3D forest semantic segmentation using multispectral LiDAR and 3D deep learning

Narges Takhtkeshha1,2*, Lauris Bocaux1, Lassi Ruoppa3, Fabio Remondino1, Gottfried Mandlburger2, Antero Kukko3, Juha
Hyyppä3
1
3D Optical Metrology (3DOM) unit, Bruno Kessler Foundation (FBK), Trento, Italy;
Email: ntakhtkeshha@fbk.eu, lbocaux@fbk.eu, remondino@fbk.eu
2
Department of Geodesy and Geoinformation, TU Wien, Vienna, Austria; Email: gottfried.mandlburger@geo.tuwien.ac.at
3
Department of Photogrammetry and Remote Sensing, Finnish Geospatial Research Institute, National Land Survey of Finland,
Espoo, Finland; Email: lassi.ruoppa@nls.fi, antero.kukko@nls.fi, juha.hyyppa@nls.fi

Keywords: Multispectral LiDAR point cloud, 3D deep learning, Semantic segmentation, Forest inventory.

Abstract
Conservation and decision-making regarding forest resources necessitate regular forest inventory. Light detection and ranging
(LiDAR) in laser scanning systems has gained significant attention over the past two decades as a remote and non-destructive
solution to streamline the labor-intensive and time-consuming procedure of forest inventory. Advanced multispectral (MS) LiDAR
systems simultaneously acquire three-dimensional (3D) spatial and spectral information across multiple wavelengths of the
electromagnetic spectrum. Consequently, MS-LiDAR technology enables the estimation of both the biochemical and biophysical
characteristics of forests. Forest component segmentation is crucial for forest inventory. The synergistic use of spatial and spectral
laser information has proven to be beneficial for achieving precise forest semantic segmentation. On the other hand, 3D deep
learning (DL) models have demonstrated outstanding performance in point cloud segmentation compared to traditional methods.
Thus, this study aims to investigate the potential of MS-LiDAR data, captured by the HeliALS system, providing high-density
multispectral point clouds to segment forests into six components: ground, low vegetation, trunks, branches, foliage, and woody
debris. Three point-wise 3D deep learning models and one machine learning model, including kernel point convolution (KPConv),
superpoint transformer (SPT), point transformer V3 (PTv3), and random forest (RF), are implemented. Our experiments confirm
the superior accuracy of the KPConv model. Additionally, various geometric and spectral feature vector scenarios are examined.
The highest accuracy is achieved by feeding all three wavelengths (1550 nm, 905 nm, and 532 nm) as the initial features into the
deep learning model, resulting in improvements of 33.73% and 32.35% in mean intersection over union (mIoU) and in mean
accuracy (mAcc), respectively. This study highlights the excellent potential of multispectral LiDAR for improving the accuracy in
fully automated forest component segmentation.

1. Introduction
By providing a holistic understanding of the forest ecosystem, forest inventory supports sustainable management practices that
balance economic, ecological, and social values. However, forest measurements are still primarily collected through labor-intensive
in-situ methods using simple tools such as calipers, measuring tapes, and hypsometers. These direct measurement techniques are
time-consuming and prone to human error. Over the past two decades, very high-density terrestrial laser scanners (TLS) have
garnered significant attention for their ability to collect detailed 3D forest structural data with millimeter-level accuracy, enabling
rapid, automated, and periodic forest inventories (Tanaka et al., 1998; Liang et al., 2016). TLS systems cover significantly smaller
areas compared to airborne laser scanning (ALS), and there is a need for multiple scans to map even small test plots. TLS typically
speeds up plot measurements, but a large amount of work is needed in the processing of the data. Conversely, recent unmanned
laser scanners (ULS) can capture dense point clouds of entire tree structures in a relatively short time, presenting substantial
potential to improve the accuracy and efficiency of detailed tree measurements and forest inventory (Jaakkola et al. 2010; Wallace
et al., 2012). Accurate measurement of certain tree attributes often depends on one or two specific parts of the tree. Tree foliage
plays a critical role in controlling forest productivity, health, and responses to climate change, as it directly influences
photosynthesis and evapotranspiration. Foliage profiles are also of particular interest for biodiversity conservation. Trunks and
branches are essential for obtaining quantitative structural models (QSMs) of trees and estimating the timber strength (Winberg et
al., 2023; Pehkonen et al., 2025). Tree trunks contribute to the estimation of tree’s diameter at breast height (DBH), biomass, stem
volume, and stem quality (Levick et al., 2016), while identifying branches is necessary to exclude them from leaf area index (LAI)
and biomass estimation processes (Calders et al., 2015). The presence of woody debris serves as a biodiversity indicator, as many
threatened species rely on decaying wood for habitat. Additionally, forest component segmentation facilitates individual tree
segmentation and tree modelling (Heinzel et al., 2018). Therefore, separating forest component is a useful pre-processing step for
all forestry applications development whether they predict or directly measure tree attributes from point clouds. Additionally, such
information would be beneficial even for simultaneous localization and mapping (SLAM) processes in mobile laser scanners to
avoid mismatching errors caused by under-storey and foliage points (Kaijaluoto et al., 2022).
 Currently, forest semantic segmentation primarily relies on the geometric information derived from LiDAR data. Some studies
have explored the impact of single-wavelength information from traditional monochromatic (mono) LiDAR systems on forest
component segmentation and demonstrated its advantage (Windrim et al., 2020; Kaijaluoto et al., 2022). Since the back-scattered
and captured light by the sensor varies depending on the material hit by the laser, incorporating laser intensity data can enhance
segmentation accuracy (Morel et al., 2020). The advent of MS-LiDAR technology, which operates across multiple channels, enables
the simultaneous collection of spatial and multispectral information (Takhtkeshha et al., 2024a). As forest components exhibit
distinct optical properties at different operating wavelengths, MS-LiDAR data can help with distinguishing between various forest
elements (Disney, 2018). The effects of mono- and multispectral information on forest component segmentation is discussed in
detail in Subsection 2.1. While the geometric information from LiDAR data facilitates the extraction of structural metrics, the
spectral information provided by MS-LiDAR systems not only enhances the accuracy of these characteristics but also offers new
critical insights into tree health monitoring. By identifying parameters such as water content and biochemical composition, MS-
LiDAR systems enable more reliable and comprehensive forest inventory practices.

To date, numerous innovative approaches have been proposed for semantically segmenting forest environments. Tao et al. (2015)
horizontally sliced point clouds to detect circle-like shapes for separating wood and leaves. Ferrara et al. (2018) applied density-
based spatial clustering (DBSCAN) to classify wood and non-wood points while Heinzel et al. (2018) employed mathematical
morphology techniques to identify tree trunks. However, these model-based methods are designed for a limited range of tree species
and architectures, making them less robust when applied to diverse tree structures. Machine learning (ML) algorithms using hand-
crafted features such as linearity, planarity, scattering and surface curvature have been explored in several studies to estimate DBH,
stem curvature, and stem volume (Yun et al., 2016; Ma et al., 2016; Hyyppä et al., 2020). Wang et al. (2017) compared four popular
ML classifiers—support vector machine (SVM), naive bayes (NB), random forest, and gaussian mixture model (GMM)—for
distinguishing wood from leaves using TLS data. Their findings identified RF as the most effective model while emphasizing the
critical role of density-related features. Despite these advancements, dense understory vegetation and complex forest structures
continue to challenge the generalizability of most conventional approaches. Even for human experts, accurately distinguishing
trunk boundaries, foliage initiation points, and other classes can be difficult. This underscores the need for more advanced
segmentation algorithms capable of handling diverse forest types, particularly evergreen tropical forests. DL models have shown
exceptional promise in image segmentation tasks. To adapt these models to point cloud data and address challenges associated with
processing irregular structures, specialized point-wise deep learning architectures have been developed since 2017 (Ruizhongtai
Qi et al., 2017). Unlike ML, the notable advantage of DL models is their ability to automatically extract and select features. Recent
studies have begun to employ DL models for segmenting various forest components (Morel et al., 2020; Windrim et al., 2020;
Krisanski et al., 2021a; Krisanski et al., 2021b; Kaijaluoto et al., 2022; Shao et al., 2023; Xiang et al., 2024; Liang et al., 2024).
These approaches unveiled the significant potential of DL models. A detailed review of DL-based studies is provided in Subsection
2.2. Although deep learning models have been applied in some studies for MS-LiDAR semantic segmentation (Takhtkeshha et al.,
2024b; Reichler et al., 2024), to the best of our knowledge, they have not yet been used for forest MS-LiDAR.
To take full advantage of MS-LiDAR technology for precise forest inventory, firstly, forest semantic segmentation is a mandatory
task. Thus, the primary objective of this paper is to introduce a reliable and automated forest semantic segmentation pipeline capable
of handling structurally diverse and complex forest point clouds. Building on existing research and leveraging recent advancements
in data acquisition and processing, we present a deep learning-based forest semantic segmentation pipeline utilizing high-density,
three-wavelength multispectral LiDAR data. This work focuses on segmenting six forest components: ground, low vegetation,
trunks, branches, foliage, and woody debris. The innovative aspects of the work are:

