleaner but less complete point cloud.
Differences in camera orientation were also reflected in the colors of points near the plant
boundaries. In the single-orbit reconstruction, some points along the leaf edges appeared brown,
consistent with the soil background visible from the slightly downward camera orientation. In the
ribbon skirt reconstruction, blue points were visible along portions of the plant, particularly in
regions observed from the approximately parallel camera orientation, where the sky formed part
9

(a)
Representative
RGB frame.
(b) Estimated camera poses
and sparse point cloud.
(c) Dense point
cloud.
(d) Tassel detail.
(e) Ear and silk detail.
(f) Ball pivoting surface at
the ear region.
(g) Ball pivoting surface at
the tassel region.
Figure 2: Three-dimensional reconstruction of an individual maize plant using two orbital trajectories
with parallel and oblique camera orientations. (a) Representative RGB frame. (b) Estimated camera
poses and sparse point cloud showing the two orbital acquisitions. (c) Dense point cloud showing
the reconstructed plant morphology. (d,e) Detailed views of the reconstructed tassel, ear, and silks.
(f,g) Surface meshes generated using the ball pivoting algorithm, showing recognizable structures in
the ear and tassel regions alongside areas of incomplete surface continuity. In both (a) and (c), the
red arrows identify a leaf torn along the midrib, with the broken midrib protruding above the leaf
blade, while the white arrows identify a leaf that moved considerably during acquisition.
of the background. These observations suggest that background colors can be incorporated into
reconstructed points near thin plant boundaries, depending on the camera viewpoint and surrounding
scene.
The quantitative analysis further characterized differences in point concentration and local
10

isolation between the two reconstructions (Table 2). The pooled median nearest-neighbor distance
was s = 0.001279reconstructed coordinate units and was used as a common reference spacing
for both point clouds. The single-orbit reconstruction had a median nearest-neighbor distance of
0.001304, compared with 0.001243 for the ribbon skirt reconstruction. Although the ribbon skirt
had a slightly smaller median nearest-neighbor distance, the single-orbit cloud had greater median
neighborhood counts at larger common radii. Atr = 5s, the median neighbor counts were 71 and 66
for the single orbit and ribbon skirt, respectively. Atr = 10s, the corresponding counts were 401
and 339. These results indicate greater local point concentration at the tested larger radii in the
single-orbit reconstruction.
The ribbon skirt reconstruction, however, had a slightly smaller proportion of locally isolated
points at each tested threshold. Atr = 3s, 2.696% of single-orbit points and 2.411% of ribbon skirt
points had no other point within the radius. Atr = 5s, 1.936% and 1.678% of points, respectively,
had at most two neighbors. Similarly, atr = 10s, 0.515% of single-orbit points and 0.419% of ribbon
skirt points had at most five neighbors. These differences are consistent with the visual observation
that the ribbon skirt contained fewer scattered points outside the plant structure, although local
isolation does not necessarily indicate that a point is an outlier. However, these differences should
be interpreted in the context of the unequal point counts. The single-orbit reconstruction contained
approximately 39% more points than the ribbon skirt reconstruction, which may partly explain the
greater number of locally isolated points. Although the isolation metrics are expressed as proportions
rather than absolute counts, differences in the total number and spatial distribution of reconstructed
points can affect the proportion falling below each neighborhood threshold. Therefore, the slightly
higher isolation percentages in the single-orbit reconstruction should not be interpreted as evidence
that its acquisition strategy inherently produces more outliers. However, the isolation metrics do
not measure the completeness of reconstructed plant organs and therefore do not capture the visible
gaps observed in the ribbon skirt reconstruction.
Overall, the equal-frame comparison revealed a trade-off between structural coverage and point
dispersion. The single-orbit configuration retained more points and showed fewer visible gaps within
the plant structure, but also introduced more stray points. The ribbon skirt configuration produced
fewer scattered points outside the plant but left more regions of the plant incompletely reconstructed.
Because the two configurations differed in both temporal sampling rate and viewpoint distribution,
the observed differences cannot be attributed to either factor independently. The results show that
an equal number of input frames does not necessarily produce equivalent reconstruction coverage or
point-cloud characteristics.
3.3 Scaling from Isolated Plants to Nursery Rows
The separated plant tiling provided an intermediate scenario between reconstructing an isolated
maize plant and a densely planted nursery range. A circumferential orbit acquired at an approximate
altitude of 5–7 m with an oblique camera orientation captured several spatially separated plants
within the same scene (Figure 4a,b). The initial dense reconstruction contained 20,105,936 points, of
which 15,986,041 remained after cropping. Individual plants remained recognizable in the resulting
point cloud, although finer morphological details were less apparent than in the close views of isolated
plants (Figure 4c). The separation between plants reduced mutual occlusion, allowing their overall
structures to remain distinguishable despite the greater viewing distance. The reconstruction also
captured surrounding field features, including tire tracks and the edge of the neighboring genetic
nursery.
We next examined reconstruction of a densely planted nursery range containing multiple maize
rows. Unlike the separated plant tiling, the nursery contained plants with overlapping leaves,
11

