论文信息
论文解读
AI 辅助整理- 方法
- 0.5米NAIP影像+GAN合成CHM,对比分水岭、RF、SVM与Mask R-
- 发现
- Mask R-CNN检测F1达0.908,深度学习较传统方法误差降低23.9%,合成CHM与LiDA
- 机会点
- 可探索GAN合成高程数据在其他树种、区域及多时相场景的泛化能力与不确定性评估。
论文摘要
俄勒冈州东部入侵树种(如西部刺柏,Juniperus occidentalis)的扩张改变了水文情势,并威胁艾草松鸡栖息地。在资源稀疏且空间分布不均的条件下实现高效恢复,需要准确的单木清查数据,但传统地面调查难以适用于大尺度景观。我们提出了一种单木树冠检测与高度估测的遥感框架,将0.5 m多光谱影像(NAIP)与合成生成的冠层高度信息相结合。我们比较了经典分割方法与机器学习方法,即二维和三维多尺度分水岭分割(MSWD)、随机森林(RF)和支持向量机(SVM),与深度学习目标检测器,特别是Faster R-CNN、YOLOv8和Mask R-CNN。为在缺乏LiDAR数据的区域获取结构属性,采用生成对抗网络(GAN)由红、近红外和山体阴影波段合成冠层高度模型(CHM)。Kolmogorov–Smirnov检验表明,GAN生成的冠层高度模型与LiDAR-derived冠层高度模型之间无显著差异。深度学习模型优于经典方法,识别误差最多降低23.9%。Mask R-CNN取得最高检测性能(F1 = 0.908),其次为YOLOv8和Faster R-CNN。此外,三维MSWD优于二维MSWD,证实了高度数据的关键作用。研究结果表明,将二维多光谱影像与GAN合成的高程数据相结合,可生成准确且具有空间显式性的单木分布图。通过消除对同步LiDAR采集的需求,该方法降低了基线数据收集的经济门槛,为土地管理者提供了可扩展且成本高效的牧场清查与恢复规划工具。
全文
Frontiers in Forests and Global ChangeFrontiers in Forests and Global Change 01 frontiersin.org
Fusing multispectral and elevation
data for tree-level landscape
inventory in Central Oregon, USA
Robert Schriver 1, John Sessions 1, Mihaela Paun 2,3 and
Bogdan M. Strimbu 1*
1Department of Forest Engineering, Resources, and Management, College of Forestry, Oregon State
University, Corvallis, OR, United States, 2National Institute of Research and Development for Biological
Sciences, Bucharest, Romania, 3Universitatea Din Bucuresti, Facultatea De Administratie Si Afaceri,
Bucharest, Romania
Expansion of invasive tree species in eastern Oregon, like Western juniper
(Juniperus occidentalis), alters hydrological regimes and threatens sage-grouse
habitat. Efficient restoration on sparse spatially distributed resources requires
accurate individual-tree inventories, but conventional ground surveys are infea-
sible for large landscapes. We present a remote sensing framework for individual
tree crown detection and height estimation that pairs 0.5 m multispectral imagery
(NAIP) with synthetically generated canopy height information. We compared
classical segmentation and machine learning methods, namely 2D and 3D multi-
scale watershed segmentation (MSWD), Random Forest (RF), and Support Vector
Machine (SVM), against deep learning object detectors, specifically Faster R-CNN,
YOLOv8, and Mask R-CNN. To capture structural attributes where LiDAR is absent,
a Generative Adversarial Network (GAN) synthesized Canopy Height Models
(CHMs) from red, near-infrared, and hillshade bands. Kolmogorov–Smirnov test
showed no significant difference between GAN-generated and LiDAR-derived
Canopy Height Models (CHMs). Deep learning models outperformed classical
approaches, reducing identification errors by up to 23.9%. Mask R-CNN achieved
the highest detection performance ( =1 0.908F ), followed by YOLOv8 and Faster
R-CNN. Additionally, 3D MSWD surpassed 2D MSWD, confirming the critical
role of height data. Our findings demonstrate that combining 2D multispectral
imagery with GAN-synthesized elevation data yields accurate, spatially explicit
individual-tree maps. By eliminating the need for concurrent LiDAR acquisitions,
this approach lowers financial barriers to baseline data collection, providing land
managers with a scalable, cost-effective tool for rangeland inventory and resto-
ration planning.
KEYWORDS
digital elevation models, generative adversarial network, National Agriculture Imagery
Program, region-based convolutional neural network, tree segmentation, YOLOv8
1 Introduction
Over the past 100 years, western juniper (Juniperus occidentalis) has spread across much
of eastern Oregon, converting large expanses of sagebrush-steppe into increasingly tree-dom-
inated landscapes (Azuma et al., 2005). This shift carries real ecological costs, as juniper expan-
sion is linked to altered hydrology ( Jeppesen, 1977), and degraded habitat for species that
depend on open sagebrush, most notably the greater sage-grouse ( Miller and Rose, 1995 ;
Gedney et al., 1999 ; Baruch-Mordo et al., 2013 ). Consequently, controlling juniper
TYPE Original Research
PUBLISHED 16 November 2026
DOI 10.3389/ffgc.2026.1954252
OPEN ACCESS
EDITED BY
Avram Sorin,
University of Craiova, Romania
REVIEWED BY
Yunhong Xie,
Beijing Forestry University, China
Yanan Liu,
Beijing University of Civil Engineering
and Architecture, China
*CORRESPONDENCE
Bogdan M. Strimbu
Bogdan.Strimbu@oregonstate.edu
RECEIVED 31 July 2026
REVISED 02 October 2026
ACCEPTED 28 October 2026
PUBLISHED 16 November 2026
CITATION
Schriver R, Sessions J, Paun M and
Strimbu BC (2026) Fusing multispectral
and elevation data for tree-level
landscape inventory in Central Oregon,
USA.
Front. For. Glob. Change 9:1954252.
doi: 10.3389/ffgc.2026.1954252
COPYRIGHT
© 2026 Schriver, Sessions, Paun and
Strimbu. This is an open-access article
distributed under the terms of the
Creative Commons Attribution License
(CC BY). The use, distribution or
reproduction in other forums is
permitted, provided the original
author(s) and the copyright owner(s) are
credited and that the original publication
in this journal is cited, in accordance
with accepted academic practice. No
use, distribution or reproduction is
permitted which does not comply with
these terms.
Schriver et al. 10.3389/ffgc.2026.1954252
Frontiers in Forests and Global Change 02 frontiersin.org
encroachment now ranks among the primary restoration goals for
land managers across the region.
Restoration programs at the landscape level cannot be planned
effectively without knowing where individual trees stand, how many
there are, and how large they have grown. However, field inventory
(Avery and Burkhart, 2015; Kershaw et al., 2017) is not cost-effective for
western juniper, since the species is usually scattered across large areas
and carries little commercial value. Remote sensing sidesteps this prob-
lem, as it can characterize inventory conditions over broad landscapes
at a fraction of the cost of ground-based measurement (Noone et al.,
2013; Nielsen et al., 2014). The last two decades saw a rapid expansion
of remote sensing techniques in forest assessments, and public geospa-
tial data have become more accessible. The National Agriculture
Imagery Program (NAIP) of the USA, for instance, provides high-res-
olution multispectral coverage across the country on a recurring basis.
Lidar acquisitions, while still incomplete, are expanding through federal
and state mapping efforts to provide accurate description of the ground
and top canopy. Furthermore, photogrammetric efforts create surface
models suitable for forest inventory. Together, spectral and elevation
data supplies the information to identify and characterize individual
trees at scale (Ke and Quackenbush, 2011; Archuleta et al., 2017). The
general remote sensing strategies extracting forest attributes are either
based on imagery or from point clouds. The approach based on two-
dimensional raster imagery is widely available and computationally
straightforward (Camps-Valls et al., 2014) but sensitive to resolution
and scale. The approach using point clouds from LiDAR or photogram-
metry captures canopy structure more directly and typically yield more
accurate tree-level estimates but at higher acquisition and processing
cost (Næsset, 2002; Zhen et al., 2016; Y ou et al., 2020).
Delineating individual crowns accurately is a prerequisite for most
tree-level inventory systems (Ayrey et al., 2017), and the methods avail-
able for doing so fall broadly into two camps. Rule-based segmentation
relies on properties such as crown geometry, local intensity maxima, or
spectral gradients to separate a tree from its neighbors, while machine-
learning approaches instead learn the relevant distinguishing features
from labeled examples. The rise of deep learning in particular has
pushed object-detection accuracy well beyond what earlier rule-based
methods could achieve, and architectures such as Faster R-CNN, YOLO,
and Mask R-CNN have since been adopted widely for forestry applica-
tions (Wang et al., 2004; Li et al., 2012; Bengio et al., 2013; Wegner et al.,
2016; He et al., 2015 , 2017: Nikitina, 2024; Pratama et al., 2025 ;
Alegbeleye et al., 2026).
Irrespective of the procedure used to identify individual trees (i.e.,
rule based or machine learning based), the canopy heigh model
(CHM) was identified as providing significant discriminating power.
If point cloud data is available, a CHM is easily obtained. However, the
absence of point clouds requires usage of imagery. Generating a CHM
from an optical image belongs to a broader class of problems known
as monocular height estimation, which has drawn considerable inter-
est within the remote sensing community. Mou and Zhu (2018) tack-
led it with IM2HEIGHT, a fully residual convolutional-deconvolutional
network trained to map a monocular image directly to a height sur-
face, while Ghamisi and Y okoya (2018) took a generative approach
with IMG2DSM, recovering a digital surface model from single imag-
ery using a conditional Generative Adversarial Network (GAN) built
on the pix2pix architecture (Isola et al., 2017). Amirkolaee and Arefi
(2019) reported similar accuracy with a deep convolutional encoder-
decoder design. Nearly all of this work, however, was developed and
tested on dense urban scenes, where tall buildings create sharp,
high-contrast vertical edges that the networks can exploit. It remains
an open question whether the same image-to-elevation translation
holds up in sparse, low-contrast arid woodland, where the target
objects are small, irregularly shaped, and spectrally close to the sur-
rounding ground. This gap is the starting point for the present study,
which further departs from prior work by evaluating the synthesized
elevation surface not as an end in itself but by how well it supports a
downstream task, specifically individual tree detection.