 The investigation of the impact of multi-wavelength (more than two) LiDAR data on forest component segmentation;
 The use of deep learning methods for the semantic segmentation of forest multispectral point clouds;
 The segmentation of forest point clouds into six forest components;
 The analysis of the influence of vegetation indices on the initial feature vector for DL-based forest semantic segmentation;
 The release of the first publicly available forest MS-LiDAR benchmark dataset (once the paper be accepted).

2. Related Work

2.1 Laser Radiometric Information in Forestry

To date, MS-LiDAR technology has demonstrated significant potential across diverse ecological applications, including enhancing
the detection and segmentation of individual trees (Huo and Lindberg, 2020; Dai et al., 2018), identifying invasive tree species
(Mielczarek et al., 2022), classifying tree species (Budei et al., 2017; Hakula et al., 2023), and estimating tree physiological and
biochemical properties such as moisture content and foliar nitrogen (Gaulton et al., 2010; Gaulton et al., 2013; Eitel et al., 2014;
Wallace et al., 2014; Junttila et al., 2018; Goodbody et al., 2020; Maltamo et al., 2020).
1
Table 1. Summary of deep learning-based forest component segmentation studies using laser scanning in the literature; the best
method is bold.
References LiDAR Forest classes Methodology(s) Spectral Remarks
data effect
3D
Lu et al., 2025 ULS 5 (Ground, low Sen-net NA Open data and code
vegetation, stem,
live branches, and
woody branches)
3D
Liang et al., 2024 TLS 3 (Trunk, branches, Point transformer, NA Open data
and foliage) DGCNN, PointNet++ (ForestSemantic);
private code
3D
Xiang et al., 2024 ULS 5 (Ground, low ForAINet -0.3% The negative spectral
vegetation, stem, effect could be due to
live branches, and combining two mono-
dead branches) spectral scanners from
different regions.
Open data and code
3D
Shao et al., 2023 ULS 5 (Ground, debris, Superpoint Graph NA Private data and code
trunks, crowns, and
others)
3D
Kaijaluoto et al., MLS 4 (Ground, RandLA-Net and 2D +1.3% Open data, private code
2022 understory, LSSegNet
trunk, and foliage)
3D
Krisanski et al., TLS 5 (Terrain, PointNet++ NA Open code, private data
2021a vegetation,
coarse woody
debris, and stem)
3D
Krisanski et al., TLS, MLS, 5 (Terrain, Modified PointNet++ NA Open data and code
2021b ALS, and vegetation,
ULS coarse woody
debris, and stem)
3D
Windrim et al., 2020 ALS 2 (Stem and foliage) Fully convolutional +2.9% Private data and code
network
3D
PointNet +2.3%
3D
Morel et al., 2020 TLS 2 (Leaf and wood) Improved PointNet++ NA Private data and code

Depending on the wavelength of the scanner, the intensity of the reflected LiDAR pulse can facilitate the differentiation of return
signals resulting from interactions with various materials (Tao et al., 2015). Li et al. (2013) illustrated the potential of integrating
two monochromatic laser scanners operating at 1548 nm and 1064 nm to separate leaves from trunks and branches. They observed
that at SWIR wavelengths, leaves exhibit approximately half the brightness of trunks due to liquid water absorption. Conversely,
at the NIR range of the electromagnetic spectrum, leaves and trunks displayed roughly equal brightness. The advantages of multi-
wavelength systems for distinguishing wood from leaves and improving tree parameter estimation have been further validated by
Danson et al. (2014), Howe et al. (2015), and Li et al. (2018). These studies utilized dual-wavelength (SWIR and NIR) full-
waveform TLS systems, such as the Salford Advanced Laser Canopy Analyser (SALCA) and the dual-wavelength Echidna LiDAR
(DWEL), alongside traditional methods like thresholding and random forest classification. The synergetic use of spatial and spectral
information not only improves classification accuracy but also reduces variance across multiple vegetation types (Li et al., 2018).
The application of spectral information in forestry is not limited to just multi-wavelength LiDAR systems.
Even incorporating a single wavelength spectral information into forest semantic segmentation tasks has demonstrated accuracy
enhancement by up to 34% (Zhu et al., 2017; Windrim et al., 2020; Kaijaluoto et al., 2022). Béland et al. (2014) demonstrated that
applying a threshold in the 1535–1550 nm spectral range effectively distinguished between leaf and woody materials. This
advantage arises because solid targets, such as wood, produce sharply peaked return pulses compared to foliage, enabling
radiometric information to aid in differentiating them. Although the impact of mono- and dual-wavelength LiDAR intensities on
forest component segmentation has been investigated, to the best of our knowledge, no studies have explored forest semantic
segmentation using more than two wavelengths. Besides, former studies are TLS-based. Notably, prior research employing MS-
LiDAR systems has focused on segmenting forest laser data into woody and non-woody elements. In contrast, our study considers
six forest classes, enhancing the 3D forest inventory.
2.2 Deep Learning for Forest Semantic Segmentation
As outlined in Section 2.1, the potential of DL models has remained unexplored in previous studies on forest semantic segmentation
using MS-LiDAR data. Since this paper proposes a novel DL-based approach for forestry multispectral point cloud segmentation,
this section reviews prior research on forest component segmentation utilizing DL, as summarized in Table 1. According to Table
1, most DL models employed in this domain are based on PointNet++ (Qi et al., 2017), while other notable models include
Superpoint Graphs (Landrieu and Simonovsky, 2018), RandLA-Net (Hu et al., 2019), point transformer (Zhao et al., 2021),
DGCNN (Wang et al., 2018), and fully convolutional networks. The additional attributes of multispectral point clouds make graph-
based DL models computationally expensive. The superiority of point transformer over PointNet++ and DGCNN has been
demonstrated by Liang et al. (2024). Kaijaluoto et al. (2022) proposed a novel method for real-time semantic segmentation of
mobile laser scanning (MLS) forest point clouds by utilizing raw, non-georeferenced 2D laser scanner profiles formatted as a
sequence of 2D rasters. Each raster series included range, reflectance, and echo deviation data at a single scan line level. Their U-
Net-based 2D convolutional neural network (LSSegNet) demonstrated slightly lower accuracy compared to the point-wise
RandLA-Net model. More recently, Lu et al. (2025) presented a KPConv-based model called Sen-net for semantic segmentation
of ULS data across five forest components, demonstrating superior performance compared to state-of-the-art deep learning models
such as RandLA-Net.