Metric One orbit, 5 fps Ribbon skirt, 2 fps
Retained points 520,601 375,318
Bounding-box diagonal 1.895069 1.857731
Median nearest-neighbor distance 0.001304 0.001243
MedianN i(5s)71 66
MedianN i(10s)401 339
Ni(3s) = 02.696% 2.411%
Ni(5s)≤21.936% 1.678%
Ni(10s)≤50.515% 0.419%
Table 2: Point-density and isolation diagnostics for the equal-frame reconstructions. Both configura-
tions used 295 frames from the same source video. The common reference spacing,s = 0.001279
reconstructed coordinate units, was calculated as the median of the pooled nearest-neighbor distances
from both point clouds.Ni(r)represents the number of neighboring points within radiusr, excluding
the query point itself. The metrics characterize local point concentration and isolation but do not
directly measure the completeness or geometric accuracy of reconstructed plant organs.
uneven plant densities, and substantial occlusion between neighboring rows. Three reconstruction
experiments were conducted on the same nursery range to examine the effects of acquisition trajectory
and frame sampling on the resulting point clouds (Figure 5).
The first experiment used a single orbital trajectory with an oblique camera orientation at
approximately 5–7 m altitude. A total of 126 frames sampled at 1 frame/second produced a dense
reconstruction containing 16,248,202 points. The overall arrangement of the nursery rows and
individual plants remained recognizable. However, the close view revealed substantial voids in the
reconstruction, particularly between densely planted rows where leaves overlapped and obscured
one another (Figure 5a). These regions appeared as gaps or bluish areas in the reconstructed point
cloud, indicating that portions of the plant structure were not recovered.
In the second experiment, a serpentine trajectory was added to the orbital acquisition to provide
additional observations of the nursery range. The serpentine acquisition included passes with
nadir and oblique camera orientations, with the oblique view directed perpendicular to the rows.
The reconstruction used 208 frames sampled at 1 frame/second and produced 25,369,115 points.
Compared with the single orbit, the additional viewpoints reduced the visible voids and recovered
more of the plant structure (Figure 5b). However, the reconstructed leaves appeared less sharply
defined. Movement between acquisition passes caused slight differences in leaf positions, resulting
in more dispersed points around some leaf surfaces. Thus, the additional trajectory improved the
visible coverage of the nursery plants, but also introduced artifacts associated with plant movement.
The third experiment retained the original orbital trajectory but increased the frame sampling
rate from 1 to 3 frames/second, resulting in 426 input frames. The resulting reconstruction showed
fewer visible voids than the lower sampling rate reconstructions, with many previously incomplete
regions becoming more densely represented (Figure 5c). However, the increased number of frames
also produced greater point dispersion around the reconstructed plant structures. This dispersion was
consistent with leaf movement captured across the additional frames, which introduced differences
between observations of the same organs.
The fourth experiment used a curlicue trajectory in a different maize field during the 2024
growing season. The trajectory was intended to provide more varied viewpoints of densely planted
maize and examine whether this acquisition pattern could reduce reconstruction artifacts. A total
of 287 frames were extracted at approximately 0.5 frames/second from a video lasting 9 minutes
and 33 seconds. The resulting dense reconstruction contained 7,811,270 points, fewer than the three
reconstructions from the 2022 nursery range. Despite the smaller point cloud, individual plants were
12

(a) Flight trajectories: top, single orbit at 5
fps; bottom, ribbon skirt at 2 fps.
(b) Single orbit, 295 frames;
520,601 points.
(c) Ribbon skirt, 295 frames;
375,318 points.
Figure 3: Comparison of two acquisition configurations using an equal number of frames from
the same UAV video. (a) Single-orbit trajectory at approximately plant height and ribbon skirt
trajectory consisting of three orbits at different elevations and camera orientations. (b) The single-
orbit reconstruction contains more points and fewer visible gaps within the plant structure, but
also shows more stray points outside the plant. Brown points along some plant boundaries are
consistent with the soil background visible from the downward camera orientation. (c) The ribbon
skirt reconstruction contains fewer scattered points outside the plant, but shows more gaps within
the reconstructed plant structure. Blue points near some plant boundaries are consistent with the sky
background visible from approximately parallel viewpoints. Both reconstructions used 295 frames.
Ground and soil points were removed manually using comparable procedures in MeshLab, while
above-ground scattered points were left unedited. The displayed point counts correspond to the
clouds after ground removal.
clearly distinguishable from their neighbors, without visible duplicated or ghost plant structures or
double-layered soil surfaces (Figure 5d; Figure 7). However, substantial voids remained, primarily
on the reconstructed soil surface. Thus, the curlicue acquisition produced recognizable individual
plant structures without visible plant duplication, although the surrounding ground surface remained
incompletely reconstructed.
The quantitative evaluation of visible reconstruction voids further characterized the differences
among the nursery range reconstructions. In the 2022 field experiment, voids occupied 17.57% of the
rendered scene from the single orbit sampled at 1 frame/second. The visible void area decreased to
5.07% with the addition of serpentine passes and to 2.68% when the original orbit was sampled at 3
frames/second. In the 2024 curlicue experiment, visible voids occupied 37.50% of the selected scene,
primarily on the soil surface. Although the curlicue reconstruction contained a larger proportion
of visible voids, individual plants remained distinguishable without visible duplicated structures.
Because the curlicue experiment was conducted in a different field and growing season, its point
counts and void percentage are not directly comparable with those of the 2022 reconstructions.
13

(a) Representative RGB frame.
(b) Estimated camera poses and
sparse point cloud.
(c) Dense point cloud.
Figure 4: Three-dimensional reconstruction of a separated plant tiling using one circumferential
orbit with an oblique camera orientation at approximately 5–7 m altitude. (a) Representative RGB
frame showing several spatially separated maize plants. (b) Estimated camera poses and sparse
point cloud. (c) Dense point cloud showing recognizable individual plants and surrounding field
features, including tire tracks and the edge of a neighboring genetic nursery.
The nursery range experiments revealed different reconstruction characteristics associated with
acquisition trajectory and frame sampling. In the 2022 experiments, the single orbit preserved the
overall nursery arrangement but contained substantial gaps in regions with overlapping vegetation.
Adding serpentine passes reduced these gaps while introducing greater point dispersion around some
leaf surfaces. Increasing the frame sampling rate along the original orbit further reduced visible
voids, but also increased point dispersion. The curlicue trajectory, evaluated in a different field
during the 2024 growing season, produced a smaller point cloud in which individual plants were
clearly distinguishable without visible duplicated or ghost plant structures. However, substantial
reconstruction voids remained, primarily on the soil surface. These observations suggest that reducing
visible plant duplication and improving the separation of neighboring plants do not necessarily
eliminate gaps in the reconstructed scene. The curlicue result provides an additional example
of reconstruction behavior rather than a controlled comparison with the three earlier acquisition
strategies.
3.4 Failure Modes and Limits of Postprocessing
The reconstruction experiments revealed several limitations that became particularly apparent when
imaging densely planted nursery ranges. Occlusion and shadowing were associated with missing
regions in both plant and soil surfaces. In the nursery reconstructions, these gaps were most visible
where neighboring plants overlapped or where portions of the scene were poorly observed from the
available camera viewpoints. Although additional viewpoints and frames reduced some of these gaps
(Section 3.3), they did not eliminate all reconstruction artifacts.
Other failures involved the placement of reconstructed structures rather than missing points. In
some nursery reconstructions, the soil appeared as separated or overlapping planes, while individual
plants were duplicated (Figure 6a,b). These structures did not correspond to the physical arrangement
of the nursery and indicated inconsistencies in the reconstruction and registration of the scene. The
repetitive appearance of maize plants, similar leaf structures, and changes in plant position between
observations presented additional challenges for matching corresponding image features. However,
the individual contributions of these factors were not evaluated separately.
In contrast, the curlicue reconstruction from the 2024 growing season showed no visible duplicated
plant structures or double-layered soil surfaces in the inspected regions (Figure 7). Red arrows
indicate representative individual plants and the single-layered soil surface. Although these artifacts
14