Crown size and canopy cover in western juniper vary substantially
across eastern Oregon (Gedney et al., 1999), yet individual tree crown
detection has still been shown to perform reasonably well even under
such sparse-canopy conditions when the goal is extracting canopy
metrics (Skurikhin et al., 2013). This study builds directly on an earlier
landscape-scale inventory of individual juniper (Schriver, 2018), and
extends it by considering two approaches, one based on neural net-
works, and one based on a combination of segmentation and classi-
fiers. The overarching aim is a method that can estimate juniper
location and height directly from remotely sensed imagery accurately
and cheaply enough, to make landscape-scale surveying practical.
2 Methods
2.1 Study area
This study was carried out across the Cherry Creek, Bridge Creek,
and Mountain Creek watersheds of central Oregon ( Figure 1), the
same landscape examined in an earlier individual-juniper inventory
(Schriver, 2018). The watersheds span roughly 150,000 ha, dominated
by rangeland and open juniper woodland, with mixed stands of pon-
derosa pine (Pinus ponderosa L.). Given the sheer size of this land-
scape and the variability in vegetation density and terrain across it,
conventional field inventories are impractical, which is the gap remote
sensing approaches are positioned to fill (North Slope Ochoco Holistic
Restoration Project; Gedney et al., 1999).
2.2 Remotely sensed data
2.2.1 Raw data
Imagery for this study came from the 2009 NAIP acquisition, cap-
tured at 0.5-m resolution across four spectral bands: red, green, blue,
and near-infrared. Preliminary testing showed that the red and near-
infrared bands carried the bulk of the useful signal for identifying
vegetation and separating woody cover from other land-cover types,
so these two bands were carried forward into all subsequent analyses
(Ke and Quackenbush, 2011; Bannari et al., 1995).
A hillshade layer, derived from a U. S. Geological Survey digi -
tal terrain model based on a LiDAR acquisition from 2011 and
2013, was added to provide environmental context. The LiDAR
acquisition was executed with a Leica ALS 60, flown at an altitude
of approximately 900 m with a side overlap of 60%. The point
cloud that served as input for the creation of the hillshade layer
has a density of 8 points/m 2, with a vertical accuracy of 2.0 cm
and 1.5 horizontal. The point clouds and derived products, like
digital terrain model, were processed in TerraScan. Slope, aspect,
and elevation have each been linked to where western juniper and
ponderosa pine establishes and how its canopy develops ( Davies
Schriver et al. 10.3389/ffgc.2026.1954252
Frontiers in Forests and Global Change 03 frontiersin.org
et al., 2010 ; Archuleta et al., 2017 ). Therefore, the hillshade prod -
uct was used as a proxy for local terrain and combined with the
spectral bands throughout both the canopy-height estimation
and tree-detection stages of the analysis. The NAIP imagery
(2009) and the LiDAR acquisitions (2011 and 2013) were sepa -
rated by two to four years. Over this interval, height growth of
western juniper, a slow-growing species on these sites ( Miller et
al., 2005 ), is expected to be small relative to the 1.8 m height
threshold used in our analysis. Ponderosa pine grows faster, so
growth-related discrepancies are likely larger in pine-dominated
areas. Growth is likely the only source of change, as no harvest or
wildlife occurred in the region, and the natural mortality was
minimal as no extreme meteorological events were present
during that period.
2.2.2 Preprocessed data
All of the predictive models in this study take 8-bit, three-band
rasters as input, and arriving at that three-band configuration required
testing which of the four NAIP bands actually contributed useful
signal. We compared six combinations, namely (1) R + G + B
(Figure 2a ), (2) R + NIR, (3) R + G + NIR, (4) R + B + NIR, (5)
G + B + NIR, and (6) R + G + B + NIR, by training a preliminary
Faster R-CNN model on each for 5,000 steps and scoring it against a
10% held-out subset; R + NIR gave the highest F1 and was carried
forward, consistent with the vegetation-index literature (Bannari et al.,
1995), while blue and green added no measurable benefit. Because
rainfall and terrain also help explain structural variation in western
juniper (Davies et al., 2010), and because only one band slot remained,
FIGURE 1
Study area overlaid on the NAIP imagery.
FIGURE 2
Example hillshade raster illustrating complex terrain. (a) True-color NAIP imagery, (b) False-color composite displaying red, near-infrared, and hillshade
bands for a portion of the same image.
Schriver et al. 10.3389/ffgc.2026.1954252
Frontiers in Forests and Global Change 04 frontiersin.org
we filled it with a hillshade layer, which encodes aspect and terrain
consistently across the landscape. The 10-m hillshade was resampled
to the 0.5-m imagery grid by nearest-neighbor interpolation and
stacked with red as band one, near infrared as band two, and hillshade
as band three (Figure 2b).
When the angle between sensor and object is large enough, tall
objects such as trees appear to lean away from the sensor in the image,
an effect known as relief displacement (Figure 3); LiDAR is largely
immune to this because it records the true position of each return
rather than a projected view. Left uncorrected, this displacement can
shift a tree’s apparent position in the imagery relative to its true loca-
tion in LiDAR-derived products such as our CHM. We therefore
aligned the two data sources following Safdarinezhad et al. (2017),
under the assumption that, since our surface model and imagery are
both dominated by vegetation, correct alignment should maximize the
overlap between above-ground elevation pixels and mapped vegeta-
tion. Applying this criterion produced consistent alignment between
the NAIP and LiDAR-derived layers across both study areas. However,
because this criterion assumes that vegetation is the same in both
datasets, trees removed or killed between acquisitions would reduce
the achievable overlap. Nevertheless, given that such changes affect a
small fraction of trees, they are unlikely to bias the estimated offset,
which is determined by the large majority of unchanged crowns.
2.3 Tree crown identification
The processing architecture contains two pipelines who share a
common data acquisition and preprocessing stage (Figure 4). Raw
inputs are NAIP four-band imagery (0.5 m resolution) and a digital
terrain model (DTM) obtained from the USGS at 10 m resolution or
from LiDAR at 1 m where available. Preprocessing fuses the red (R)
and near-infrared (NIR) bands with a hillshade raster derived from
the DTM, producing a three-band raster stack at 0.5 m resolution. The
hillshade component encodes aspect and local terrain slope (Davies
et al., 2010 ), improving discrimination between juniper and pines
trees and shadow or low-canopy areas on complex terrain. This three-
band raster serves as input to the GAN and to both downstream
detection models.
A Generative Adversarial Network (Isola et al., 2017) occupies a
central position in both pipelines. The Generative Adversarial
Network (GAN) is trained on paired raster-to-CHM chips derived
from areas where LiDAR data are available (dashed Lidar CHM box
in Figure 1), learning to translate the three-band raster into a synthetic
canopy height model (CHM). During inference, the GAN generates a
CHM for any tile covered by NAIP imagery, regardless of whether
LiDAR data exists for that location. This is the key step that extends
both downstream detection methods to the full 150,000 ha study area,
of which most lacks LiDAR coverage.
Pipeline A uses the CHM and three-band raster as joint inputs
to a multi-scale watershed segmentation algorithm adapted from
Wang et al. (2004). The algorithm first creates binary masks from a
normalized difference vegetation index (NDVI) and a CHM height,
isolating woody vegetation. A circular maxima filter identifies tree -
top candidates in the CHM; a pixel-wise majority filter retains only
those candidates that are local maxima at two or more scales, sup -
pressing within-crown noise. Watershed segmentation seeded at
these treetops delineates crown boundaries. The output is a set of
georeferenced crown polygons; each associated with an estimated
tree location (centroid) and tree height (maximum CHM value
within the polygon).
Pipeline B feeds the same three-band raster, but with the CHM
replacing the hillshade, into three deep learning object detectors: a
Faster Region-based Convolutional Neural Network ( Ren et al.,
2017), a YOLOv8 (Redmon et al., 2016 , Jocher et al., 2023 ), and a
Mask Region-based Convolutional Neural Network (He et al., 2017).
Faster R-CNN employs a two-stage architecture with a ResNet-50
backbone, generating bounding box predictions with location coor-
dinates and confidence scores. Mask R-CNN builds upon this two-
stage paradigm by adding a parallel mask prediction branch, yielding
both bounding boxes and pixel-wise crown masks that delineate loca-
tion and crown area. YOLOv8, by comparison, adopts a one-stage,
anchor-free design that directly outputs bounding boxes with confi-
dence estimates.
Computations were executed on a workstation equipped with an
AMD Ryzen 95950× 16-core CPU, installed on a MSI MAG B550
Tomahawk Max motherboard. The system has 128 GB DDR4 RAM, a
FIGURE 3
Displacement of image objects caused by non-nadir view.
Schriver et al. 10.3389/ffgc.2026.1954252
Frontiers in Forests and Global Change 05 frontiersin.org
NVIDIA GeForce RTX 3070 Ti graphics card with 8 GB GDDR6X
memory, and two SSD drives, the system drive (i.e., C-drive) 2 TB and a
work drive (i.e., D drive) of 8 TB. The cost of producing the system (parts
only), at the time of acquisition (i.e., 2024) was approximately 3,500 USD.
Pipeline A is a classical, rule-based segmentation method that
does not rely on deep learning, making it computationally efficient
and interpretable. Pipeline B leverages deep learning to learn hierar-
chical features directly from the input rasters, potentially offering
FIGURE 4
Methodological framework used for individual tree detection.
Schriver et al. 10.3389/ffgc.2026.1954252
Frontiers in Forests and Global Change 06 frontiersin.org
greater robustness to variations in crown size, shape, and canopy
density compared to the rule-based watershed approach. The outputs
from all detectors enable a comprehensive comparison of detection
accuracy, localization precision, and delineation quality.