Among the reviewed studies, only three out of nine DL-based papers have examined the effect of laser intensity. Importantly, none
of them have explored the impact of multispectral laser information. Windrim et al. (2020) and Kaijaluoto et al. (2022) substantiated
that incorporating single-wavelength LiDAR intensity data can enhance stem-foliage separation accuracy by up to 2.9%.
Conversely, Xiang et al. (2024) reported that while involving monochromatic LiDAR intensity improved segmentation accuracy
for ground and low vegetation, it reduced the accuracy for stem and dead branch segmentation. A slight decrease is also observed
for live branch segmentation accuracy. The -0.3% decrease in mIoU of segmentation may be attributed to the characteristics of the
FOR-Instance dataset, which comprises a mix of mono-spectral data from five countries that are collected using two different
sensors, acting at different wavelengths—RIEGL VUX-1 UAV (1550 nm) and RIEGL miniVUX-1 UAV (905 nm).

While most existing studies have focused on binary leaf-wood separation, some recent research, such as that by Shao et al. (2023),
Xiang et al. (2024), and Lu et al. (2025), has expanded the classification to include multiple forest components. Shao et al. (2023)
identified ground, debris, trunks, crowns, and other elements, whereas Xiang et al. (2024) and Lu et al. (2025) categorized forest
structures into ground, low vegetation, stems, live branches, and dead branches. In this paper, we selected three state-of-the-art
deep learning models and one machine learning algorithm for segmenting six forest components: KPConv (Thomas et al., 2019),
SPT (Robert et al., 2023), PTv3 (Wu et al., 2024), and RF. According to the literature, RF has been proved as the best machine
learning classifier for forest component segmentation (Wang et al., 2017). Superpoint transformer and point transformer V3 are
new 3D deep learning algorithms that have not yet been used for forest semantic segmentation. KPConv has shown promising
results in forest component segmentation in the recent study by Lu et al. (2025). It is important to note that PointNet++ has not
been implemented in this study because PointNet++ is among the first 3D deep learning models, and the superiority of the point
transformer over PointNet++ for forest component segmentation has already been demonstrated by Liang et al. (2024). The details
of the semantic segmentation process are thoroughly presented in Subsections 4.1- 4.6. Detailed information on these DL models
is presented in Section 4.

3. Study Area and Dataset

The case studies are four biodiverse natural forest regions located in Espoonlahti in southern Finland (Figure 1a). In Espoonlahti
area, there are close to 20 different tree species available, the most common are pine, spruce, birch, maple, aspen, rowan, oak, lime,
and alder. High-density point cloud dataset is acquired in July 2023 in leaf-on condition by the helicopter-mounted HeliALS
multispectral LiDAR system developed by the Finnish Geospatial Research Institute (FGI). This MS-LiDAR system is a
combination of three single-wavelength laser scanners, namely VUX-1HA, miniVUX-1DL, and VQ-840-G (Hakula et al. 2023).
Thus, this system can capture 3D data in three wavelengths: 1550 nm, 905 nm, and 532 nm. Detailed information on the HeliALS
system is reported in Table 2.

(a) (b)
Figure 1. Study plots (a) in Espoonlahti, Finland, and the corresponding distribution of forest components (b).
Figure 2. Multispectral forest point clouds (a, c, e, h), visualized using pseudo-coloring based on the intensity values of the SWIR, NIR, and
green scanners mapped to the red, green, and blue channels, respectively. The corresponding normalized height are shown in (b, d, f, i).
Subfigures (g) and (j) present the manually generated ground truth for the test areas.
Ground truth data are created manually in six classes (ground, low vegetation, trunk, branches, foliage, and woody debris) using
CloudCompare v2.13.0's segmentation tool to train and evaluate DL models. A total of 9601494 points are annotated. The dataset
is divided into training and test sets with a ratio of 4:1. Because of the limited availability of training data and the complexity of
annotating forest point clouds across six classes, validation data is not included in this study. Due to the complex nature of forest
ecosystems, the classes within our dataset are not uniformly represented (Figure 1b). On average, woody debris, trunks, and
branches account for only 0.23%, 1.51%, and 2.52% of the dataset, respectively. In comparison, low vegetation covers 6.54% of
the forest, while 20.14% of the points correspond to the forest floor (ground). Most of the points (69.08%) are attributed to foliage.
Hence, we have a highly class-imbalanced dataset. The multispectral point cloud dataset used in this study is presented in Figure 2
and is publicly available as the first forest MS-LiDAR benchmark dataset for use in future research.

Table 2. Specification of the multispectral LiDAR dataset over the forestry area.

Scanner VUX-1HA miniVUX-1DL VQ-840-G

Wavelength (nm) 1550 905 532
Average point density 530 163 604
(point/m2)
Pulse repetition rate 1017 100 200
(kHz)
Beam divergence (mrad) 0.5 0.5 × 1.6 1
Flight altitude (m) ~ 100

3.1 Multispectral LiDAR vs. Monochromatic LiDAR

Figure 3 depicts the geometry difference between the multispectral and monochromatic point clouds. While monochromatic laser
scanner components of some MS-LiDAR systems, like Optech Titan and RIEGL VQ-1560i-DW, have roughly the same point
density, the varying pulse repetition rates of individual monochromatic laser scanners in the HeliALS system result in differences
in the geometries of individual laser wavelengths. Among the monochromatic laser scanners, VUX-1HA provides the most accurate
representation of individual trees, though it lacks full corresponding points for the low vegetation and woody debris. Conversely,
the VQ-840-G offers superior geometric representation of the low vegetation and woody debris, but has sparse points for branches,
ground, and particularly trunks. The miniVUX-1DL performs the worst in scanning forest components. According to Figure 3d,
the multispectral HeliALS system not only enhances the spectral richness of point clouds but also provides the most comprehensive
representation of all forest classes thanks to its three scanners. To more thoroughly assess the advantages of MS-LiDAR systems
over conventional mono-LiDAR scanners, we conducted an ablation study using different single-reflectance values. The results are
presented in Subsection 5.2.

Figure 3. Surveyed geometry of forest components: mono-scanners (VUX-1HA (a), miniVUX-1DL (b), VQ840-G (c)) vs.
MS-LiDAR system (d).
 4. Methodology
Traditional machine learning methods usually rely on extracting representative geometric features such as verticality, linearity,
planarity, sphericity, and principal components to facilitate the forest point cloud semantic segmentation. As previously mentioned,
this study aims to explore the potential of radiometric information of MS-LiDAR data for boosting forest semantic segmentation.
The spectral histograms of the six forest classes are illustrated in Figure 4. According to these subfigures, all six forest components
possess unique multispectral distributions. Notably, although the spectral behaviours of low vegetation and foliage are roughly
similar in the Green and NIR wavelengths, they act differently in SWIR wavelength. As a result, we propose a spatial-multispectral
pipeline for forest semantic segmentation using 3D deep learning and MS-LiDAR data which synergistically integrates geometric
and multispectral laser information. We employ laser reference values instead of amplitude values, as the former are range-corrected
(Pfennigbauer and Ullrich, 2010).

Figure 4. Distribution of reflectance values of forest components across different laser wavelengths. Reflectance values are
scaled between 0 and 1 for visualization purposes.

Figure 5 provides an overview of the proposed methodology. In the first step, outliers in the point clouds from each monochromatic
laser scanner are removed using the SOR (statistical outlier filter) tool in CloudCompare, with default parameters. Subsequently,
multispectral point clouds are generated by merging the individual monochromatic laser data using the
preMergeChannelsPointclouds module in OPALS software (Mandlburger et al., 2009; Pfeifer et al., 2014). The number of neighbors
is configured to one, with a search radius of 0.25 meters in 3D space. Points for which complete spectral information (i.e., all three
wavelengths) cannot be obtained from neighboring points are discarded. The average point density of the resulting multispectral
point clouds is 1198.19 points/m², approximately twice as high as the densest mono-scanner (VQ-840-G). Subsequently, the
coordinates and reflectance values of the multispectral point clouds are fed into deep learning models. We implemented three state-
of-the-art 3D deep learning models—KPConv, SPT, and PTv3 — along with established random forest machine learning algorithm
to compare them for spatial-multispectral forest semantic segmentation.