(a) Single orbit, 126 frames at 1 fps.
(b) Orbit and serpentine, 208 frames at
1 fps.
(c) Single orbit, 426 frames at 3 fps.
(d) Curlicue, 287 frames at
approximately 0.5 fps.
Figure 5: Close views of four reconstruction experiments conducted on densely planted maize nursery
ranges. Panels (a)–(c) show reconstructions from the same nursery range in 2022, while panel (d)
shows the curlicue reconstruction from a different field with the same planting layout in 2024. (a) A
single orbital trajectory with an oblique camera orientation produced 16,248,202 points from 126
frames. Substantial voids are visible between densely planted rows. (b) Combining orbital and
serpentine trajectories produced 25,369,115 points from 208 frames. Additional viewpoints reduced
the visible voids, but leaf movement between passes produced less sharply defined structures. (c)
Increasing the sampling rate of the original orbital trajectory to 3 frames/second produced 77,361,305
points from 426 frames. Many visible voids were reduced, although greater point dispersion appeared
around the plant structures. (d) The curlicue trajectory produced 7,811,270 points from 287 frames
sampled at approximately 0.5 frames/second. Individual plants appeared more distinguishable, but
substantial reconstruction voids remained.
were not observed in the selected views, substantial reconstruction voids remained, particularly on
the soil surface.
Plant movement also affected reconstruction at the individual plant scale. As observed in
Section 3.1, moving leaves could produce ghost organs or dispersed points around leaf margins, tips,
and tassels. Ambient wind and UAV rotor downwash were possible contributors to this movement,
although their individual effects were not measured. Unlike gaps caused by missing observations,
these artifacts introduced conflicting representations of the same plant structure. Consequently, the
reconstructed point cloud could contain both missing and duplicated geometry.
Color masking was evaluated as a way to reduce background content before three-dimensional
reconstruction. However, the color characteristics of maize organs overlapped with those of the
15

(a) Misregistered soil surfaces.
(b) Duplicated plant structures.
Figure 6: Examples of geometric reconstruction artifacts observed in densely planted maize nursery
ranges. (a)Separatedsoilsurfacesindicatinginconsistentreconstructionoftheground. (b)Duplicated
plant structures that do not correspond to distinct physical plants. These artifacts illustrate errors in
scene reconstruction and registration that are separate from gaps caused by occlusion or insufficient
observations.
surrounding environment. The ExG andLab*masks could suppress soil and other background
regions, but also removed genuine plant tissue with brown, yellow, orange, or red coloration.
This was particularly problematic for senescent leaves, tassels, and other structures that were not
predominantly green.
The effect was visible in the reconstruction generated from ExG-masked images (Figure 8). The
initial point cloud contained 2,842,193 points, which decreased to 463,607 after cropping to the plant
region and 378,646 after removing black boundary points. These reductions reflected successive
processing operations rather than improvements in reconstruction accuracy. Although masking
suppressed portions of the background, it also removed image information needed to reconstruct
genuine plant structures. The resulting point cloud lacked portions of the tassel and brace roots, as
indicated by the red arrows in Figure 8.
Applying the ball pivoting algorithm to generate a triangular surface mesh did not restore the
missing structures (Figure 8c). Ball pivoting connected available points into surface faces, but could
not reconstruct organs whose image information had been removed by masking. The resulting
surface therefore retained gaps corresponding to missing plant structures. This example illustrates
a limitation of color based background suppression: reducing unwanted scene content can also
eliminate biologically meaningful structures when plant and background colors overlap.
(a) Individual plants.
(b) Soil surface.
Figure 7: Detailed views of the 2024 curlicue reconstruction. (a) Individual maize plants are distin-
guishable without visible duplicated or ghost plant structures. Red arrows indicate representative
plants. (b) The reconstructed soil surface appears as a single layer, without visible double-layering.
The red arrow indicates the reconstructed soil surface.
16