2.3.1 Canopy height models obtained with
generative adversarial network
For portions of the study area where LiDAR data were unavail -
able, CHMs were estimated using a Generative Adversarial
Network (GAN), with multispectral NAIP imagery serving as the
only predictor. Estimating canopy height from spectral informa -
tion represents a paired image-to-image translation problem, an
application for which GAN architectures have proven particularly
effective ( Isola et al., 2017 ). Consequently, we adopted the pix2pix
framework ( Isola et al., 2017 ) and customized the original archi -
tecture by configuring the U-Net generator and PatchGAN dis -
criminator to accept three input bands (red, near-infrared, and
hillshade) and produce a single-band CHM output.
Training data were created by extracting paired image and
CHM tiles measuring 256 × 256 pixels from areas with LiDAR cov -
erage. This process yielded approximately 50,000 paired samples,
of which 70% were allocated to model training and the remaining
30% were reserved for evaluation. The measures used to assess the
procedures using the 70%–30% split (e.g., Kolmogorov–Smirnov
test and the root mean square error) should be interpreted as an
assessment of the GAN’s ability to reproduce the height distribu -
tion under this chip-random split, rather than as a strictly spatially
independent validation.
Because most pixels in the CHM represent ground or low vegeta-
tion and therefore have values close to zero, we implemented a modi-
fied L1 loss function ( Equation 1 ) that distinguishes between
low-height and tree-height pixels. A threshold of 1.8 m, derived from
field observations, was used to separate the two classes:
<= −≥ truetrue predicted true0 if 1.8 mModified L Loss if 1.8 myyy y1 (1)
The selected threshold is supported by both empirical field mea-
surements and the observed height-frequency distribution (Figure 5),
where the abundance of canopy heights begins to increase noticeably.
In addition, analyses of crown dimensions indicated that this value
represents a meaningful transition point between low vegetation and
established trees. Given that nearly all non-zero CHM values in the
study area are associated with woody vegetation, the 1.8 m threshold
provides an effective criterion for distinguishing trees from bare
ground and herbaceous cover (Isola et al., 2017).
Each band was independently rescaled by linear stretch to the
band’s full dynamic range, so that reflectance (R, NIR), hillshade, and
CHM values occupied comparable numeric ranges despite their dif-
ferent physical units.
2.3.2 Multi-scale watershed detection, random
forests and support vector machine
Many methods were proven successful in identification of indi -
vidual tree from remotely sensed data ( Chen et al., 2006 ;
Jakubowski et al., 2013 ; Längkvist et al., 2016 ; Strand et al., 2006 ;
Y aseen, 2024), such as random forest, support vector machines,
watershed, or even maximum likelihood. For this study we focused
on three methods which are routinely used as a start in many tree-
level investigations: an enhanced a multi-scale watershed segmen -
tation algorithm adapted from Wang et al. (2004) , random forest,
as implemented by Riddell et al. (2020) , and support vector
machine, as proposed by Ballanti et al. (2016) .
The enhanced multi-scale watershed segmentation procedure
integrates the generated CHM with NAIP imagery through six
sequential processing stages ( Figure 6). The NDVI threshold was
determined through a systematic evaluation of candidate values
ranging from 0.2 to 0.8 at intervals of 0.1. Each threshold was tested
using a held-out dataset of 30 image chips, and performance was
assessed based on omission error (false negatives divided by total
positives) and commission error (false positives divided by total
detections) generated by the multi-scale watershed algorithm.
Among the evaluated values, an NDVI threshold of 0.6 produced the
lowest combined omission and commission error (0.42), outperform-
ing thresholds of 0.5 and 0.7, which yielded total errors of 0.49 and
0.51, respectively. Furthermore, receiver operating characteristic
analysis identified 0.6 as the optimal threshold on the precision–
recall curve for the study dataset. To further reduce false detections,
a height mask derived from the CHM was applied, retaining only
features exceeding 1.8 m and thereby restricting detections to woody
vegetation.
2.3.3 Faster R-CNN
Convolutional object detectors are now well established for gen-
eral-purpose object recognition (Ren et al., 2017) and have carried
that success over into remote sensing applications (Tang et al., 2017).
FIGURE 5
Truncated CHM histogram.
Schriver et al. 10.3389/ffgc.2026.1954252
Frontiers in Forests and Global Change 07 frontiersin.org
We accordingly identified tree crowns in the orthophotos with the
Faster R-CNN implementation available in the TensorFlow Object
Detection Framework ( Huang et al., 2017 ), built on a ResNet-50
backbone pre-trained on ImageNet (Deng et al., 2009) and fine-tuned
on our tree crown dataset via stochastic gradient descent (learning
rate 0.003, momentum 0.9, weight decay 0.0005) for 20,000 steps with
FIGURE 6
Improved multi-scale watershed segmentation.
Schriver et al. 10.3389/ffgc.2026.1954252
Frontiers in Forests and Global Change 08 frontiersin.org
a batch size of 1 on a single NVIDIA Tesla P100 GPU. The network
architecture itself was left unmodified; only the anchor box configu-
ration and confidence threshold were adapted to the dataset.
Ground truth for training and testing was built by manually
interpreting the NAIP imagery, starting from a random sample of
2,500 m2 (50 m × 50 m) chips. This chip footprint reflects a balance
among three constraints: (1) it comfortably captures 5–10 trees given
the study area’s typical stand density of roughly 3–6 trees per
2,500 m2; (2) it fits the TensorFlow Object Detection Framework’s
standard 600 × 600-pixel input, since a 50 × 50 m chip at 0.5 m/pixel
yields a 100 × 100-pixel image that can be padded to 600 × 600; and
(3) a sensitivity check against 25 × 25 m and 75 × 75 m chips changed
F1 by less than 2%, supporting the robustness of this choice. A
bounding box was drawn for every tree whose crown fell more than
half within a chip, and boxes smaller than 9 m2 were excluded, since
a crown that size corresponds to a tree below the 1.8 m height thresh-
old. That 9 m2 cutoff follows from field-measured allometry: juniper
and ponderosa pine crowns in east Oregon are approximately circu-
lar, and at 0.5 m/pixel the smallest reliably detectable crown diameter
is about 3.4 m (a 7-pixel span), which works out to roughly 9 m 2 of
projected area. The 9 m2 area is empirically proven as the crown size
associated with a 1.8 m tree in this population and consistent with
published allometric tables for the species (Gedney et al., 1999). We
digitized all boxes in LabelImg (Tzutalin, 2015), ultimately process-
ing 750 chips into 8,111 individual tree annotations.
Following Li et al. (2017) and Liu et al. 2016, the annotated boxes
were divided randomly into a 70% training set and a 30% test set, with
no intermediate validation partition. We considered this justified
because (1) the 70% training partition already contained 5,677 annota-
tions across 525 chips, enough for stable gradient estimates; (2) hyper-
parameters such as learning rate, anchor sizes, and the NMS threshold
were tuned by 5-fold cross-validation inside the training partition rather
than against a held-out validation set; and (3) the 30% test partition was
reserved exclusively for final metric reporting, avoiding any leakage, a
protocol consistent with standard practice in remote-sensing object
detection (Tang et al., 2017), though a three-way split would give a
lower-variance estimate of generalization error and is worth adopting
in future work with larger datasets. We expanded the effective training
set through geometric augmentation: because nadir imagery is rotation-
invariant (Cheng et al., 2016), images could be rotated up to 270° and
flipped along either axis without compromising the validity of the
ground truth, and the TensorFlow Object Detection Framework was
further configured to randomly crop and pad images back to their origi-
nal size.
Augmentation of this kind is a well-documented remedy for overfit-
ting on small datasets (Ronneberger et al., 2015), and although our 750
chips considerably exceed the roughly 40 images used in the original
U-Net study (Ronneberger et al., 2015), the single-class (‘tree’) design
and the wide intra-class variability of juniper crowns across seasons and
terrain still leave room for augmentation to help. An ablation comparing
models with and without augmentation bore this out, with a 4.3% gain
in F1 when augmentation was applied.
Because anchor box size and aspect ratio strongly influence Faster
R-CNN’s accuracy on small objects (Eggert et al., 2017), we derived both
from a histogram analysis of the side lengths and aspect ratios of the
8,111 digitized crowns. Crown side lengths peaked between 4 and 6 m
(8–12 pixels at 0.5 m/pixel), spanned 3.4–18 m overall, and stayed below
12 m for 95% of trees; from this we set six anchor sizes, namely 32, 48,
64, 96, 128, and 192 pixels, equivalent to 16, 24, 32, 48, 64, and 96 m on
the ground, spanning the small-to-medium crown range without
redundant overlap. Adding further anchors in preliminary tests raised
per-image inference time by 15–20% while improving F1 by less than
0.2%, so six was retained. Since almost every crown in the dataset was
nearly square, an aspect ratio of 1.0 accounted for the bulk of cases, and
we added ratios of 0.75 and 1.25 to cover the occasional elon -
gated crown.
The Faster R-CNN backbone was initialized from a Microsoft
Common Objects in Context benchmark pretrained checkpoint (Ball
et al., 2023), with all layers fine-tuned on our ground-truth tree crown
dataset. We treated the red, NIR, and CHM bands as a generic three-
channel input without remapping to conventional RGB statistics.
Therefore, the convolutional filters were allowed to adapt to the resulting
false-color composite during fine-tuning.
2.3.4 One-stage detection: YOLO
Faster R-CNN belongs to the family of two-stage, proposal-
based detectors, which initially generate candidate regions that
are subsequently classified. One-stage detectors instead regress
bounding boxes and class probabilities directly from the image
in a single forward pass, an approach introduced by Redmon et
al. (2016) and refined in subsequent versions of the YOLO archi -
tecture. Because the proposal stage is eliminated, one-stage detec -
tors are typically an order of magnitude faster at inference, a
property that is material when detection must be carried out over
a 150,000 ha landscape rather than a single scene.