Moreover, we investigated the effect of MS LiDAR information by considering various initial feature vector scenarios, including
solely reliance on 3D coordinates, single-channel reflectance, dual-channel reflectance, reflectance values of all channels, and the
inclusion of a vegetation index (VI). Detailed information on the integration of vegetation indices is provided in Section 4.1.
The results are validated using established metrics, including overall accuracy (OA), mean accuracy, intersection over union (IoU),
and mean IoU. These metrics are based on the true positives (TP), true negatives (TN), false positives (FP), and false negatives
(FN). TP represents the number of points that belong to a specific forest component in the real world and are correctly predicted
by the model. TN refers to the number of points that the model correctly identifies as not belonging to that forest component. FP
indicates the number of points that the model incorrectly classifies as belonging to a certain forest component. FN are the points
that the model fails to identify correctly as part of a certain forest class. Furthermore, to compare deep learning models and initial
feature vectors tailored for forest inventory purposes, we propose a new evaluation metric that assigns twice the importance to key
forest components—trunks, branches, foliage, and woody debris. Given that mIoU is the primary metric for assessing semantic
segmentation, we introduce weighted IoU (wIoU). In this approach, the weight of each class is doubled for trunks, branches, foliage,
and woody debris to better reflect their significance in forest inventory. All experiments are conducted on a Linux Mint 21.3
machine, equipped with a 32-core AMD EPYC 9354 CPU and an NVIDIA L40 GPU with 48.31 GB of onboard memory. The
entire code for our study is developed in Python and is publicly available at https://github.com/3DOM-FBK/3D-forest-semantic-
segmentation.
 Figure 5. Proposed pipeline for precise forest inventory based on semantic segmentation of MS-LiDAR data.

4.1 Common Considerations in Semantic Segmentation

To normalize the features, the point clouds are centered along the planimetric axes (X and Y) according to Equations 1 and 2.
Heights are normalized by subtracting each point's height from the local minimum height using OPALS software and then making
sure that elevations are positive by subtracting the minimum value from height values (Equation 3). Other spectral and geometric
features are normalized by applying a combination of robust (outlier-resistant) scaling and min-max scaling as described in
Equation 4. In this formula, f denotes the spectral feature, and IQR (interquartile difference) is the difference between the 75th and
25th quantiles.

X = X - mean (X) (1)

Y = Y - mean (Y) (2)

Z= Z – min (Z) (3)

f - mdeian(f) (4)
f_normalized =
IQR(f)

where IQR(f) = Q75 (f) - Q25 (f)

f_normalized - min (f_normalized)

f_normalized =
max (f_normalized) - min (f_normalized)

Most advanced 3D deep learning models have built-in data augmentation tools. In some cases, both geometric and spectral
augmentations are considered. For example, KPConv randomly drops a specific percentage of spectral information and introduces
noise. However, since reflectance values correspond to the characteristics of the target being hit and remain stable under varying
illumination conditions thanks to active LiDAR sensors—unlike images—typical radiometric augmentation techniques such as
jittering, value dropping, and random noise can negatively impact reflectance values. As a result, these techniques may cause
models to rely more on geometric features. Therefore, we deactivated such radiometric augmentations in the models we used.
4.2 KPConv
KPConv is a distinctive architecture that positions convolutional weights in space using kernel points, rather than relying on multi-
layer perceptron (MLP) encodings (Thomas et al., 2019). We utilized the open-source PyTorch implementation of KPConv 1. To
mitigate class imbalance, class frequency-based weighting is applied to the Cross-Entropy loss function during segmentation. For
parameterizing the KPConv model in the forest semantic segmentation task, we empirically set the initial subsampling grid size to
2 cm, the input sphere radius to 4 cm, the number of kernel points to 30, and the batch size to 6. Furthermore, we configured the
convolution behavior to Gaussian and the aggregation function to “closest”. The model is trained for 300 epochs with a learning
rate of 0.01. The default stochastic gradient descent (SGD) optimizer is used. The remaining parameters are set to their default
values.

4.3 Superpoint Transformer

The PyTorch-based Superpoint Transformer2 is publicly available. This model incorporates a fast implementation of the cut pursuit
algorithm for partitioning point clouds into a hierarchical superpoint structure, along with a self-attention mechanism that captures
relationships between superpoints across multiple scales (Robert et al., 2023). This model stands out for its high computational
efficiency, which is particularly important when processing high-density and high-dimensional multispectral point clouds. To tackle
the class imbalance issue, we weighed the Cross-Entropy loss by computing a weighted average of the per-sample losses,
encouraging the model to pay more attention to underrepresented classes. Additionally, we weighed the sampling of our data when
creating batches to oversample under-represented classes and subsample over-represented ones.

SPT utilizes a set of point feature vectors for hierarchical partitioning of the point cloud and training. By default, these vectors
include several hand-crafted features that are automatically extracted by SPT. In our approach, we selected normalized height,
reflectance at three wavelengths, verticality, linearity, planarity, scattering, and curvature as the point features used for both
partitioning the point cloud into superpoints and for training. The voxel resolution is set to 1 cm. The number of nearest neighbors
is set to 25, with a search radius of 1 meter. The most critical and sensitive parameters in SPT are regularization, spatial_weight,
and cutoff, which serve as inputs to the Cut Pursuit partitioning algorithm. We empirically set the values of regularization,
spatial_weight, and cutoff to [0.3,0.6,0.8], [1e-1, 1e-1, 1e-1], and [10, 10, 10], respectively, based on observations from example
datasets. The model is trained for 400 epochs with a learning rate of 0.1, a weight decay of 0.01, and the default AdamW optimizer.

4.4 Point Transformer V3

Point transformer V3 represents the latest advancement in the point transformer model series. A key innovation of this architecture
is its use of the self-attention mechanism for point cloud segmentation, first introduced by Zhao et al. (2021). Compared to its
predecessors, PTv3 is specifically designed to address the traditional trade-off between accuracy and efficiency by removing the
reliance on relative positional encoding and replacing complex attention patch interaction mechanisms with a streamlined approach
optimized for serialized point clouds (Wu et al., 2024). We used open-source PTv3 3. To address the class imbalance problem, we
employed the Lovász-softmax loss function. The training is conducted over 800 epochs using the default AdamW optimizer with
a batch size of 4, a learning rate of 0.01, and a weight decay of 0.05.

4.5 Random Forest

Random forest classifier has demonstrated superior performance compared to other machine learning models in the literature.
Considering the features chosen by Wang et al., 2017 and our experimental results, we selected 15 hand-crafted features extracted
using CloudCompare’s “compute geometric features” tool, experimentally applied at 1 m local neighbourhood radius. These
features include linearity, planarity, verticality, eigenentropy, surface variation, anisotropy, omnivariance, sum of eigenvalues,
second eigenvalue, first and second elements of principal component analysis (PCA). In addition, reflectance values of the three
wavelengths and NDVI NIR-SWIR as supplementary features. Detailed information on the reasons for considering NDVI NIR-SWIR as
the vegetation index is provided in Subsection 4.6. The main parameters for the RF classifier are set as follows: number of
estimators = 200, maximum depth = 50, maximum features = log2, minimum samples split = 2, and minimum samples leaf = 10.
The class weight is set to “balanced” to deal with the imbalanced dataset.