(a) ExG-masked RGB
frame.
(b) Reconstructed point
cloud.
(c) Ball pivoting surface
mesh.
Figure 8: Effects of ExG masking on the reconstruction of an individual maize plant. (a) Represen-
tative RGB frame after applying the ExG mask. (b) Reconstructed point cloud showing missing
plant structures following background suppression. The upper red arrow identifies the missing tassel,
while the lower red arrow identifies missing brace roots. These structures were excluded because
their colors did not satisfy the green vegetation mask. (c) Triangular surface mesh generated using
ball pivoting. The missing structures remain absent from the surface reconstruction because the
corresponding image information was removed before reconstruction.
4 Discussion
4.1 From Visual Reconstruction to Quantitative Plant Phenotyping
The reconstruction of individual maize plants demonstrates that consumer UAV RGB imagery can
preserve substantial morphological information under field conditions without requiring a fixed
imaging platform or controlled background. The reconstructed point clouds retained recognizable
leaves, stems, brace roots, tassel branches, ears, and silks. These observations are important because
the geometry of these organs provides information that is difficult to recover from conventional
two-dimensional images alone. For example, leaf orientation, the spatial arrangement of leaves around
the stem, and the position of reproductive structures are inherently three-dimensional characteristics.
However, recognizing an organ in a point cloud is different from establishing that its dimensions or
geometry have been reconstructed accurately. The present results demonstrate the visibility of these
structures, but do not yet establish the accuracy required for quantitative phenotyping.
Previous studies have demonstrated the potential of three-dimensional plant reconstruction for
extracting morphological traits under more controlled acquisition conditions. Li et al. (2022a) recon-
structed maize seedlings using multiple views and reported strong correlations between reconstructed
traits and manual measurements, including plant height, leaf length, leaf width, and relative leaf
area. They further demonstrated organ segmentation for extracting structural traits from maize
point clouds (Li et al., 2022b), while Yang et al. (2024) investigated stem and leaf segmentation
17

across a dataset of 428 maize plants. These studies establish that three-dimensional representations
can support quantitative measurements when the reconstructed geometry is sufficiently reliable. Our
experiments address a different part of this problem: whether recognizable plant structures can
be recovered from freely acquired UAV imagery in an agricultural field, where the plants are not
stationary, the background is uncontrolled, and neighboring plants may obscure one another.
This distinction becomes particularly important when moving from isolated plants to nursery
rows. At the individual plant scale, the camera could observe the plant from multiple directions with
relatively little interference from neighboring vegetation. As the acquisition expanded to separated
plants and densely planted nursery ranges, the reconstruction retained the general plant arrangement
but progressively lost the continuity and separation of individual organs. Similar challenges have
been considered by Zermas et al. (2020), who reconstructed small groups of maize plants from UAV
and handheld RGB imagery and extracted traits including plant count, height, leaf area index, leaf
length, and leaf angle. Our results support the feasibility of reconstructing maize from aerial RGB
imagery, but also show why extending organ measurements to dense nursery populations remains
challenging. A point cloud may represent the overall arrangement of a nursery while still lacking the
geometry needed to distinguish individual leaves or measure their dimensions.
Thisdifferencesuggeststhattheusefulnessofareconstructionshouldbeevaluatedaccordingtothe
intended phenotype rather than through a single measure of point cloud quality. A reconstruction
containing incomplete leaf surfaces may still support plant localization, row identification, or
approximate canopy measurements, whereas estimating leaf area, leaf curvature, or the angle
between a leaf and the stem requires more complete and correctly associated organ geometry.
Similarly, a point cloud may be sufficient for identifying certain structures, while other measurements
may require a surface mesh or explicit organ segmentation. The appropriate representation therefore
depends on the biological question. Future phenotyping workflows should distinguish between
traits that can be extracted reliably from incomplete point clouds and those that require additional
reconstruction or validation.
4.2 Viewpoint Diversity, Plant Movement, and Reconstruction Completeness
One of the central findings of this study is that increasing the number of observations does not
necessarily produce a better reconstruction in every respect. In the equal frame comparison, the
single orbit and ribbon skirt configurations used the same number of images but distributed their
observations differently. The single orbit produced more points and fewer visible gaps within the plant
structure, although it also contained more scattered points outside the plant. These observations
show that point density, point isolation, and visible structural continuity describe different properties
of a reconstruction and should not be treated as interchangeable measures of quality.
The difference can be understood in terms of the information supplied by the camera viewpoints.
Repeated observations from a similar elevation can provide substantial overlap for surfaces already
visible from that trajectory, producing dense point coverage in those regions. However, additional
frames from similar viewpoints cannot recover surfaces that remain hidden. Changing camera
elevation or orientation can expose previously concealed structures, but distributing a fixed number
of frames across more viewpoints also changes the amount of overlap available within each viewing
region. Consequently, greater viewpoint diversity does not guarantee a denser or more continuous
point cloud. The balance between observing new surfaces and maintaining sufficient overlap for
reliable correspondence is particularly important for maize, where leaves are thin, curved, and
frequently overlap.
These findings are consistent with earlier work demonstrating the importance of viewing geometry
in crop reconstruction. Che et al. (2020) reported that nadir and oblique imagery could support
18

maize height estimation, while oblique observations provided advantages for representing canopy
layers in voxel based leaf area index estimation. Niu et al. (2024) also reported improvements in
maize reconstruction using oblique imagery compared with nadir observations, together with reduced
height accuracy at coarser spatial resolution. Our experiments extend this discussion to individual
organ visibility and the spatial distribution of reconstructed points. Rather than identifying one
camera orientation as universally preferable, the results indicate that different viewpoints contribute
different information about the plant. Parallel views can reveal vertical plant profiles, while elevated
and oblique observations can expose upper leaves, tassels, and surfaces that are poorly visible from
lower positions.
The nursery range experiments further illustrate why the relationship between additional obser-
vations and reconstruction quality is not straightforward. The initial orbital reconstruction preserved
the row arrangement but contained substantial voids between overlapping plants. Adding serpentine
passes reduced these gaps by introducing observations from additional directions. Increasing the sam-
pling rate of the original orbit also reduced visible voids. However, both approaches produced greater
point dispersion around some plant structures. The additional observations therefore contributed
information about previously incomplete regions while also introducing inconsistencies associated
with changes in plant position during acquisition.
The curlicue reconstruction from the 2024 growing season provided a different outcome. Although
substantial voids remained, primarily on the soil surface, individual plants were clearly distinguishable
without visible duplicated or ghost plant structures, and the inspected soil surfaces showed no visible
double-layering. The reconstruction contained fewer points than the three earlier nursery range
reconstructions, suggesting that a larger point cloud is not necessarily associated with better
separation of neighboring plants. This observation also illustrates that the completeness of the
reconstructed ground surface and the ability to distinguish individual plants are different aspects of
reconstruction quality. However, because the curlicue acquisition was conducted in a different field
and growing season, the absence of visible plant duplication cannot be attributed to the trajectory
alone.
Unlike rigid objects, maize plants do not maintain exactly the same geometry throughout a
flight. Leaves bend, twist, and move independently, particularly near their tips and in younger
upper leaves. When the camera returns to a plant after several seconds, the same leaf may occupy
a different position. Structure from Motion and multiview stereo reconstruction generally assume
that corresponding observations describe a static scene. When this assumption is violated, the
reconstruction may contain duplicated leaf surfaces, ghost organs, or scattered points rather than a
single consistent representation. The observed artifacts therefore reflect a fundamental difficulty
in reconstructing living plants: increasing the number of viewpoints can improve visibility, but
acquiring those viewpoints over time also increases the opportunity for the plant to change shape.
This creates a trade-off between spatial coverage and temporal consistency. A longer acquisition
can observe more of the plant, but the resulting images may no longer describe one fixed plant
geometry. Conversely, a shorter acquisition may reduce differences in organ position while leaving
some surfaces insufficiently observed. The optimal strategy is therefore unlikely to depend on frame
count alone. It should consider camera geometry, overlap, acquisition duration, and the movement
of the target plant. Future experiments using synchronized cameras or repeated acquisitions under
measured wind conditions could help separate the effects of viewpoint diversity from those of temporal
movement. The present study identifies these competing factors, but does not independently quantify
their contributions.
19