We trained YOLOv8 on the identical three-band raster (red, NIR,
hill shade) using the same 750 chips and the same 8,111 digitized
bounding boxes described above, and the same 70/30 training/testing
partition, so that the comparison with Faster R-CNN is not confounded
by differences in data. The network was initialized from COCO-
pretrained weights and fine-tuned for 20,000 steps with stochastic gradi-
ent descent, a learning rate of 0.003, momentum of 0.9, and weight
decay of 0.0005, mirroring the Faster R-CNN protocol. Anchor-free box
prediction removes the need for the anchor-size configuration required
by Faster R-CNN. Detections were retained at a confidence threshold of
0.5 and scored as true positives at an intersection over union (IoU)
greater than 0.1, identical to the Faster R-CNN evaluation, and preci-
sion, recall, F1, omission error, and commission error were computed
with the same matching routine.
2.3.5 Instance segmentation: mask R-CNN
Both Faster R-CNN and YOLO return axis-aligned bounding
boxes. A box is a coarse descriptor of a tree crown, which is approxi-
mately circular, and roughly 21% of the area of a bounding box cir-
cumscribing a circular crown is background rather than crown. Mask
R-CNN (He et al., 2017) extends Faster R-CNN with a third output
branch that predicts a per-instance segmentation mask in parallel with
the box and class branches, at little additional computational cost. For
an inventory task in which crown projected area is itself a quantity of
interest, and in which crown area is used to infer tree height through
the allometric relationship described above, a mask is a more direct
estimator than a box.
Training Mask R-CNN requires polygon annotations rather
than boxes. Rather than re-digitize the 8,111 trees, we derived
instance masks for the existing ground truth by intersecting each
digitized bounding box with the crown polygon produced by the
Schriver et al. 10.3389/ffgc.2026.1954252
Frontiers in Forests and Global Change 09 frontiersin.org
multi-scale watershed segmentation of the same chip, retaining
the connected component whose centroid lies closest to the box
center. Chips for which this procedure failed to yield a single
coherent polygon were flagged and the mask was digitized manu -
ally. Mask R-CNN was then trained with a ResNet-50 backbone
pre-trained on ImageNet ( Deng et al., 2009 ), under the same
optimizer, schedule, and train/test partition used for Faster
R-CNN and YOLOv8. Detection metrics were computed on the
predicted boxes so as to remain comparable with the other detec -
tors; crown area was additionally evaluated against the digitized
crown polygons.
2.4 Assessment of tree identification
Each tree’s location was taken as the centroid of its crown
polygon, and a detection was scored as a true positive if that
centroid fell within a 2 m buffer of the corresponding ground
truth polygon. Detections were matched to their nearest ground
truth box; where more than one detected point mapped to the
same ground truth object, only the nearest was retained as a true
positive and the rest were counted as false positives. The proce -
dures were assessed using three standard metrics, namely preci -
sion ( Equation 2 ), recall ( Equation 3 ), and F1 score ( Equation 4 ):
= +true positiveprecision true positive false positve (2)
= +true positiverecall true positive false negative (3)
precision recall12 precision recallF ∗= ∗ + (4)
Because CNN based procedures and the multi-scale watershed
detector (MSWD) produce geometrically different outputs (i.e., axis-
aligned bounding boxes versus irregular crown polygons derived from
watershed delineation) we applied a matching criterion appropriate to
each output type, while keeping the assignment logic (nearest-match,
one-to-one correspondence, single true positive per ground truth)
identical across both methods so that the resulting precision, recall,
and F1 values remain comparable.
We selected a comparatively low IoU threshold (i.e., 0.1),
rather than the conventional 0.5 used in standard object-detec -
tion benchmarks, because ground-truth boxes were digitized as
coarse rectangular bounds around each crown rather than precise
crown outlines. A stricter IoU threshold would penalize correctly
localized detections for a box-shape mismatch that does not
reflect a true detection error. We note that this choice makes our
reported metrics not directly comparable to IoU@0.5-based
results from other tree-detection studies.
Because MSWD output is a crown polygon rather than a
bounding box, an IoU-based comparison to the rectangular
ground truth would conflate localization accuracy with a shape
mismatch that is not meaningful for this method. Therefore, we
reduced each detected crown polygon to a point by taking its
centroid and counted a detection as a true positive if that point
fell within a ground-truth bounding box or within a 2 m buffer
around it. The buffer was necessary because some ground-truth
boxes captured trees only partially within their 50 × 50 m chip,
which could place a valid detection’s centroid just outside the
labeled box. Each detected point was assigned to its spatially clos -
est ground-truth box; where multiple detected points mapped to
the same ground-truth box, only the nearest was scored as a true
positive and the rest as false positives, which is the same one-to-
one assignment rule applied to CNN’s box-based matches.
Because both procedures answer the same underlying question,
namely “Was a given real tree correctly detected, without double-
counting?, ” using the geometric test appropriate to each method’s
output, the precision, recall, and F1 are comparable despite dif -
fering matching geometry.
Considering that chips are small relative to the overall study area and
were drawn densely across it, random selection of chips does not guaran-
tee spatial separation between training and testing chips, namely chips
from the training and testing sets could be geographically adjacent or, in
principle, drawn from the same local stand. Given that high-resolution
imagery exhibits spatial autocorrelation in canopy structure over these
distances, the estimated precision, recall, and F1 may be optimistic relative
to performance on a truly spatially independent test set.
2.4.1 Tree crown measurement
Detection is only part of the picture: crown shape and projected
area, recoverable from watershed segmentation using either two- or
three-dimensional inputs (Zhen et al., 2016 ), add information of
direct value to forest management. In our implementation, each
detected treetop serves as a sink and the local-maxima-filtered raster
as the elevation surface for a watershed delineation of crown extent
(Wang et al., 2004), an approach Poznanovic et al. (2014) previously
validated for estimating juniper crown area from spectral imagery, and
tree height is then read off as the maximum CHM value inside the
delineated crown.
TABLE 1 Kolmogorov–Smirnov distances evaluating the departure between generated and reference histograms for models trained on distinct datasets (values closer to 0 means the distributions
match closely, whereas value near 1 suggests different distributions).
Histogram Model
Mod. L1 + Both Mod. L1 + Ochoco Mod. L1 + Pine Creek
Wheeler Ochoco + Pine Creek 0.011 0.022 0.037
Pine Creek 0.003 0.001 0.025
Wheeler Ochoco 0.026 0.107 0.010
Bold values indicate models whose training and evaluation used the same dataset combination.
Schriver et al. 10.3389/ffgc.2026.1954252
Frontiers in Forests and Global Change 10 frontiersin.org
3 Results
3.1 Generative adversarial network
GAN output is notoriously hard to evaluate in general, since
appropriate metrics depend heavily on the generation task at hand.
Because our generated pixels represent physical height rather than an
abstract image quality, we opted for a quantitative approach centered
on comparing height histograms (0.3 m bins) from the LiDAR-derived
and GAN-generated CHMs via a two-sample Kolmogorov–Smirnov
(KS) test. For woody vegetation above 1.8 m, this test found no signifi-
cant difference between the two distributions ( p > 0.9; Table 1,
Figure 7). We further examined how the choice of loss function
affected this agreement, comparing our modified L1 loss against the
standard pix2pix L1 loss and against a variant with the adversarial
term removed (Table 2).
Kolmogorov Smirnov provides an overall picture of the accuracy
of the predicted CHM, but local differences can be large across the
landscape. To assess the spatial variation, we executed a systematic
sampling with 50 plots of 1,000 m2. The results show that areas domi-
nated by ponderosa pine exhibit approximately twice as large differ-
ence than areas dominated by western juniper. However, the number
of plots with majority ponderosa pine was minute compared to the
plots with mostly juniper (i.e., 6 ponderosa pine vs. 44 juniper).
Because absolute RMSE scales with the magnitude of the underly-
ing height values, we also considered RMSE normalized by the mean
ground-truth height of each distribution. For the Wheeler Ochoco
distribution, mean height was approximately double that of Pine
Creek, and the corresponding RMSE values scale similarly, suggesting
that much of the apparent difference in model accuracy across areas
reflects differences in canopy height rather than differences in model
fit. When expressed as a percentage of mean height, the models’ rela-
tive errors were considerably more consistent across areas than the
absolute RMSE values suggest.
The evaluation presented in this section characterizes the GAN-
generated CHM at two levels only: the distribution of heights across
the landscape (Kolmogorov–Smirnov test, Tables 1, 2) and the pixel-
wise agreement with the LiDAR-derived CHM (RMSE, Table 3).
Neither analysis assesses the height assigned to any individual tree.
Agreement between height distributions does not imply agreement
at the tree level, because over- and under-estimates at different loca-
tions can offset one another and leave the overall histogram
unchanged. Moreover, the pixel-wise RMSE values in Table 3 are of
the same order of magnitude as the heights of many juniper trees in
the study area, indicating that local errors may be substantial for a
given tree even when the distributions agree. Because paired field
measurements of individual tree height were not available, the accu-
racy of the GAN-derived height of individual trees remains unquan-
tified, and the heights reported for detected trees should be
interpreted as landscape-level estimates rather than validated per-
tree measurements.
3.2 Individual tree crown detection
Adding the synthetic CHM to the multi-scale watershed algo -
rithm increased F1 from 0.644 (2D MSWD) to 0.733 (3D MSWD),
a 13.8% relative improvement. The gain came almost entirely from
precision (0.533 to 0.761), as the CHM height mask removed shrubs
and herbaceous vegetation that the 2D algorithm detected as trees;
commission error fell from 0.467 to 0.239 (a 48.8% reduction).
Recall declined from 0.813 to 0.706, indicating that some trees were
excluded by the 1.8 m height mask, likely those whose synthetic
height was underestimated ( Table 4 ). The 3D RF achieves an
F1-score of 0.707 with a balanced recall of 0.731 and precision of
0.683, while 3D SVM lags slightly behind with an F1 of 0.675, a
recall of 0.697, and a precision of 0.654. These results confirm unso-
phisticated machine learning classifications offer a viable, interpre -
table alternative, but falls short of the accuracy achieved by deep
learning-based object detectors.