4.6 Integration of Vegetation Index

As the RIEGL reflectance values are in dB, to calculate vegetation indices, we first converted reflectance values to linear units by
Equation 10 (Pfennigbauer and Ullrich, 2010). We choose an appropriate vegetation index from computed five vegetation indices
frequently used in the literature and compared their sensitivity to forest components. Using reflectance values, we calculated the
normalized difference vegetation indices (NDVI NIR-Green and NDVI NIR-SWIR), chlorophyll vegetation index (CVI), green ratio
vegetation index (GRVI), and green difference vegetation index (GDVI), as expressed in Equations 5-10 (Luo et al., 2022;
Goodbody et al., 2020; Pan et al., 2019). Based on the results presented in Figure 6, the NDVI NIR-SWIR demonstrates superior
capability in distinguishing forest components. Hence, NDVI NIR-SWIR is considered the selected vegetation index to be fed into the
best-performing deep learning model, alongside the reflectance values.

1 https://github.com/HuguesTHOMAS/KPConv-PyTorch
2 https://github.com/drprojects/superpoint_transformer
3 https://github.com/Pointcept/PointTransformerV3
 reflectance db (5)
reflectance linear =10 10
(NIR - Green) (6)
NDVINIR-Green =
(NIR + Green)
(NIR - SWIR) (7)
NDVINIR-SWIR =
(NIR + SWIR)
NIR × Green (8)
CVI=
SWIR2
NIR (9)
GRVI=
Green
GDVI = NIR - Green (10)

Figure 6. The sensitivity of the different vegetation indices to distinguishing forest components.

5. Results

5.1 Performance of Different Deep Learning and Machine Learning Models

Figure 7 reports the accuracy of forest semantic segmentation achieved by the mentioned DL and ML classifiers. According to this
figure, KPConv demonstrates a substantial performance advantage over the other classifiers. Deep learning models outperform the
traditional random forest machine learning model, improving wIoU by approximately 49.24%. Among the deep learning models,
KPConv achieved the highest accuracy, followed by PTv3 and SPT, which exhibited the next highest accuracies, respectively. The
pretrained KPConv weights are available in our GitHub repository and can be used as initial weights for training in future studies.
Since KPConv outperforms the other classifiers, we consider this model for further spectral ablation study.

Figure 8 presents the confusion matrices for multispectral point cloud segmentation using KPConv across two test plots, and Figure
9 illustrates the segmentation results produced by KPConv. In both test plots, segmentation accuracy exceeds 70% across all classes,
except for branches in the first test plot. In Figure 9, each row depicts a segmented forest element, with colors indicating ground
truth labels. According to the achieved results, foliage, ground, low vegetation, and woody debris are segmented with high (more
than 92%) TP. Moreover, trunk points are adequately segmented as trunks (TP > 77%), so that trunks are correctly detected in
general. Most false positives in ground segmentation are due to misclassification as low vegetation. Typically, low points are
assumed to be ground points, and the above points can be low vegetation. Therefore, this error reflects the difficulty in
distinguishing the forest floor when it is obscured by low vegetation. In the first test plot, low vegetation is predominantly
misclassified as ground, while in the second plot, it is often confused with foliage. This is likely due to the similar spectral
characteristics of foliage and low vegetation, particularly when foliage appears at lower canopy levels. Trunks are frequently
misclassified as foliage or branches, as these elements usually emerge directly from the trunk. Branches are commonly confused
with foliage, and vice versa, due to their irregular spatial distribution among one another. Woody debris is mostly misclassified as
ground, followed by low vegetation, as all occur at comparable height levels.
 Figure 7. Multispectral forest point clouds semantic segmentation using different classifiers.

(a) (b)
Figure 8. Confusion matrices of the KPConv-based (the best classifier according to Figure 7) multispectral forest point clouds semantic
segmentation; (a) test plot 1, (b) test plot 2.
Ground
vegetation
Low
Trunks
Branches
Foliage
Woody debris

Figure 9. Visual results of multispectral LiDAR-based forest component segmentation using the best classifier. (a, c, e,
g, i, k) correspond to test plot 1, while (b, d, f, h, j, l) are associated with test plot 2. Point clouds of detected
components are colored based on the ground truth.
 5.2 Multispectral Data Effect
We investigated the effect of incorporating spectral information on forest semantic segmentation through an ablation study,
beginning with using all three available reflectance values in conjunction with NDVI NIR-SWIR as the vegetation index. The best-
performing model, KPConv (see Figure 7), is employed for this spectral ablation study, and the corresponding results are presented
in Table 3. Our experiments demonstrate that the incorporation of spectral information consistently improves the mAcc. While the
segmentation performance for low vegetation and foliage improved across all scenarios, the inclusion of the NIR wavelength in
the NIR and NIR + green combinations resulted in a marginal decline in ground segmentation accuracy, with a decrease of up to
0.31%. Trunk segmentation accuracy substantially improved with the addition of spectral information, except when SWIR and NIR
are used together, which resulted in a slight decline of 0.54%. In contrast, branch segmentation accuracy decreased by up to 1.66
% when using NIR, SWIR + NIR, and SWIR + green combinations. Coordinate-based information alone yielded satisfactory results
for detecting woody debris. Nevertheless, the synergistic use of SWIR and NIR, and especially the combination of all three
wavelengths (NIR, SWIR, and Green), significantly enhanced woody debris segmentation accuracy by as much as 49.43%. Among
all spectral configurations, the combination of all three wavelengths emerged as the most effective, yielding the highest overall
improvement across all classes, with gains of 33.09% in wIoU and 32.35% in mAcc. Conversely, the combination of SWIR and
green wavelengths resulted in the poorest spectral performance. Although it enhanced the mAcc by 4.06%, it led to a 2.35%
decrease in the wIoU.

Table 3. Spectral ablation study on forest semantic segmentation based on the best-performing model (KPConv).

Feature vector IoU (%) mIoU wIoU mAcc OA

(%) (%) (%) (%)
Ground Low Trunk Branches Foliage Woody
vegetation debris
Coordinates 92.54 25.41 60.87 5.64 88.59 38.15 51.87 50.45 57.08 89.82
+SWIR 92.70 28.46 64.87 11.12 88.92 26.60 52.11 50.42 63.20 89.67
+NIR 92.50 35.55 64.31 3.98 89.29 29.59 52.54 50.24 62.10 90.30
+Green 92.60 36.47 63.10 6.35 89.44 28.30 52.71 50.35 62.81 90.59
+SWIR + NIR 92.87 33.77 60.33 3.77 89.13 38.53 53.07 51.02 58.74 90.34
+SWIR + Green 92.81 27.90 63.19 4.47 88.77 23.73 50.15 48.10 61.14 89.95
+NIR + Green 92.23 40.19 65.08 7.55 89.80 22.75 52.93 50.28 64.08 90.66
+SWIR + NIR + Green 98.61 93.20 79.09 57.85 97.27 87.58 85.60 83.54 89.43 97.68
+SWIR +NIR +Green + VI 92.58 29.11 63.35 5.77 88.79 33.44 52.17 50.44 61.68 89.98

6. Discussion
The efficiency of different DL and ML models for forest semantic sgemnation across six classes compared to each other in this
study. Among the four classifiers implemented, KPConv surpassed the others with a 33.94% improvement in wIoU. Notably,
KPConv demonstrated strong performance in segmenting complex forest components, including branches and woody debris. This
aligns with findings by Robert et al. (2023), who reported that KPConv outperformed SPT in capturing intricate local geometries
and improved segmentation of classes such as poles, fences, and vegetation in urban environments—further supporting our results.
Nevertheless, KPConv demonstrated the poorest time efficiency, taking approximately ten times longer than SPT and PTv3. PTv3
outperformed SPT in segmenting trunk and woody debris classes. However, for the remaining classes, SPT achieved the second
highest score after KPConv model. It is worth noting that we made considerable efforts to fairly allocate time for hyperparameter
tuning across all models. However, there remains room for further improvement, particularly in identifying more optimal
hyperparameters for the SPT, PTv3, and RF models.