4.3 Missing Geometry, Reconstruction Artifacts, and the Limits of Postprocess-
ing
The reconstruction failures observed in this study represent different types of information loss and
geometric inconsistency. Occlusion produces missing observations when leaves, stems, or soil surfaces
are hidden from the available viewpoints. Plant movement produces conflicting observations when the
same organ occupies different positions across frames. Incorrect image correspondence or registration
can produce duplicated plants or separated soil surfaces that do not correspond to the physical scene.
Although these failures may all appear as irregularities in a point cloud, they have different causes
and cannot necessarily be corrected using the same processing strategy.
The distinction is especially relevant when interpreting point cloud density. A densely recon-
structed region is not necessarily geometrically accurate, because some of its points may represent
moving organs, background surfaces, or incorrect correspondences. Conversely, a region with rela-
tively few points may correspond to a genuine thin structure, such as a tassel branch or leaf tip.
This limitation is consistent with findings from UAV LiDAR studies, where increasing point density
beyond a certain level did not necessarily improve crop height estimation (Luo et al., 2021). Although
LiDAR and image based reconstruction generate points through different processes, both demonstrate
why point count alone is insufficient for evaluating the usefulness of three-dimensional plant data. For
photogrammetric reconstructions, point density must be considered alongside structural continuity,
organ visibility, and geometric consistency.
Postprocessing introduces another challenge because removing unwanted points can also remove
biologically meaningful information. In our masking experiment, ExG- andLab*-based approaches
reduced portions of the background but also excluded plant structures that were not predominantly
green. The missing tassel and brace roots illustrate this problem. A green vegetation mask
implicitly assumes that the relevant plant structures have a particular color, yet maize organs vary
in pigmentation with developmental stage, genotype, environmental conditions, and tissue condition.
Senescent leaves, reproductive organs, and some phenotypic abnormalities may therefore be removed
together with soil or other background material. This is particularly problematic for phenotyping
studies because the removed structures may be the very traits researchers intend to measure.
The same consideration applies to surface reconstruction. Ball pivoting converted the dense
point clouds into triangular meshes and made some morphological structures easier to distinguish,
including silks, leaf midribs, and tassel branches. However, the resulting surfaces were constrained by
the available point cloud. Missing points could produce holes, while duplicated or displaced points
could lead to irregular surface connections. In the original reconstruction experiments, MeshLab
ball pivoting provided visually finer representations of certain plant structures than the alternative
surface reconstruction methods examined. Nevertheless, a visually smoother or more continuous
surface should not be interpreted as evidence of greater geometric accuracy without an independent
reference. Surface reconstruction can change how the available geometry is represented, but it cannot
establish the correct shape of an organ that was never adequately observed.
These results suggest that preprocessing and reconstruction should be evaluated with respect
to the biological information they preserve, not only the background points they remove or the
visual appearance of the final model. For example, excluding soil may simplify an isolated plant
reconstruction, but cropping too close to the ground can eliminate brace roots, which are relevant
to plant anchorage and lodging resistance. Similarly, filtering scattered points may improve the
appearance of a leaf while removing legitimate observations of thin or damaged tissue. A phenotyping
workflow therefore requires a distinction between artifacts and biological variation. This distinction
becomes especially important when reconstructing plants with unusual morphology, because an
automated procedure designed around typical plant appearance may remove precisely the structures
20