Faster R-CNN, operating on the same 3D inputs, delivers a
marked improvement in precision (0.913), at the cost of a more
modest recall of 0.735. This translates to an F1-score of 0.814 and
a commission error of just 0.087, meaning that when Faster
R-CNN detects a tree, it is almost certainly correct. However, this
high precision comes with two notable trade-offs. First, the two-
stage region proposal architecture imposes a substantial
FIGURE 7
Truncated histograms of estimated and generated heights.
TABLE 2 Kolmogorov–Smirnov distances assessing the model histograms trained with different loss functions (values closer to 0 means the distributions match closely, whereas value near 1
suggests different distributions).
Distribution Model
Mod. L1 + Both L1 + Both L1 + No Adv. Loss + Both
Wheeler Ochoco + Pine Creek 0.011 0.023 0.016
Pine Creek 0.003 0.020 0.027
Wheeler Ochoco 0.026 0.027 0.038
Bold values indicate the recommended model (Modified L1 Loss, trained on both datasets).
Schriver et al. 10.3389/ffgc.2026.1954252
Frontiers in Forests and Global Change 11 frontiersin.org
computational burden, making inference over large landscapes,
such as the 150,000 ha study area, potentially prohibitive in terms
of time and resources. Second, and perhaps more critically, Faster
R-CNN depends on a large body of manually annotated ground
truth for training; digitizing the reference dataset used here con -
sumed roughly 48 h of a skilled operator’s time, an investment
that will not pencil out for every management objective. YOLOv8,
as a one-stage, anchor-free detector, was hypothesized to sacrifice
localization accuracy for speed, particularly given the small size
and frequent adjacency of juniper crowns, conditions under
which proposal-free detectors have historically struggled.
However, in our study YOLOv8 achieves a recall of 0.816, which
surpasses both Faster R-CNN (0.735) and 3D MSWD (0.706),
while maintaining a strong precision of 0.860 and an F1-score of
0.837 ( Table 4). Its commission error of 0.140 is slightly higher
than Faster R-CNN’s 0.087 ( Figure 8 ) but considerably lower than
that of 3D MSWD (0.239) and RF (0.317). Mask R-CNN, which
extends Faster R-CNN with a parallel mask prediction branch,
emerges as the overall best-performing method across most met -
rics. With a recall of 0.912, the highest among all methods, a
precision of 0.905, and an F1-score of 0.908 ( Table 4 ), Mask
R-CNN achieves a balance between omission and commission
errors (0.088 and 0.095, respectively).
The metrics assess whether trees were correctly located; but do
not evaluate the accuracy of the height attributed to each
detected tree.
TABLE 3 Pixel-wise root mean square RMSE (m) between generated and ground-truth CHM values for models trained with different loss functions and datasets.
Distribution Model
Mod. L1 + Both Mod. L1 + Ochoco Mod. L1 + Pine
Ck
L1 + Both L1 + No
Adv. + Both
Ochoco + Pine Ck 5.71 5.86 5.78 4.97 3.99
Pine Creek 1.78 1.75 3.09 1.77 1.30
Wheeler Ochoco 8.61 9.63 8.90 7.94 5.87
Bold values indicate lowest RMSE (best pixel-wise accuracy) for each dataset row.
TABLE 4 Individual tree detection metrics across all seven methods.
Algorithm Recall Precision F1 Omission Commission
2D MSWD 0.813 0.533 0.644 0.187 0.467
3D MSWD 0.706 0.761 0.733 0.294 0.239
3D RF 0.731 0.683 0.707 0.269 0.317
3D SVM 0.697 0.654 0.675 0.303 0.346
3D Faster R-CNN 0.735 0.913 0.814 0.265 0.087
3D YOLOv8 0.816 0.860 0.837 0.184 0.140
3D Mask R-CNN 0.912 0.905 0.908 0.088 0.095
2D = imagery without elevation component; 3D = imagery with CHM. MSWD = multiscale watershed detection; RF = random forests; SVM = support vector machine. Bold values indicate
best performance per metric.
FIGURE 8
Detection performance (precision, recall, F1) across all seven methods.
Schriver et al. 10.3389/ffgc.2026.1954252
Frontiers in Forests and Global Change 12 frontiersin.org
3.3 Aggregated tree statistics
Because our tree-level detections can be aggregated to any coarser
scale, we benchmarked that aggregated output against the canopy
cover mapping of Noone et al. (2013) and Nielsen et al. (2014), using
our own digitized crown polygons as ground truth. Our polygons
were originally digitized on a 50-m grid at an offset that did not match
the 30-m grid used in those studies, so we first re-expressed them as
estimated canopy cover on a 30-m grid following the resampling
method of Nielsen et al. (2014): a grid cell was retained for compari-
son only if at least 50% of it was covered by a digitized image chip, and
the canopy cover attributed to each retained cell was normalized by
the fraction of that cell actually overlapped by the chip. Averaging the
absolute difference between this ground truth and each method esti-
mated canopy cover, our approach had a mean absolute error of
4.02%, compared to 8.63% for the random forest implemented by
Noone et al. (2013) and Nielsen et al. (2014) , an improvement of
nearly 60%.
4 Discussion
This study asked whether a canopy height model synthesized
from multispectral imagery can substitute for LiDAR in individual
tree detection across a landscape where LiDAR coverage is incom -
plete. Our findings address this question in four directions. First, the
synthetic CHM reproduced the LiDAR-derived height distribution
closely enough to serve as a structural input for detection, although
its accuracy for individual trees remains unquantified. Second,
adding the synthetic CHM improved detection substantially, mainly
by reducing false positives from low vegetation. Third, Mask R-CNN
provided the most accurate detection, while YOLOv8 offers a speed
advantage for landscape-scale mapping. Finaly, several factors limit
how directly the framework transfers to other landscapes, including
the single-region training data, the temporal mismatch between
imagery and LiDAR data acquisition, and the use of a single
tree class.
4.1 Generative adversarial network
Developing a generalized canopy-height model from multiple
LiDAR acquisitions introduce challenges because datasets may differ
in acquisition date, environmental conditions, and vegetation compo-
sition. Nevertheless, models trained using observations from multiple
locations produced height distributions that more closely matched
reference data than models calibrated using a single site (Figure 7).
This finding suggests that greater variability in the training dataset
improved generalization and reduced sensitivity to local conditions,
consistent with observations reported for multi-temporal forest data-
sets (Hulet et al., 2014).
The Ochoco West project area is dominated by dense ponderosa
pine stands, whereas the Pine Creek project area is defined by low
density juniper savannah, with substantially lower canopy heights.
Because RMSE is sensitive to the absolute magnitude of the target
variable, the larger heights of ponderosa pine inflate RMSE relative to
Pine Creek even when the model’s relative accuracy is comparable;
large individual errors are also disproportionately magnified by the
squared-error term, further exaggerating differences driven by a small
number of outliers. This magnitude sensitivity likely explains the
apparent conflict between the model trained with L1 loss, no adver-
sarial component, and the combined dataset who achieved the lowest
RMSE across all three evaluation distributions but the poorest agree-
ment with the ground-truth height distribution by the KS test. Because
our primary interest is in reproducing the overall distribution of
canopy heights, which determines downstream estimates of tree
count, volume, and value, rather than minimizing pixel-level squared
error, we selected the model with the best KS agreement (Modified L1
loss, combined dataset) as the final GAN used in this study, despite its
comparatively higher RMSE. We report RMSE primarily to illustrate
this discrepancy and to caution against relying on RMSE alone when
evaluating GAN-generated CHMs for forestry applications, rather
than treating it as evidence of superior model performance.
The GAN’s utility for this application would compare its height
estimate for each individual detected tree against that same tree’s field-
measured height. We did not have access to such paired data and were
restricted to the aggregate histogram-based comparisons. Therefore,
the height values feeding the tree segmentation procedures carry an
unquantified individual-tree-level error that our aggregate metrics
cannot characterize.
A further source of uncertainty is the two-to-four-year interval
between the NAIP imagery and the LiDAR acquisitions. During train-
ing, tree mortality may teach the GAN to underpredict height for
some crowns that are visibly present. Height growth, particularly in
ponderosa pine, would work in the opposite direction, with the refer-
ence CHM slightly exceeding the heights that correspond to the imag-
ery. Because the same temporally mismatched LiDAR CHM was used
to evaluate the GAN, both the KS distances and the RMSE values
contain an unknown contribution from real canopy change rather
than model error alone. The direction of this effect on the reported
metrics is not determined a priori. The tree detection evaluation is not
affected in the same way, because ground truth was digitized from the
same 2009 NAIP imagery used as model input.
4.2 Limitations of the GAN-generated CHM
at the individual-tree level
The most direct assessment of the GAN’s utility for tree-level
inventory would compare the height it assigns to each detected tree
against that tree’s field-measured height. Such paired data were not
available, and our evaluation is therefore restricted to distribution-
level (KS) and pixel-level (RMSE) comparisons with the LiDAR-
derived CHM. This limitation has practical consequences, as the
height attributed to each tree, taken as the maximum CHM value
within its crown carries an unquantified error that may be consider-
ably larger than the close agreement of the height distributions sug-
gests. This error is likely not uniform across the landscape; our
systematic plot sample showed that differences between generated and
LiDAR-derived heights in ponderosa pine–dominated plots were
approximately twice those in juniper-dominated plots, although only
six pine-dominated plots were available. Finally, the chip-random
70/30 split used to train and evaluate the GAN does not guarantee
spatial independence, so even the distribution-level agreement may
be optimistic for areas far from LiDAR coverage.
Consequently, the GAN-generated CHM is best suited to applica-
tions that depend on aggregate height structure, such as masking
woody vegetation above 1.8 m, supporting tree detection, or
Schriver et al. 10.3389/ffgc.2026.1954252
Frontiers in Forests and Global Change 13 frontiersin.org
estimating landscape-level height distributions, rather than to applica-
tions that require accurate heights for specific trees, such as individ-
ual-tree volume estimation or prioritizing individual trees for removal
based on size. Future work should collect field-measured heights for
a stratified sample of trees, or use held-out LiDAR tiles as a surrogate,
to quantify per-tree height bias and error and to determine how this
error varies with species, crown size, and terrain.