According to Table 3, incorporating additional spectral information does not necessarily lead to improved performance. Notably,
the combination of all three wavelengths with the vegetation index resulted in a lower wIoU compared to both the SWIR + NIR
combination and the scenario of using all three wavelengths. On the other hand, the mono-spectral SWIR and green wavelengths
showed greater accuracy than the bi-spectral combination of SWIR and green. Therefore, selecting the optimal spectral
configuration is tricky and depends on the specific target classes of interest. To visually examine the effect of spectral configuration
on forest semantic segmentation, we compare multispectral (including three wavelengths), bi-spectral, mono-spectral, and non-
spectral point clouds in two zoomed plots. The combination of SWIR and NIR and the use of SWIR alone are selected as the bi-
spectral and mono-spectral configurations, respectively, as they demonstrated the highest accuracy according to Table 3. As shown
in Figure 10, the transition to multispectral point clouds resulted in a noticeable reduction in segmentation errors. Remarkably, the
segmentation accuracy for ground, low vegetation, and foliage consistently improved. In both zoomed-in plots, while trunks are
accurately segmented across all configurations, the mono-spectral setup outperforms the bi-spectral and coordinate-only
configurations in segmenting branches and woody debris.

Figure 10. Visual comparison of multispectral (b and g), bi-spectral (SWIR and NIR; c and h), mono-spectral (SWIR; d and i), and
coordinates only (e and j) point cloud data for forest component segmentation.

7. Conclusions
Although tree species reflect light differently, our experiments demonstrate the significant potential of laser reflectance in three
wavelengths to enhance the semantic segmentation of forest point clouds by 33.09% in wIoU compared to solely using the
coordinates of point clouds. Notably, our results show that even using the single-spectral information of a mono-LiDAR system
boosts forest semantic segmentation, encouraging the deployment of laser reflectance, which has mostly received limited attention
in current LiDAR forestry studies. Since MS-LiDAR simultaneously captures both 3D spatial and spectral information, and
semantic segmentation serves as a foundational step in automated forest inventory, our proposed methodology presents a promising
avenue for precise, single-source, and automated forest management. By improving the segmentation accuracy of forest
components, MS-LiDAR data not only support more accurate and advanced measurements of biophysical attributes such as DBH,
stem curve, and branching structure, but also enable the estimation of biochemical properties, including leaf water content, forest
health, productivity status, and species classification.
Our proposed pipeline yields promising results for forest component segmentation, which can be further enhanced through post-
processing techniques. For example, the segmented trunks can be utilized as initial points for defining stem shape using well-
established geometrical methods. Besides, by detecting the trunks, the proposed framework streamlines the important and
challenging task of individual tree segmentation. Integrating laser scanners with precise and AI-powered forest segmentation
pipelines, like the one proposed, can significantly accelerate 3D forest inventory by reducing point cloud volume through the
removal of unwanted forest components as specified by the end user.

In this study, we considered six forest components: ground, low vegetation, trunks, branches, foliage, and woody debris. Particular
attention should be given to first-order branches in future research, as their location influences trunk and wood quality. In addition,
a limited number of point clouds are annotated and used for training and evaluation of the results. The generalizability of the trained
model on unseen datasets, as well as potential improvements using additional training data should be explored in future research.
Moreover, our published pre-trained deep learning model weights, presented for both mono- and multispectral LiDAR, can facilitate
future studies by serving as initial weights.

We utilized range-corrected reflectance values provided by RIEGL’s V-Line laser scanners. Although these reflectance values yield
promising results in forest semantic segmentation, it is imperative to explore robust, incidence angle–independent radiometric
calibration solutions, particularly for complex multi-target environments such as forest environments in future studies.
Incorporating radiometrically calibrated data may further enhance the accuracy and reliability of forest component segmentation
using MS-LiDAR data.

Acknowledgements
The research is funded by European Spatial Data Research (EuroSDR). Moreover, this project has received funding from the
European Union – NextGenerationEU instrument and is funded by the Research Council of Finland under grant number 353264,
and UNITE Flagship (337656, 348644, 359404) and Scan4est research infrastructure (346382).

References

Béland, M., Baldocchi, D., Widlowski, J.-L., Fournier, R., Verstraete, M., 2014: On seeing the wood from the leaves and the role
of voxel size in determining leaf area distribution of forests with terrestrial LiDAR. Agric. For. Meteorol., 184, 82–97.

Budei, B.C., St-Onge, B., Hopkinson, C., Audet, F.-A., 2017: Identifying the genus or species of individual trees using a three-
wavelength airborne LiDAR system. Remote Sens. Environ., 204, 632–647.

Calders, K., Newnham, G., Burt, A., Murphy, S., Raumonen, P., Herold, M., Culvenor, D., Avitabile, V., Disney, M., Armston, J.,
Kaasalainen, M., 2015: Nondestructive estimates of above-ground biomass using terrestrial laser scanning. Methods Ecol. Evol.,
6, 198–208.

Cote, J.-F., Widlowski, J.-L., Fournier, R., Verstraete, M., 2009: The structural and radiative consistency of three-dimensional tree
reconstructions from terrestrial lidar. Remote Sens. Environ., 113, 1067–1081.

Dai, W., Yang, B., Dong, Z., Shaker, A., 2018: A new method for 3D individual tree extraction using multispectral airborne LiDAR
point clouds. ISPRS J. Photogramm. Remote Sens., 144, 265–276.

Danson, F., Gaulton, R., Armitage, R., Disney, M., Gunawan, O., Lewis, P., Pearson, G., Ramirez, A., 2014: Developing a dual-
wavelength full-waveform terrestrial laser scanner to characterize forest canopy structure. Agric. Forest Meteorol., 198–199, 7–14.

Disney, M., 2018: Terrestrial LiDAR: a three‐dimensional revolution in how we look at trees. New Phytol., 222, 21–24.

Eitel, J.U.H., Magney, T.S., Vierling, L.A., Dittmar, G., 2014: Assessment of crop foliar nitrogen using a novel dual-wavelength
laser system and implications for conducting laser-based plant physiology. ISPRS J. Photogramm. Remote Sens., 97, 229–240.

Ferrara, R., Virdis, S., Ventura, A., Ghisu, T., Duce, P., Pellizzaro, G., 2018: An automated approach for wood-leaf separation from
terrestrial LIDAR point clouds using the density-based clustering algorithm DBSCAN. Agric. For. Meteorol., 262, 122–134.

Gaulton, R., Danson, F.M., Ramirez, F.A., Gunawan, O., 2013: The potential of dual-wavelength laser scanning for estimating
vegetation moisture content. Remote Sens. Environ., 132, 32–39.

Gaulton, R., Pearson, G., Lewis, P., Disney, M., 2010: The Salford Advanced Laser Canopy Analyser (SALCA): A multispectral
full waveform LiDAR for improved vegetation characterisation. Remote Sens., 2, 1079–1097.

Girardeau-Montaut, D., 2024: CloudCompare, v2.11.alpha. Available online: https://www.danielgm.net/cc/ (accessed on 18

October 2024).

Goodbody, T., Tompalski, P., Coops, N., Hopkinson, C., Treitz, P., van Ewijk, K., 2020: Forest Inventory and Diversity Attribute
Modelling Using Structural and Intensity Metrics from Multi-Spectral Airborne Laser Scanning Data. Remote Sens., 12, 2109.
Hakula, A., Ruoppa, L., Lehtomäki, M., Yu, X., Kukko, A., Kaartinen, H., Taher, J., Matikainen, L., Hyyppä, E., Luoma, V.,
Holopainen, M., Kankare, V., Hyyppä, J., 2023: Individual tree segmentation and species classification using high-density close-
range multispectral laser scanning data. ISPRS Open J. Photogramm. Remote Sens., 9, 100039.

Hancock, S., Lewis, P., Foster, M., Disney, M., Muller, J.-P., 2012: Measuring forests with dual wavelength lidar: A simulation
study over topography. Agric. For. Meteorol., 161, 123–133.