that distinguish an abnormal phenotype.
4.4 Toward a Public Dataset of Field Grown Plants Reconstructed from RGB
UAV Imagery
Beyond demonstrating the feasibility of reconstructing individual maize plants and nursery ranges,
these experiments provide a starting point for developing a publicly accessible dataset of three-
dimensional plant reconstructions derived from consumer UAV RGB imagery. Such a dataset would
address a different need from collections of images or point clouds acquired under controlled laboratory
conditions. Field grown plants experience changing illumination, wind, background variation,
mutual occlusion, and differences in growth stage and morphology. These conditions complicate
reconstruction, but they are also the conditions under which many agricultural phenotyping systems
must eventually operate. A dataset that preserves these challenges would provide a useful resource
for studying the robustness and limitations of reconstruction algorithms.
Existing plant reconstruction and segmentation studies demonstrate the value of structured
three-dimensional data for quantitative phenotyping. For example, the maize dataset examined by
Yang et al. (2024) supported learning stem and leaf segmentation across hundreds of plants, while
the studies by Li et al. (2022a,b) demonstrated the extraction of morphological information from
reconstructed maize. These efforts establish the importance of having plant representations that
can be segmented and measured. A complementary field dataset could focus on the acquisition
conditions and reconstruction variability encountered when those representations are generated from
freely flown UAV imagery. Its contribution would not simply be to provide additional point clouds,
but to connect the reconstructed geometry with the images, camera viewpoints, field conditions, and
processing decisions that produced it.
The experiments in this study illustrate why retaining the complete acquisition and processing
history would be important. The equal frame comparison showed that the same number of images can
produce point clouds with different structural coverage and point distributions. The nursery range
experiments showed that additional trajectories or frames can reduce visible voids while introducing
greater dispersion around moving leaves. The masking experiment demonstrated that preprocessing
can remove genuine plant structures before reconstruction begins. If only the final point clouds
were retained, much of the information needed to understand these differences would be lost. A
reproducible dataset should therefore include the original RGB images or video, frame selection
information, camera calibration and estimated poses, reconstruction parameters, intermediate point
clouds, surface meshes where available, and records of masking, cropping, and manual editing.
A useful dataset should also represent variation in the plants themselves. The present experiments
include isolated plants, separated plant tilings, and densely planted nursery rows, each presenting
different reconstruction challenges. Expanding this collection across genotypes, developmental
stages, planting densities, and environmental conditions would allow researchers to examine whether
reconstruction methods generalize beyond a small number of favorable examples. Including plants
with senescent leaves, damaged tissue, reproductive structures, or unusual morphology would
be particularly valuable. Such variation would help determine whether an algorithm preserves
biologically meaningful differences or preferentially reconstructs plants that conform to a typical
appearance.
An important distinction is needed between a dataset of reconstructed plants and a dataset
containing validated three-dimensional ground truth. The point clouds generated in this study
are reconstruction outputs, not independently verified measurements of the physical plants. They
could support algorithm development and comparisons of reconstruction behavior, but they should
not be treated as exact geometric references. Developing a benchmark for reconstruction accuracy
21

would require an additional validation component, such as calibrated reference objects, independent
measurements of selected organs, or higher fidelity reference scans. Repeated acquisitions of the same
plants could also help quantify reconstruction variability under different flight conditions. These
measurements would make it possible to distinguish errors caused by acquisition geometry, plant
movement, and the reconstruction algorithm itself.
The longer term opportunity is to connect image acquisition, three-dimensional reconstruction,
organ segmentation, and biological measurements within a common dataset. For example, a
reconstructed plant could be associated with annotations identifying stems, leaves, tassels, ears,
and brace roots, together with measurements of selected morphological traits. Such annotations
would support the development of algorithms that identify organs in incomplete point clouds and
evaluate which traits can be recovered reliably. The original images and camera poses would also
allow researchers to test alternative reconstruction methods on the same observations, rather than
comparing outputs generated from unrelated datasets. This would make it possible to investigate
whether improvements arise from acquisition design, reconstruction algorithms, segmentation, or
subsequent processing.
The present study does not yet provide the scale, replication, or independent geometric vali-
dation needed for such a benchmark. Its contribution is to identify the acquisition configurations,
reconstruction artifacts, and processing decisions that a field reconstruction dataset should represent
and document. The results also suggest that unsuccessful or incomplete reconstructions should
not automatically be excluded from the dataset. Point clouds containing ghost organs, missing
surfaces, or registration errors would provide useful examples for developing methods that detect
reconstruction failures and estimate uncertainty. Preserving both successful and unsuccessful cases
would make the collection more representative of field conditions and more informative for evaluating
algorithm robustness.
A publicly available dataset built around these principles could help bridge the gap between
controlled three-dimensional plant phenotyping and reconstruction from agricultural field imagery.
It would also provide a common basis for comparing future methods without requiring every
research group to repeat the same field acquisitions. The potential value of consumer UAV RGB
reconstruction therefore extends beyond producing visually recognizable models of individual plants.
With documented acquisition conditions, appropriate annotations, and independent validation, these
reconstructions could contribute to a reproducible framework for studying plant morphology and
developing field phenotyping methods.
4.5 Study Limitations and Future Validation
The results should be interpreted within the experimental scope of this study. The reconstruc-
tions were evaluated through visual inspection, quantitative point distribution measurements, and
image-based void analysis, but without independent three-dimensional ground truth or validated
measurements of individual organs.
The experiments also involved different plants, acquisition trajectories, sampling rates, and field
conditions, which were not varied independently in a fully controlled design. In particular, the
equal frame comparison represents one pair of reconstructions rather than replicated evidence of the
general superiority of either acquisition configuration. The observed differences therefore identify
practical reconstruction behavior under the conditions examined, but do not establish universal
relationships between camera geometry, frame count, and geometric accuracy.
The quantitative evaluation of visible reconstruction voids also has limitations. The void
percentages were calculated from selected rendered views rather than directly from the three-
dimensional geometry. Consequently, these measurements depend on the viewing angle, scene
22

boundaries, and the visibility of gaps in the point clouds. Although they provide a quantitative
description of visible reconstruction voids, they do not establish the geometric completeness or
accuracy of individual plant organs. Furthermore, the curlicue experiment was conducted in a
different field and growing season, so its void percentage provides a descriptive measure of that
reconstruction rather than a controlled comparison with the three earlier acquisition strategies.
Future work should prioritize repeated acquisitions under documented environmental conditions,
independent geometric measurements, and quantitative evaluation of organ completeness and re-
construction consistency. Controlled comparisons of camera orientation, flight altitude, temporal
sampling, and acquisition duration would help identify which factors most strongly affect the recon-
struction of different plant organs. Automated organ segmentation and uncertainty assessment would
then provide a basis for determining whether the resulting geometry supports reliable phenotypic
measurements. These steps are necessary to move from visually recognizable field reconstructions
toward validated three-dimensional plant data suitable for quantitative analysis.
5 Conclusions
Consumer UAV RGB video processed using COLMAP successfully recovered recognizable structures
of individual maize organs under field conditions, including leaves, tassels, ears, silks, and brace roots.
Reconstructions of separated plant groups and nursery rows further demonstrated the potential
to extend 3D reconstruction beyond individual plants, while revealing the increasing challenges of
occlusion and preserving fine structural detail. The comparison using an equal number of frames
demonstrated that different image sampling configurations produce distinct point density and
isolation characteristics even when the number of frames and source video remain constant. These
findings highlight the importance of considering temporal sampling and viewpoint distribution
together when designing UAV acquisitions for 3D plant reconstruction. The masking experiments
further showed that preprocessing decisions can affect the preservation of biologically important
structures, particularly non-green tissues and fine organs. Together, these findings provide practical
guidance for acquiring and processing UAV imagery for 3D reconstruction of maize under field
conditions. They also establish a foundation for future work on automated organ segmentation,
geometric validation, and the development of publicly accessible 3D maize datasets for computer
vision and agricultural research.
Conflict of Interest Statement
The authors declare that the research was conducted in the absence of any commercial or financial
relationships that could be construed as a potential conflict of interest.
Author Contributions
Conceptualization, methodology, validation, resources, and writing, D.E.K. and T.K.; software, data
curation, and visualization, D.E.K.; supervision, project administration, and funding acquisition,
T.K. All authors have read and agreed to the submitted version.
Funding
This research received no external funding.
23