4.3 Deep learning object detectors
The three deep learning object detectors evaluated in this study,
Faster R-CNN, YOLOv8, and Mask R-CNN, outperform the classical
segmentation and machine learning approaches, though they differ
in specific strengths and trade-offs (Table 4). YOLOv8, as a one-stage,
anchor-free detector, eliminates the region proposal stage entirely,
offering a fundamentally different trade-off. Its commission and
omission errors represent a balanced profile between false positives
and false negatives (Figure 8). The architectural efficiency of YOLOv8
suggests substantial gains in inference speed, making it particularly
attractive for landscape-scale deployment across the 150,000 ha study
area. Mask R-CNN extends the two-stage Faster R-CNN architecture
with a parallel mask prediction branch. Contrary to our initial expec-
tation that the mask branch would primarily benefit crown area
delineation rather than detection, Mask R-CNN achieves the top two
recall, F1-score, and precision among all methods. Its omission error
and commission error of below 0.1 represent the most balanced per-
formance (Figure 7). The mask branch appears to implicitly regular-
ize the detector, improving discrimination of adjacent crowns in
dense canopies, a critical advantage for tree detection in natural
forests.
The performance hierarchy among deep learning methods is clear
but context-dependent: Mask R-CNN leads across most metrics,
YOLOv8 offers the best balance of accuracy and likely speed, while
Faster R-CNN delivers superior precision for applications where false
positives are most costly. All three substantially outperform the clas-
sical approaches, with the gap between the best classical method and
Mask R-CNN representing a 23.9% relative improvement.
Chief among the practical costs of the deep learning methods is
ground truth generation: producing bounding box and mask annota-
tions is a substantial up-front investment that will not be justified for
every management objective. The multi-scale watershed approach
needs no manual annotation at all, and Random Forest and SVM
require only coarse labels, so where the accuracy gain from deep
learning does not offset the annotation burden, the 3D watershed
method remains a practical fallback—while applications that demand
precision-driven or exhaustive tree counts still benefit substantially
from Mask R-CNN and YOLOv8.
4.4 Tree classification
Species classification was initially examined as a complementary
objective because modern object-detection networks can simultane-
ously localize and categorize targets. However, western juniper and
ponderosa pine exhibit similar responses within the available spectral
bands, and the three-channel input configuration limited our ability
to incorporate additional ecological variables that might improve dis-
crimination. Preliminary experiments therefore provided limited
success in separating the two species. To maintain consistency among
all evaluated methods, analyses were conducted using a single tree
class. Future research could investigate the incorporation of environ-
mental variables or higher-dimensional spectral datasets to improve
species-level classification (Wegner et al., 2016).
A limitation of our evaluation procedure for both the CNN-based
detector and the GAN-based CHM generator is that training and test-
ing partitions were split randomly at the chip level rather than by spa-
tially contiguous block. Given the tendency of high-resolution remote
sensing imagery to exhibit spatial autocorrelation, meaning that
neighboring pixels and nearby chips often share highly similar canopy
structure, illumination, and terrain context, the partitioning strategy
does not rule out the possibility that information from training chips
influenced performance on nearby or adjacent test chips, which would
tend to inflate the accuracy metrics relative to performance on a fully
spatially independent test set. A more conservative evaluation would
withhold entire spatial blocks (e.g., sub-basins or contiguous tiles) for
testing, ensuring a minimum separation distance between training
and test regions. We did not adopt this approach here due to the lim-
ited number of contiguous high-quality LiDAR and ground-truth
regions available, which would have substantially reduced the training
data available under a block-holdout scheme.
4.5 Small object detection
Tree crowns combine two properties that are individually difficult
for object detectors: they are small relative to the full image and often
sit close together or overlap. In proposal-based architectures such as
Faster R-CNN, the downsampling used during feature extraction can
cause small, closely spaced objects to merge or disappear from the
feature map altogether (Huang et al., 2017; Hu and Ramanan, 2017),
which we suspect is a major contributor to Faster R-CNN’s compara-
tively high omission error (0.265), the trees it misses tend to be small
or growing in dense clusters.
YOLOv8’s anchor-free design may improve detection of variable-
sized crowns, reflected in its lower omission error of 0.184. Mask
R-CNN’s pixel-wise mask branch appears to provide the most robust
solution for small and adjacent crowns, achieving the highest recall
(0.912) and lowest omission error (0.088). The mask branch enforces
spatial coherence at fine resolutions, enabling better discrimination of
neighboring crowns.
Among classical methods, 2D MSWD achieves recall comparable
to YOLOv8 (0.813 vs. 0.816) but with much lower precision (0.533 vs.
0.860) and higher commission error (0.467 vs. 0.140), indicating that
while segmentation detects many trees, it generates excessive false
positives from shrubs and herbaceous vegetation. The addition of the
CHM (3D MSWD) improves precision to 0.761 at the cost of recall
dropping to 0.706. Random Forest and SVM offer intermediate
results, indicating that classical machine learning classification of
watershed segments offers only modest improvements over pure seg-
mentation and remains inferior to deep learning for small object
detection.
4.6 Band selection and input configuration
The selection of red and NIR bands was motivated by systematic
trial-and-error evaluation based on F1 scores. The R + NIR configura-
tion consistently achieved the highest F1, followed by R + G + NIR,
Schriver et al. 10.3389/ffgc.2026.1954252
Frontiers in Forests and Global Change 14 frontiersin.org
and R + NIR + hillshade (the configuration used in this study), which
achieved an F1 of 0.908. The blue and green bands introduced spectral
noise from background vegetation, reducing precision without
improving recall. The hillshade band added terrain context, improving
precision by approximately 2.1% in complex topography (Davies et al.,
2010). These spectral and terrain inputs were consistent across all 3D
methods (Table 4).
The CHM, generated by a GAN from spectral imagery, provided
critical structural information that distinguished woody vegetation
from herbaceous cover. The impact is most clearly demonstrated by
comparing 2D and 3D MSWD: the CHM reduced commission error
from 0.467 to 0.239 (48.8% reduction) at the cost of a modest recall
decline from 0.813 to 0.706, improving F1 from 0.644 to 0.733. For
Random Forest and SVM, CHM-derived features (e.g., segment mean
height) were among the most important predictors. A conclusion
similar with the 3D MSWD algorithm can be reached for the deep
learning methods, as the combination of spectral (i.e., red and NIR)
and structural (i.e., Syntenic CHM) inputs contributed to strong per-
formances across all three architectures.
The clearest evidence for the value of height information comes
from the multi-scale watershed algorithm, where the synthetic CHM
reduced commission error by 48.8% at the cost of an 11-point drop in
recall. The CHM acts mainly as a filter that removes shrubs and her-
baceous vegetation that are spectrally similar to juniper crowns at
0.5 m resolution. For RF and SVM, CHM-derived features such as
segment mean height were among the most important predictors.
4.7 Synthesis and operational implications
The choice of method for operational mapping depends on three
factors: the accuracy required, the computational budget, and the
availability of annotated training data. Mask R-CNN is the most suit-
able choice where accuracy is the priority, such as complete tree
counts for restoration planning. Its higher recall appears to come from
the mask branch improving separation of adjacent crowns, which is
also where Faster R-CNN loses most trees: small or clustered crowns
that merge during feature downsampling. YOLOv8 is the most suit-
able choice for repeated or very large-area mapping. All three CNNs
require annotated training data; the reference dataset here took
approximately 48 h of a skilled operator’s time to digitize. Where such
data cannot be produced, 3D MSWD requires no annotation and pro-
vides moderate accuracy (F1 = 0.733), making it a practical fallback.
Because RF and SVM performed no better than 3D MSWD while still
requiring labels, we do not recommend them for this application.
Mask R-CNN is the method of choice for applications demanding
exhaustive tree counts and accurate crown delineation, such as carbon
accounting, habitat mapping, and regeneration surveys, offering the
highest recall (0.912), highest F1 (0.908), and second largest precision
(0.905), with omission and commission errors both below 0.10.
YOLOv8 offers an attractive trade-off for rapid, landscape-scale
surveys where inference speed is paramount, such as emergency
response, large-area inventory, or multi-temporal monitoring, deliver-
ing competitive recall (0.816) and precision (0.860) with an F1 of
0.837, likely at a fraction of the computational cost of two-stage
methods.
Faster R-CNN may be preferred for applications where precision
is the overriding concern, such as verifying rare species presence
where false positives are unacceptable, as it achieves the highest
precision (0.913) and lowest commission error (0.087), despite its
recall limitations. 3D MSWD, 3D RF , and 3D SVM serve as pragmatic
baselines when ground truth data are unavailable or resources are
constrained.
Future work should quantify inference speed per unit area for
each method to enable full cost–benefit analysis in operational con-
texts spanning 150,000 ha. The superiority of Mask R-CNN over
Faster R-CNN in recall warrants further investigation into whether
the mask branch implicitly regularizes the detector. For classical
approaches, incorporating more sophisticated features, such as texture
measures, shape descriptors, and multi-scale structural metrics, could
potentially improve RF and SVM performance, reducing the gap with
deep learning methods.
5 Conclusion
In this study, we evaluated two distinct methodological pipe -
lines for individual tree crown detection in eastern Oregon’s eco -
systems, which are dominated by western juniper and Ponderosa
pine. The first pipeline employed classical segmentation and
machine learning approaches: multi-scale watershed segmenta -
tion (MSWD) in both 2D and 3D configurations, followed by
Random Forest (RF) and Support Vector Machine (SVM) classi -
fiers. The second pipeline considered three convolutional object
detectors, Faster R-CNN, YOLOv8, and Mask R-CNN, to detect
tree crowns directly from imagery augmented with a generated
Canopy Height Model.