Heinzel, J., Huber, M.O., 2018: Constrained spectral clustering of individual trees in dense forest using terrestrial laser scanning
data. Remote Sens., 10(7), 1056.

Howe, G., Hewawasam, K., Douglas, E., Martel, J., Li, Z., Strahler, A., Schaaf, C., Cook, T., Chakrabarti, S., 2015: Capabilities
and performance of dual-wavelength Echidna® lidar. J. Appl. Remote Sens., 9, 095979.

Hu, Q., Yang, B., Xie, L., Rosa, S., Guo, Y., Wang, Z., Trigoni, N., Markham, A., 2019: RandLA-Net: Efficient semantic
segmentation of large-scale point clouds (CVPR 2020 Oral).

Huo, L., Lindberg, E., 2020: Individual tree detection using template matching of multiple rasters derived from multispectral
airborne laser scanning data. Int. J. Remote Sens., 41, 9525–9544.

Hyyppä, E., Yu, X., Kaartinen, H., Hakala, T., Kukko, A., Vastaranta, M., Hyyppä, J., 2020: Comparison of backpack, handheld,
under-canopy UAV, and above-canopy UAV laser scanning for field reference data collection in boreal forests. Remote Sens., 12,
3327.

Jaakkola, A., Hyyppä, J., Kukko, A., Yu, X., Kaartinen, H., Lehtomäki, M. and Lin, Y., 2010. A low-cost multi-sensoral mobile
mapping system and its feasibility for tree measurements. ISPRS J. of Photogramm. Remote Sens, 65(6), 514-522.

Junttila, S., Sugano, J., Vastaranta, M., Linnakoski, R., Kaartinen, H., Kukko, A., Holopainen, M., Hyyppä, H., Hyyppä, J., 2018:
Can leaf water content be estimated using multispectral terrestrial laser scanning? A case study with Norway spruce seedlings.
Front. Plant Sci., 9, 299.

Jupp, D.L., Culvenor, D., Lovell, J., Newnham, G., Strahler, A., Woodcock, C., 2009: Estimating forest LAI profiles and structural
parameters using a ground-based laser called “Echidna®”. Tree Physiol., 29, 171–181.

Kaijaluoto, R., Kukko, A., El Issaoui, A., Hyyppä, J., Kaartinen, H., 2022: Semantic segmentation of point cloud data using raw
laser scanner measurements and deep neural networks. ISPRS Open J. Photogramm. Remote Sens., 3, 100011.

Krisanski, S., Taskhiri, M.S., Gonzalez, S., Herries, D., Muneri, A., Gurung, M., Montgomery, J., Turner, P., 2021a: Forest structural
complexity tool—An open source, fully-automated tool for measuring forest point clouds. Remote Sens., 13(22), 4677.

Krisanski, S., Taskhiri, M.S., Gonzalez, S., Herries, D., Turner, P., 2021b: Sensor agnostic semantic segmentation of structurally
diverse and complex forest point clouds using deep learning. Remote Sens., 13, 1413.

Landrieu, L., Simonovsky, M., 2018: Large-scale point cloud semantic segmentation with superpoint graphs. CVPR,
10.1109/CVPR.2018.00479.

Li, W., Guo, Q., Jakubowski, M.K., Kelly, M., 2012: A new method for segmenting individual trees from the lidar point cloud.
Photogramm. Eng. Remote Sens., 78, 75–84.

Li, Z., Douglas, E., Strahler, A., Schaaf, C., Yang, X., Wang, Z., Yao, T., Zhao, F., Saenz, E., Paynter, I., Woodcock, C., Chakrabarti,
S., Cook, T., Martel, J., Howe, G., Jupp, D., Culvenor, D., Newnham, G., Lovell, J., 2013: Separating leaves from trunks and
branches with dual-wavelength terrestrial lidar scanning. Proc. IEEE Int. Geosci. Remote Sens. Symp., 3383–3386.

Li, Z., Schaefer, M., Strahler, A., Schaaf, C., Jupp, D., 2018: On the utilization of novel spectral laser scanning for three-dimensional
classification of vegetation elements. Interface Focus, 8, 20170039.

Liang, X., Kankare, V., Hyyppä, J., Wang, Y., Kukko, A., Haggrén, H., Yu, X., Kaartinen, H., Jaakkola, A., Guan, F., Holopainen,
M., Vastaranta, M., 2016: Terrestrial laser scanning in forest inventories. ISPRS J. Photogramm. Remote Sens., 115, 36–45.

Liang, X., Qi, H., Deng, X., Chen, J., Cai, S., Zhang, Q., Wang, Y., Kukko, A., Hyyppä, J., 2024: ForestSemantic: A dataset for
forest semantic learning from close-range sensing. Geo-spatial Inf. Sci..

Lindberg, E., Holmgren, J., Olofsson, K., Olsson, H., 2012: Estimation of tree lists from airborne laser scanning by combining
single-tree and area-based methods. Int. J. Remote Sens., 33, 8–19.

Lovell, J., Jupp, D.L.B., Culvenor, D.S., Coops, N., 2003: Using airborne and ground-based ranging lidar to measure canopy
structure in Australian forests. Can. J. Remote Sens., 29, 607–615.
Lu H., Li B., Yang G., Fan G., Wang H., Pang Y., Wang Z., Lian Y., Xu H., Huang H., 2025: Towards a point cloud understanding
framework for forest scene semantic segmentation across forest types and sensor platforms. Remote Sens. Environ., 318, 114591.
Luo, B., Yang, J., Song, S., Shi, S., Gong, W., Wang, A., Du, L., 2022: Target classification of similar spatial characteristics in
complex urban areas by using multispectral LiDAR.

Ma, L., Zheng, G., Eitel, J.U., Moskal, L.M., He, W., Huang, H., 2016: Improved salient feature-based approach for automatically
separating photosynthetic and nonphotosynthetic components within terrestrial lidar point cloud data of forest canopies. IEEE
Trans. Geosci. Remote Sens., 54(2), 679–696.

MacArthur, A., MacDonald, P., 2021: Spectral information recovery from waveform LiDAR using optimized wavelet
decomposition parameters. Remote Sens. Environ., 264, 112598.

Maltamo, M., Räty, J., Korhonen, L., Kotivuori, E., Kukkonen, M., Peltola, H., Kangas, J., Packalen, P., 2020: Prediction of forest
canopy fuel parameters in managed boreal forests using multispectral and unispectral airborne laser scanning data and aerial images.
Eur. J. Remote Sens., 53, 245–257.

Mandlburger, G., Otepka, J., Wagner, W., Karel, W., Pfeifer, N., 2009: Orientation and processing of airborne laser scanning data
(OPALS) - Concept and first results of a comprehensive ALS software.

Mielczarek, D., Sikorski, P., Archiciński, P., Ciężkowski, W., Zaniewska, E., Chormanski, J., 2022: The use of an airborne laser
scanner for rapid identification of invasive tree species Acer negundo in riparian forests. Remote Sens., 15, 212.

Morel, J., Bac, A., Kanai, T., 2020: Segmentation of unbalanced and in-homogeneous point clouds and its application to 3D scanned
trees. Vis. Comput., 36, 2419–2431.

Newnham, G., Armston, J., Calders, K., Jupp, D.L.B., 2015: Terrestrial laser scanning for plot-scale forest measurement. Curr. For.
Rep., 1, 239–251.

Ni-Meister, W., Jupp, D., Dubayah, R., 2001: Modeling lidar waveforms in heterogeneous and discrete canopies. IEEE Trans.
Geosci. Remote Sens., 39, 1943–1958.

Pan, S., Guan, H., Yu, Y., Li, J., Peng, D., 2019: A comparative land-cover classification feature study of learning algorithms: DBM,
PCA, and RF using multispectral LiDAR data. IEEE J. Sel. Topics Appl. Earth Obs. Remote Sens.