Acknowledgments
We thank Chris Browne for field management and Tim Krock, Hadi AliAkbarpour, Kannappan
Palaniappan, Dexa Akbar, and Vinny for helpful discussions. We acknowledge support from the
Department of Electrical Engineering and Computer Science for D.E.K. and an anonymous gift in
aid of maize research.
Data Availability Statement
Videos and reconstructions associated with this study are available athttps://www.dropbox.com/
sh/x4ucxbbjtm6xbmu/AABYrR_V2fFeZQsPhHnB8v5_a?dl=0.
References
Adak, A., Murray, S. C., and Anderson, S. L. (2023). Temporal phenomic predictions from
unoccupied aerial systems can outperform genomic predictions.G3: Genes Genom. Genet.13,
jkac294. doi:10.1093/g3journal/jkac294
Aktar, R., Kharismawati, D. E., Palaniappan, K., Aliakbarpour, H., Bunyak, F., Stapleton, A. E.,
et al. (2020). Robust mosaicking of maize fields from aerial imagery.Appl. Plant Sci.8, e11387.
doi:10.1002/aps3.11387
Araus, J. L. and Cairns, J. E. (2014). Field high-throughput phenotyping: the new crop breeding
frontier.Trends Plant Sci.19, 52–61. doi:10.1016/j.tplants.2013.09.008
Bernardini, F., Mittleman, J., Rushmeier, H., Silva, C., and Taubin, G. (1999). The ball-pivoting
algorithm for surface reconstruction.IEEE Trans. Vis. Comp. Graph.5, 349–359. doi:10.1109/
2945.817351
Che, Y., Wang, Q., Xie, Z., Zhou, L., Li, S., Hui, F., et al. (2020). Estimation of maize plant
height and leaf area index dynamics using an unmanned aerial vehicle with oblique and nadir
photography.Ann. Bot.126, 765–773. doi:10.1093/aob/mcaa097
Cignoni, P., Callieri, M., Corsini, M., Dellepiane, M., Ganovelli, F., and Ranzuglia, G. (2008).
MeshLab: an open-source mesh processing tool. InEurographics Italian Chapter Conference,
eds. V. Scarano, R. De Chiara, and U. Erra (The Eurographics Association). doi:10.2312/
LocalChapterEvents/ItalChap/ItalianChapConf2008/129-136
Furbank, R. T. and Tester, M. (2011). Phenomics — technologies to relieve the phenotyping
bottleneck.Trends Plant Sci.16, 635–644
Gibbs, J. A., Pound, M. P., French, A. P., Wells, D. M., Murchie, E. H., and Pridmore, T. P. (2020).
Active vision and surface reconstruction for 3d plant shoot modelling.IEEE/ACM Trans. Comput.
Biol. Bioinform.17, 1907–1917. doi:10.1109/TCBB.2019.2896908
Gollob, C., Ritter, T., Kraßnitzer, R., Tockner, A., and Nothdurft, A. (2021). Measurement of forest
inventory parameters with Apple iPad Pro and integrated LiDAR technology.Remote Sens.13,
3129. doi:10.3390/rs13163129
24