Our results demonstrate the critical value of structural infor -
mation for accurate tree detection. The integration of a CHM,
generated by a Generative Adversarial Network (GAN) from
spectral imagery, substantially improved performance across all
methods. For the multi-scale watershed algorithm, the addition
of 3D data improved the F1-score from 0.644 (2D MSWD) to
0.733 (3D MSWD), representing a 13.8% relative improvement,
driven primarily by a dramatic reduction in commission error
from 0.467 to 0.239. The Random Forest and SVM classifiers
achieved intermediate performance with F1-scores of 0.707 and
0.675, respectively, indicating that machine learning classifica -
tion offers modest recall improvements over pure segmentation
but at the cost of increased commission errors.
The deep learning object detectors substantially outperformed
all classical methods. Mask R-CNN achieved the highest overall
performance with an F1-score of 0.908, demonstrating that the
mask branch not only provides precise crown delineation but also
improves detection accuracy through implicit spatial regulariza -
tion. YOLOv8 achieved an F1-score of 0.837 with a balanced recall
and precision, offering an attractive trade-off between accuracy
and the inference speed critical for landscape-scale deployment.
Faster R-CNN delivered the lowest commission error (0.087),
making it the preferred choice for applications where false posi -
tives are particularly costly, though its omission error of 0.265 limit
its utility for exhaustive tree counting. Collectively, the deep learn -
ing methods reduced tree identification errors by substantial mar -
gins relative to the best classical approach, with Mask R-CNN
achieving a 23.9% F1 improvement over 3D MSWD and a 28.4%
improvement over 3D RF .
Schriver et al. 10.3389/ffgc.2026.1954252
Frontiers in Forests and Global Change 15 frontiersin.org
Synthesizing a CHM with a GAN is, in itself, a transferable
methodological contribution: recovering three-dimensional
canopy structure from two-dimensional spectral imagery lets
structural information feed into tree detection workflows even
where no concurrent LiDAR acquisition exists. However, the syn -
thetic CHM was validated only at the distribution and pixel levels.
Because paired field-measured heights were unavailable, its accu -
racy for individual trees remains unknown. The heights assigned
to individual detected trees should therefore be treated as indica -
tive rather than validated. Before the GAN-derived CHM is used
for tree-level height estimation, biomass modeling, or other appli -
cations that depend on per-tree height, it should be validated
against field-measured tree heights.
Despite the strong performance of deep learning methods, their
adoption in operational contexts must consider the substantial anno-
tation effort required. For applications where such resources are
unavailable, the 3D multi-scale watershed approach offers a prag -
matic baseline with reasonable performance at no annotation cost.
For applications demanding maximum accuracy, Mask R-CNN and
YOLOv8 provide compelling options, with the choice between them
depending on whether exhaustive recall (Mask R-CNN) or inference
speed (YOLOv8) is prioritized.
Future work should extend this comparison by quantifying
inference speed per unit area for operational deployment across
the 150,000 ha study area, enabling a full cost–benefit analysis.
Additionally, the integration of environmental covariates for
species-level classification, investigation of multi-scale architec -
tures for improved small object detection, and exploration of
transfer learning to other forest types and regions represent
promising directions. The results presented here provide an
empirical foundation for method selection in forest assessments,
demonstrating both the capabilities and the limitations of current
tree detection technologies.
Data availability statement
The raw data supporting the conclusions of this article will be
made available by the authors, without undue reservation.
Author contributions
RS: Formal analysis, Methodology, Software, Visualization,
Writing – original draft, Investigation. JS: Conceptualization, Funding
acquisition, Investigation, Project administration, Supervision,
Writing – original draft. MP: Methodology, Writing – original draft.
BS: Conceptualization, Formal analysis, Investigation, Methodology,
Software, Supervision, Validation, Writing – original draft.
Funding
The author(s) declared that financial support was received for this
work and/or its publication. The study was funded by the SUN grant
of the US Department of Agriculture number 2014-38502-22598 and
by the College of Forestry, Oregon State University.
Acknowledgments
We gratefully acknowledge the staff of the Wheeler Soil and Water
Conservation District for their valuable technical assistance and field
support.
Conflict of interest
The author(s) declared that this work was conducted in the
absence of any commercial or financial relationships that could be
construed as a potential conflict of interest.
Generative AI statement
The author(s) declared that Generative AI was not used in the
creation of this manuscript.
Any alternative text (alt text) provided alongside figures in this
article has been generated by Frontiers with the support of artificial
intelligence and reasonable efforts have been made to ensure accuracy,
including review by the authors wherever possible. If you identify any
issues, please contact us.
Publisher’s note
All claims expressed in this article are solely those of the authors
and do not necessarily represent those of their affiliated organizations,
or those of the publisher, the editors and the reviewers. Any product
that may be evaluated in this article, or claim that may be made by its
manufacturer, is not guaranteed or endorsed by the publisher.
References
Alegbeleye, O. M., Meddens, A. J. H., Rotimi, Y . O., and Ibeh, K. G. (2026). Urban tree
crown detection based on deep learning and high-resolution aerial imagery: PTCNet
for Pullman, W A, USA. Remote Sens. Appl. Soc. Environ. 41, 101818. doi: 10.1016/j.
rsase.2025.101818
Amirkolaee, H. A., and Arefi, H. (2019). Height estimation from single aerial images
using a deep convolutional encoder-decoder network. ISPRS J. Photogramm. Remote Sens.
149, 50–66. doi: 10.1016/j.isprsjprs.2019.01.013
Archuleta, C.-A., Constance, M., Arundel, E. W ., Lowe, A. J., Mantey, K. S., and
Phillips, L. A. (2017) in The National Map Seamless Digital Elevation Model Specifications,
ed. U.S. Geological Survey (Reston, V A: US Geological Survey).
Avery, T. E., and Burkhart, H. (2015). Forest Measurements. 5th Edn Long Grove, IL:
Waveland Press.
Ayrey, E., Fraver, S., Kershaw, J. A., Kenefic, L. S., Hayes, D., Weiskittel, A. R., et al.
(2017). Layer stacking: a novel algorithm for individual forest tree segmentation from
Schriver et al. 10.3389/ffgc.2026.1954252
Frontiers in Forests and Global Change 16 frontiersin.org
LiDAR point clouds. Can. J. Remote. Sens. 43, 16–27. doi:
10.1080/07038992.2017.1252907
Azuma, D. L., Hiserote, B. A., and Dunham, P . A. (2005). The Western Juniper Resource of
Eastern Oregon. Resource Bulletin PNW-RB-248. Portland, OR: USDA Forest Service,
Pacific Northwest Research Station.
Ball, J. G. C., Hickman, S. H. M., Jackson, T. D., Koay, X. J., Hirst, J., Jay, W ., et al. (2023).
Accurate delineation of individual tree crowns in tropical forests from aerial RGB imagery
using mask R-CNN. Remote Sens. Ecol. Conserv. 9, 641–655. doi: 10.1002/rse2.332
Ballanti, L., Blesius, L., Hines, E., and Kruse, B. (2016). Tree species classification using
hyperspectral imagery: a comparison of two classifiers. Remote Sens. 8:445. doi: 10.3390/
rs8060445
Bannari, A., Morin, D., Bonn, F ., and Huete, A. R. (1995). A review of vegetation indices.
Remote Sens. Rev. 13, 95–120. doi: 10.1080/02757259509532298
Baruch-Mordo, S., Poznanovic, A. J., Evans, J. S., Severson, J. P ., Naugle, D. E.,
Maestas, J. D., et al. (2013). Saving sage-grouse from the trees: a proactive solution to
reducing a key threat to a candidate species. Biol. Conserv. 167, 233–241. doi: 10.1016/j.
biocon.2013.08.017
Bengio, Y ., Courville, A., and Vincent, P . (2013). Representation learning: a review and
new perspectives. IEEE Trans. Pattern Anal. Mach. Intell. 35, 1798–1828. doi: 10.1109/
TPAMI.2013.50
Camps-Valls, G., Tuia, D., Bruzzone, L., and Benediktsson, J. A. (2014). Advances in
hyperspectral image classification: earth monitoring with statistical learning methods.
IEEE Signal Process. Mag. 31, 45–54. doi: 10.1109/msp.2013.2279179
Chen, Q., Baldocchi, D., Gong, P ., and Kelly, M. (2006). Isolating individual trees in a
savanna woodland using small footprint LiDAR data. Photogramm. Eng. Remote. Sens.
72, 923–932. doi: 10.14358/pers.72.8.923
Cheng, G., Zhou, P ., and Han, J. (2016). Learning rotation-invariant convolutional neural
networks for object detection in VHR optical remote sensing images. IEEE Trans. Geosci.
Remote Sens. 54, 7405–7415. doi: 10.1109/tgrs.2016.2601622
Davies, K. W ., Petersen, S. L., Johnson, D. D., Bracken Davis, D., Madsen, M. D.,
Zvirzdin, D. L., et al. (2010). Estimating juniper cover from National Agriculture
Imagery Program (NAIP) imagery and evaluating relationships between potential
cover and environmental variables. Rangel. Ecol. Manag. 63, 630–637. doi: 10.2111/
rem-d-09-00129.1
Deng, J., Dong, W ., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. (2009). “ImageNet: a large-
scale hierarchical image database, ” in Proceedings of the 2009 IEEE Conference on
Computer Vision and Pattern Recognition, (Miami, FL: IEEE), 248–255.
Eggert, C., Brehm, S., Winschel, A., Zecha, D., and Lienhart, R. (2017). “ A closer
look: small object detection in faster R-CNN, ” in Proceedings of the 2017 IEEE
International Conference on Multimedia and Expo Workshops (ICMEW) , (Venice,
Italy: IEEE), 421–426.
Gedney, D. R., Azuma, D. L., Bolsinger, C. L., and McKay, N. (1999). Western Juniper in
Eastern Oregon. Resource Bulletin PNW-RB-220. Portland, OR: USDA Forest Service,
Pacific Northwest Research Station.
Ghamisi, P ., and Y okoya, N. (2018). IMG2DSM: height simulation from single imagery
using conditional generative adversarial net. IEEE Geosci. Remote Sens. Lett. 15, 794–798.
doi: 10.1109/LGRS.2018.2806945
He, K., Gkioxari, G., Dollár, P ., and Girshick, R. (2017). “Mask R-CNN, ” in Proceedings of
the 2017 IEEE International Conference on Computer Vision (ICCV), (Venice, Italy: IEEE),
2980–2988.