Pehkonen, M., Vastaranta, M., Holopainen, M., Hyyppä, J., Kukko, A., Pyörälä, J., 2025: Identification and segmentation of branch
whorls and sawlogs in standing timber using terrestrial laser scanning and deep learning. Forestry: Int. J. Forest Res., cpaf006.

Pfeifer, N., Mandlburger, G., Otepka, J., Karel, W., 2014: OPALS—a framework for airborne laser scanning data analysis. Comput.
nviron. Urban. Syst., 45, 125–136.

Pfennigbauer, M., Ullrich, A., 2010: Improving quality of laser scanning data acquisition through calibrated amplitude and pulse
deviation measurement. Proc. SPIE, 7684, 10.1117/12.849641.

PreMergeChannelsPointclouds Module, OPALS, TU Wien, 2024: Available online:

https://opals.geo.tuwien.ac.at/html/nightly/preMergeChannelsPointclouds.html. (Last access: 22 Apr 2025).

Qi, C.R., Yi, L., Su, H., Guibas, L.J., 2017: PointNet++: deep hierarchical feature learning on point sets in a metric space. Proc.
31st Int. Conf. Neural Inf. Process. Syst. (NIPS'17), Curran Associates Inc., Red Hook, NY, USA, 5105–5114.

Raumonen, P., Kaasalainen, M., Åkerblom, M., Kaasalainen, S., Kaasalainen, T., 2013: Automatic reconstruction of tree models
from terrestrial laser scanner data. ISPRS J. Photogramm. Remote Sens., 88, 144–151.

Reichler, M., Taher, J., Manninen, P., Kaartinen, H., Hyyppä, J., Kukko, A., 2024: Semantic segmentation of raw multispectral laser
scanning data from urban environments with deep neural networks. ISPRS Open J. Photogramm. Remote Sens., 12, 100061.

Robert, D., Raguet, H., Landrieu, L., 2023: Efficient 3D semantic segmentation with Superpoint Transformer. Proc. IEEE/CVF Int.
Conf. Comput. Vis. (ICCV), 2023, 1785–1795.

Salas, C., Ene, L., Gregoire, T., Næsset, E., Gobakken, T., 2010: Modelling tree diameter from airborne laser scanning derived
variables: A comparison of spatial statistical models. Remote Sens. Environ., 114, 1277–1285.

Shao, J., Habib, A., Fei, S., 2023: Semantic segmentation of UAV LiDAR data for tree plantations. Int. Arch. Photogramm. Remote
Sens. Spatial Inf. Sci., XLVIII-1/W2-2023, 1901–1906.

St-Onge, B., Vega, C., Fournier, R., Hu, Y., 2008: Mapping canopy height using a combination of digital stereo-photogrammetry
and lidar. Int. J. Remote Sens., 29, 3343–3366.
Takhtkeshha, N., Mandlburger, G., Remondino, F., Hyyppä, J., 2024a: Multispectral light detection and ranging technology and
applications: A review. Sensors, 24(5), 1669.

Takhtkeshha, N., Bayrak, O.C., Mandlburger, G., Remondino, F., Kukko, A., Hyyppä, J., 2024b: Automatic annotation of 3D
multispectral LiDAR data for land cover classification. IGARSS 2024 – 2024 IEEE Int. Geosci. Remote Sens. Symp., 8645–8649.

Tanaka, T., Yamaguchi, J., Takeda, Y., 1998: Measurement of forest canopy structure with a laser plane range-finding method—
development of a measurement system and applications to real forests. Agric. For. Meteorol., 91(3–4), 149–160.

Thomas, H., Qi, C.R., Deschaud, J.-E., Marcotegui, B., Goulette, F., Guibas, L., 2019: KPConv: Flexible and Deformable
Convolution for Point Clouds. Proc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), Seoul, Korea (South), 6410–6419.

Ting, Y., An, F., Li, W., Sun, Y., Cao, L., Xue, L., 2016: A novel approach for retrieving tree leaf area from ground-based LiDAR.
Remote Sens., 8, 942.

Vormisto, J., Tuomisto, H., Oksanen, J., Ruokolainen, K., 2000: Palm distribution patterns in Amazonian rainforests are explained
by habitat heterogeneity and dispersal limitation. J. Ecol., 88, 939–951.

Walker, I., Ferreira, M.J.N., 1985: On the responses of the shrimp Macrobrachium amazonicum to pollution, hydrographic aspects,
and chemical factors. Amazoniana, 9, 411–415.

Wallace, A., McCarthy, A., Nichol, C., Ren, X., Morak, S., Martinez-Ramirez, D., Woodhouse, I., Buller, G., 2014: Design and
evaluation of multispectral LiDAR for the recovery of arboreal parameters. IEEE Trans. Geosci. Remote Sens., 52, 10.

Wallace, L., Lucieer, A., Watson, C., Turner, D., 2012: Development of a UAV-LiDAR system with application to forest inventory.
Remote Sens., 4, 1519–1543.

Wang, D., 2020: Unsupervised semantic and instance segmentation of forest point clouds. ISPRS J. Photogramm. Remote Sens.,
165, 86–97.

Wang, D., Brunner, J., Ma, Z., Lu, H., Hollaus, M., Pang, Y., Pfeifer, N., 2018: Separating tree photosynthetic and non-
photosynthetic components from point cloud data using dynamic segment merging. Forests, 9, 252.

Wang, D., Hollaus, M., Pfeifer, N., 2017: Feasibility of machine learning methods for separating wood and leaf points from
terrestrial laser scanning data. ISPRS Ann. Photogramm. Remote Sens. Spatial Inf. Sci., IV-2/W4, 157–164.

Wang, Y., Sun, Y., Liu, Z., Sarma, S.E., Bronstein, M.M., Solomon, J.M., 2019: Dynamic Graph CNN for Learning on Point Clouds.
ACM Trans. Graph., 38, 5, 146.

Whitmore, T.C., 1984: Tropical rainforests of the Far East. Oxford University Press, Oxford, UK.

Wilson, E. O. (1999). The diversity of life. W. W. Norton & Company, New York, NY, USA.

Winberg, O., Pyörälä, J., Yu, X., Kaartinen, H., Kukko, A., Holopainen, M., Holmgren, J., Lehtomäki, M., Hyyppä, J., 2023: Branch
information extraction from Norway spruce using handheld laser scanning point clouds in Nordic forests. ISPRS Open J.
Photogramm. Remote Sens., 9, 100040.

Windrim, L., Bryson, M., 2020: Detection, segmentation, and model fitting of individual tree stems from airborne laser scanning
of forests using deep learning. Remote Sensing, 12, 1469.

Wright, S.J., 2005: Tropical forests in a changing environment. Trends Ecol. Evol., 20, 553–560.

Wu, X., Jiang, L., Wang, P.-S., Liu, Z., Liu, X., Qiao, Y., Ouyang, W., Tong, H., Zhao, H., 2024: Point Transformer V3: Simpler,
faster, stronger. CVPR, 4840–4851.

Xiang, B., Wielgosz, M., Kontogianni, T., Peters, T., Puliti, S., Astrup, R., 2024: Automated forest inventory: Analysis of high-
density airborne LiDAR point clouds with 3D deep learning. Remote Sens. Environ., 305, 114078.

Zagt, R.J., Werger, M.J.A., 1998: Community structure and the demography of primary forest species in relation to secondary
succession in tropical rainforests of Guyana. Trop. Forest Resour. Dynam. Conserv., 140, 59–84.

Zhao, H.S., Jiang, L., Jia, J.Y., Torr, P., Koltun, V., 2021: Point Transformer. In 2021 IEEE/CVF International Conference on
Computer Vision (ICCV), Montreal, QC, Canada, 16239–16248.

Zhu, X., Skidmore, A.K., Darvishzadeh, R., Niemann, K.O., Liu, J., Shi, Y., Wang, T., 2018: Foliar and woody materials