Hu, P., Chapman, S. C., Wang, X., Potgieter, A., Duan, T., Jordan, D., et al. (2018). Estimation of
plant height using a high throughput phenotyping platform based on unmanned aerial vehicle and
self-calibration: Example for sorghum breeding.Eur. J. Agron.95, 24–32. doi:10.1016/j.eja.2018.
02.004
Hui, F., Zhu, J., Hu, P., Meng, L., Zhu, B., Guo, Y., et al. (2018). Image-based dynamic quantification
and high-accuracy 3D evaluation of canopy structure of plant populations.Ann. Bot.121, 1079–
1088. doi:10.1093/aob/mcy016
Kharismawati, D. E., Akbarpour, H. A., Aktar, R., Bunyak, F., Palaniappan, K., and Kazic, T.
(2020).CorNet: unsupervised deep homography estimation for agricultural aerial imagery. In
16th European Conference on Computer Vision 2020 (ECCV2020), eds. V. Ferrari, B. Fisher,
C. Schmid, and E. Trucco. 402–419. doi:10.1007/978-3-030-65414-6_28
Kharismawati, D. E., Dhakal, S., McCusker, C. E., Wilson, J. R., Ohlson, E. W., and Khanal, S.
(2026). When ground-truth fidelity matters: An orchestrated uas framework for wheat streak
mosaic virus detection using vision transformers and machine learning.arXivdoi:10.48550/arXiv.
2609.12169
Kharismawati, D. E. and Kazic, T. (2025a).DroneZaic: a robust end-to-end pipeline for mosaicking
freely flown aerial video of agricultural fields.The Plant Phenom. J.8, e70033. doi:10.1002/ppj2.
70033
Kharismawati, D. E. and Kazic, T. (2025b).MaizeStandCounting (MaSC): Automated and accurate
maize stand counting from uav imagery using image processing and deep learning.arXivdoi:
10.48550/arXiv.2510.07580
Kharismawati, D. E. and Kazic, T. (2026).Maize Seedling Detection Dataset (MSDD): A Curated
High-Resolution RGB Dataset for Seedling Maize Detection and Benchmarking with YOLOv9,
YOLO11, YOLOv12 and Faster-RCNN.Agronomy16, 1605. doi:10.3390/agronomy16161605
Li, Y., Liu, J., Zhang, B., Wang, Y., Jingfa, J. Y., Zhang, X., et al. (2022a). Three-dimensional
reconstruction and phenotype measurement of maize seedlings based on multi-view image sequences.
Fron. Pl. Sci.13. doi:10.3389/fpls.2022.974339
Li, Y., Wen, W., Miao, T., Wu, S., Yu, Z., Wang, X., et al. (2022b). Automatic organ-level point
cloud segmentation of maize shoots by integrating high-throughput data acquisition and deep
learning.Comput. Electron. Ag.193, 106702. doi:10.1016/j.compag.2022.106702
Liebisch, F., Kirchgessner, N., Schneider, D., Walter, A., and Hund, A. (2015). Remote, aerial
phenotyping of maize traits with a mobile multi-sensor approach.Plant Meth.11, 9. doi:
10.1186/s13007-015-0048-8
Lin, Y. and Habib, A. (2021). Quality control and crop characterization framework for multi-
temporal UAV LiDAR data over mechanized agricultural fields.Remote Sens. Env.256, 112299.
doi:10.1016/j.rse.2021.112299
Liu, Y., Cen, C., Che, Y., Ke, R., Ma, Y., and Ma, Y. (2020). Detection of maize tassels from UAV
RGB imagery with faster R-CNN.Remote Sens.12, 338. doi:10.3390/rs12020338
Luo, S., Liu, W., Zhang, Y., Wang, C., Xi, X., Nie, S., et al. (2021). Maize and soybean heights
estimation from unmanned aerial vehicle (UAV) LiDAR data.Comput. Electron. Ag.182, 106005.
doi:10.1016/j.compag.2021.106005
25

Minervini, M., Scharr, H., and Tsaftaris, S. A. (2015). Image analysis: the new bottleneck in plant
phenotyping.IEEE Sig. Proc. Mag.32, 126–131. doi:10.1109/MSP.2015.2405111
Niu, Y., Han, W., Zhang, H., Zhang, L., and Chen, H. (2024). Estimating maize plant height
using a crop surface model constructed from uav rgb images.BioSystems Eng.241, 56–67.
doi:10.1016/j.biosystemseng.2024.04.003
Schönberger, J.L.andFrahm, J.(2016). Structure-from-Motionrevisited. InCVPR 2016: Proceedings
of the 2016 IEEE Conference on Computer Vision and Pattern Recognition(New York: IEEE),
4104–4113. doi:10.1109/CVPR.2016.445
Schönberger, J. L. and Frahm, J. (2022–present).COLMAP(GitHub)
Schönberger, J. L., Zheng, E., Pollefeys, M., and Frahm, J. (2016). Pixelwise view selection for
unstructured Multi-View Stereo. InEuropean Conference on Computer Vision (ECCV). 501–518
Sweet, D. D., Tirado, S. B., Springer, N. M., Hirsch, C. N., and Hirsch, C. D. (2022). Opportunities
and challenges in phenotyping row crops using drone-based rgb imaging.The Plant Phenom. J.5,
e20044. doi:10.1002/ppj2.20044
Velumani, K., Lopez-Lozano, R., Madec, S., Guo, W., Gillet, J., Comar, A., et al. (2021). Estimates
of maize plant density from UAV RGB images using faster-RCNN detection model: Impact of the
spatial resolution.Plant Phe.2021, 1–16. doi:10.34133/2021/9824843
Wang, X., Zhang, R., Song, W., Han, L., Liu, X., Sun, X., et al. (2019). Dynamic plant height QTL
revealed in maize through remote sensing phenotyping using a high-throughput unmanned aerial
vehicle (UAV).Sci. Rep.9, 3458. doi:10.1038/s41598-019-39448-z
Wu, S., Wen, W., Gou, W., Lu, X., Zhang, W., Zheng, C., et al. (2022). A miniaturized phenotyping
platform for individual plants using multi-view stereo 3d reconstruction.Fron. Pl. Sci.13, 897746.
doi:10.3389/fpls.2022.897746
Yang, X., Miao, T., Tian, X., Wang, D., Zhao, J., Lin, L., et al. (2024). Maize stem–leaf segmentation
framework based on deformable point clouds.ISPRS J. Photogram. Rem. Sens.211, 49–66. doi:
10.1016/j.isprsjprs.2024.03.025
Yin, T., Qi, J., Cook, B. D., Morton, D. C., Wei, S., and Gastellu-Etchegorry, J. (2020). Modeling
small-footprint airborne Lidar-derived estimates of gap probability and leaf area index.Remote
Sens.12, 4. doi:10.3390/rs12010004
Zermas, D., Morellas, V., Mulla, D., and Papanikolopoulos, N. (2020). 3d model processing for
high throughput phenotype extraction – the case of corn.Comput. Electron. Ag.172, 105047.
doi:10.1016/j.compag.2019.105047
Zhu, R., Sun, K., Yan, Z., Yan, X., Yu, J., Shi, J., et al. (2020a). Analysing the phenotype
development of soybean plants using low-cost 3D reconstruction.Sci. Rep.10, 7055. doi:
10.1038/s41598-020-63720-2
Zhu, W., Sun, Z., Yang, T., Li, J., Peng, J., Zhu, K., et al. (2020b). Estimating leaf chlorophyll
content of crops via optimal unmanned aerial vehicle hyperspectral data at multi-scales.Comput.
Electron. Ag.178, 105786. doi:10.1016/j.compag.2020.105786
26

论文原文与出版信息

arXiv (2026)
查看论文原文

继续探索