He, K., Zhang, X., Ren, S., and Sun, J. (2015). Deep residual learning for image recogni-
tion. arXiv [Preprint]. arXiv:1512.03385. doi: 10.48550/arXiv.1512.03385
Hu, P ., and Ramanan, D. (2017). “Finding tiny faces, ” in Proceedings of the 2017 IEEE
Conference on Computer Vision and Pattern Recognition (CVPR), (Honolulu, HI: IEEE),
1522–1530.
Huang, J., Rathod, V ., Sun, C., Zhu, M., Korattikara, A., Fathi, A., et al. (2017). “Speed/
accuracy trade-offs for modern convolutional object detectors, ” in Proceedings of the 2017
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), (Honolulu, HI:
IEEE), 7310–7311.
Hulet, A., Roundy, B. A., Petersen, S. L., Bunting, S. C., Jensen, R. R., and Roundy, D. B.
(2014). Utilizing National Agriculture Imagery Program data to estimate tree cover and
biomass of piñon and juniper woodlands. Rangel. Ecol. Manag. 67, 563–572. doi: 10.2111/
rem-d-13-00044.1
Isola, P ., Zhu, J., Zhou, T., and Efros, A. A. (2017). “Image-to-image translation with
conditional adversarial networks, ” in Proceedings of the 2017 IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), (Honolulu, HI: IEEE), 5967–5976.
Jakubowski, M. K., Li, W ., Guo, Q., and Kelly, M. (2013). Delineating individual trees from
LiDAR data: a comparison of vector- and raster-based segmentation approaches. Remote
Sens. 5, 4163–4186. doi: 10.3390/rs5094163
Jeppesen, D. J. (1977). “Competitive moisture consumption by the western juniper
(Juniperus occidentalis), ” in Proceedings of the Western Juniper Ecology and Management
Workshop, General Technical Report PNW-74, (Portland, OR: USDA Forest Service, Pacific
Northwest Research Station), 83–90.
Jocher, G., Chaurasia, A., and Qiu, J. (2023). Ultralytics YOLOv8 [Computer Software].
Available online at: https://github.com/ultralytics/ultralytics (Accessed July 17, 2026).
Ke, Y ., and Quackenbush, L. J. (2011). A review of methods for automatic individual tree-
crown detection and delineation from passive remote sensing. Int. J. Remote Sens. 32,
4725–4747. doi: 10.1080/01431161.2010.494184
Kershaw, J. A., Ducey, M. J., Beers, T. W ., and Husch, B. (2017). Forest Mensuration. 5th
Edn Hoboken, NJ: Wiley Blackwell.
Längkvist, M., Kiselev, A., Alirezaie, M., and Loutfi, A. (2016). Classification and segmen-
tation of satellite orthoimagery using convolutional neural networks. Remote Sens. 8:329.
doi: 10.3390/rs8040329
Li, W ., Fu, H., Yu, L., and Cracknell, A. (2017). Deep learning based oil palm tree detec-
tion and counting for high-resolution remote sensing images. Remote Sens. 9:22. doi:
10.3390/rs9010022
Li, W ., Guo, Q., Jakubowski, M. K., and Kelly, M. (2012). A new method for segmenting
individual trees from the lidar point cloud. Photogramm. Eng. Remote. Sens. 78, 75–84.
doi: 10.14358/pers.78.1.75
Liu, W ., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.-Y ., et al. (2016). “SSD: single
shot multibox detector, ” in Computer Vision – ECCV 2016, (Cham: Springer), 21–37.
Miller, R. F ., Bates, J. D., Svejcar, T. J., Pierson, F . B., and Eddleman, L. E. (2005). Biology,
Ecology, and Management of Western Juniper (Juniperus occidentalis). Technical Bulletin,
vol. 152 Corvallis, OR: Oregon State University Agricultural Experiment Station.
Miller, R. F ., and Rose, J. A. (1995). Historic expansion of Juniperus occidentalis (western
juniper) in southeastern Oregon. Great Basin Nat. 55, 37–45.
Mou, L., and Zhu, X. X. (2018). IM2HEIGHT: height estimation from single monocular
imagery via fully residual convolutional-deconvolutional network [Preprint].
arXiv:arXiv:1802.10249. doi: 10.48550/arXiv.1802.10249
Næsset, E. (2002). Determination of mean tree height of forest stands by digital photo-
grammetry. Scand. J. For. Res. 17, 446–459. doi: 10.1080/028275802320435469
Nielsen, E., Poznanovic, A. J., and Popper, K. (2014). Accuracy Comparison of Tree
Mapping Methods in Eastern Oregon. Portland, OR: Institute for Natural Resources,
Portland State University.
Nikitina, A. D. (2024). Automatic segmentation of tree crowns in pine forests using Mask
R-CNN on RGB imagery from UAVs. Вопросы Лесной Науки (Forestry Science), 7, 146,
doi: 10.31509/2658-607x-202473-150.
Noone, M., Kagan, J. S., and Nielsen, E. (2013). Western Juniper (Juniperus occidentalis)
and Invasive Annual Grass Mapping in Eastern Oregon. Portland, OR: Institute for Natural
Resources.
Pratama, Y ., Idris, M. Y ., Rasywir, E., Toscany, A. N., Saputra, C., and Bustami, M. I.
(2025). Task-Aligned Pipeline for UAV-Based Palm Tree Age Estimation Using YOLOv8,
CNN Embedding, and Regression Models. In 2025 International Conference on Artificial
Intelligence and Technological Solutions (ICAITech) 2025 Nov 20 (pp. 161–166). IEEE.
Poznanovic, A. J., Falkowski, M. J., Maclean, A. L., Smith, A. M. S., and Evans, J. S. (2014).
An accuracy assessment of tree detection algorithms in juniper woodlands. Photogramm.
Eng. Remote. Sens. 80, 627–637. doi: 10.14358/pers.80.7.627
Redmon, J., Divvala, S., Girshick, R., and Farhadi, F . (2016). Y ou Only Look Once:
Unified, Real-Time Object Detection. In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), Las Vegas BV USA: IEEE. pp. 779–788. doi:
10.1109/CVPR.2016.91
Ren, S., He, K., Girshick, R., and Sun, J. (2017). Faster R-CNN: towards real-time object
detection with region proposal networks. IEEE Trans. Pattern Anal. Mach. Intell. 39,
1137–1149. doi: 10.1109/TPAMI.2016.2577031
Riddell, A. P ., Fitzgerald, S. A., Qi, C., and Strimbu, B. M. (2020). Classification strategies
for unbalanced binary maps: finding ponderosa pine (Pinus ponderosa) in the Willamette
Valley. Remote Sens. 12:3325. doi: 10.3390/rs12203325
Ronneberger, O., Fischer, P ., and Brox, T. (2015). “U-net: convolutional networks for
biomedical image segmentation, ” in Medical Image Computing and Computer-Assisted
Intervention – MICCAI 2015, eds. N. Navab, J. Hornegger, W . M. Wells and A. F . Frangi
(Cham: Springer International Publishing), 234–241.
Safdarinezhad, A., Mokhtarzade, M., and Valadan Zoej, M. J. (2017). Coregistration of
satellite images and airborne LiDAR data through the automatic bias reduction of RPCs.
IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 10, 749–762. doi: 10.1109/
jstars.2016.2585588
Schriver, R. (2018). Landscape Inventory and Harvest Strategies for Individual Western
Juniper in Eastern Oregon. MS Thesis. Corvallis (OR): Oregon State University.
Skurikhin, A. N., Garrity, S. R., McDowell, N. G., and Cai, D. M. (2013). Automated tree
crown detection and size estimation using multi-scale analysis of high-resolution satellite
imagery. Remote Sens. Lett. 4, 465–474. doi: 10.1080/2150704x.2012.749361
Strand, E. K., Smith, A. M. S., Bunting, B. C., Vierling, L. A., Hann, D. B., and Gessler, P . E.
(2006). Wavelet estimation of plant spatial patterns in multitemporal aerial photography.
Int. J. Remote Sens. 27, 2049–2054. doi: 10.1080/01431160500444764
Tang, T., Zhou, S., Deng, Z., Zou, H., and Lei, L. (2017). Vehicle detection in aerial images
based on region convolutional neural networks and hard negative example mining.
Sensors 17:336. doi: 10.3390/s17020336
Schriver et al. 10.3389/ffgc.2026.1954252
Frontiers in Forests and Global Change 17 frontiersin.org
Tzutalin (2015). LabelImg. Git Repository. Available online at: https://github.com/tzuta-
lin/labelImg (Accessed 2018).
Wang, L., Gong, P ., and Biging, G. S. (2004). Individual tree-crown delineation and tree-
top detection in high-spatial-resolution aerial imagery. Photogramm. Eng. Remote. Sens.
70, 351–357. doi: 10.14358/pers.70.3.351
Wegner, J. D., Branson, S., Hall, D., Schindler, K., and Perona, P . (2016). “Cataloging
public objects using aerial and street-level images — urban trees, ” in Proceedings of the
2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), (Las Vegas,
NV: IEEE), 6014–6023.
Y aseen, M. (2024). What is YOLOv8: An In-Depth Exploration of the Internal Features of
the Next-Generation Object Detector. Technical Report.
Y ou, C. C., Lim, S. P ., Lim, S. C., Tan, J. S., Lee, C. K., and Khaw, Y . M. J. (2020). “ A survey
on surface reconstruction techniques for structured and unstructured data, ” in Proceedings
of the 2020 IEEE Conference on Open Systems (ICOS), (Kota Kinabalu, Malaysia: IEEE),
37–42. doi: 10.1109/ICOS50156.2020.9293685
Zhen, Z., Quackenbush, L. J., and Zhang, L. (2016). Trends in automatic individual tree
crown detection and delineation—evolution of LiDAR data. Remote Sens. 8:333. doi:
10.3390/rs8040333
论文原文与出版信息
Frontiers in Forests and Global Change