Forest Stem Extraction and Modeling (FoSEM): A LiDAR-Based Framework for Accurate Tree Stem Extraction and Modeling in Radiata Pine Plantations

Abstract

Accurate characterization of tree stems is critical for assessing commercial forest health, estimating merchantable timber volume, and informing sustainable value management strategies. Conventional ground-based manual measurements, although precise, are labor-intensive and impractical at large scales, while remote sensing approaches using satellite or UAV imagery often lack the spatial resolution needed to capture individual tree attributes in complex forest environments. To address these challenges, this study provides a significant contribution by introducing a large-scale dataset encompassing 40 plots in Western Australia (WA) with varying tree densities, derived from Hovermap LiDAR acquisitions and destructive sampling. The dataset includes parameters such as plot and tree identifiers, DBH, tree height, stem length, section lengths, and detailed diameter measurements (e.g., DiaMin, DiaMax, DiaMean) across various heights, enabling precise ground-truth calibration and validation. Based on this dataset, we present the Forest Stem Extraction and Modeling (FoSEM) framework, a LiDAR-driven methodology that efficiently and reliably models individual tree stems from dense 3D point clouds. FoSEM integrates ground segmentation, height normalization, and K-means clustering at a predefined elevation to isolate stem cores. It then applies circle fitting to capture cross-sectional geometry and employs MLESAC-based cylinder fitting for robust stem delineation. Experimental evaluations conducted across various radiata pine plots of varying complexity demonstrate that FoSEM consistently achieves high accuracy, with a DBH RMSE of 1.19 cm (rRMSE = 4.67%) and a height RMSE of 1.00 m (rRMSE = 4.24%). These results surpass those of existing methods and highlight FoSEM’s adaptability to heterogeneous stand conditions. By providing both a robust method and an extensive dataset, this work advances the state of the art in LiDAR-based forest inventory, enabling more efficient and accurate tree-level assessments in support of sustainable forest management.

1. Introduction

Forest resources in the southwest of Western Australia hold significant commercial value, primarily driven by the timber industry’s demand for high-quality wood products [1,2]. Radiata pine plantations are extensively cultivated to supply essential raw materials for construction, packaging, and manufacturing sectors, thereby underpinning both local and national economic growth [3,4,5]. Accurate and efficient inventory management of these commercial forests is crucial for optimizing resource utilization, planning sustainable harvesting schedules, and maximizing economic returns [6,7]. Detailed knowledge of tree stem parameters, such as diameter at breast height (DBH), tree height, and stem curvature, is fundamental for assessing timber quality, determining suitable harvestable volumes, and forecasting future yields [8,9,10]. Furthermore, precise forest inventories enable the identification of optimal harvesting strategies, ensuring that wood products meet industry standards while maintaining the health and productivity of the plantations [11,12]. Therefore, the development of scalable and accurate methodologies for tree stem modeling is indispensable for supporting the commercial objectives of radiata pine plantations, enhancing the efficiency of forest management practices, and sustaining the economic viability of the timber sector in the region.

Traditional methods for forest tree data collection include primarily ground-based manual measurements, and UAV- or satellite-based measurements. Manual ground measurements involve direct measurements of tree parameters such as the diameter at breast height (DBH), tree height, and canopy spread using tools such as clinometers, calipers, and measuring tapes [13,14,15,16]. Although this approach allows for the extraction of detailed information, such as wood volume and stems per hectare (stocking), it is inherently limited by several factors. Moreover, manual measurements are relatively labor intensive, time-consuming, and costly, particularly when applied to large forest areas [17,18,19,20,21]. In addition, challenging terrains and remote locations further restrict the feasibility of this method. To address manual measurements’ inefficiencies, satellite imagery and unmanned aerial vehicles (UAVs) have been widely adopted for forest monitoring [17,22,23,24]. These methods utilize multispectral, hyperspectral, or LiDAR sensors mounted on satellites or UAVs to capture imagery or point cloud data over extensive forested regions, providing a broad spatial perspective. However, despite these advantages, these techniques have significant limitations. First, satellite imagery suffers from limited spatial resolution, making it challenging to identify and measure individual tree attributes, especially in complex forest environments [23,24,25]. Second, the predominantly vertical perspective of remote sensing imagery does not capture horizontal cross-sectional details of tree stems, leading to a reduced precision in estimating parameters such as DBH [22,26,27,28]. All these factors impact its operational efficiency in large-scale forest monitoring.

Recently, the handheld LiDAR device has emerged as a powerful tool for capturing high-resolution 3D point cloud data in forested environments [29,30,31,32,33,34]. Its adaptability allows it to function as a handheld device, making it particularly effective for navigating the diverse and complex terrains commonly found in forested areas [31,32,33,34]. However, leveraging raw 3D point cloud data for tree stem modeling and parameter estimation in such challenging environments presents significant obstacles. The heterogeneity of vegetation types and the rugged, uneven terrain of Western Australia contribute to inaccuracies in plant height measurements [35,36,37]. Additionally, the high density, irregular spatial distribution, and significant size variability of trees introduce substantial noise and occlusion within the point cloud data [36,38]. These complexities make individual tree point cloud extraction and stem and crown segmentation more challenging and hinder the accurate extraction of stem features [30,39]. Consequently, existing methods often struggle to achieve the desired level of accuracy and efficiency when processing such complex datasets. Existing methods like Terrestrial Laser Scanning (TLS) and Handheld Mobile Laser Scanning (HMLS) provide advancements, yet they struggle with scalability, motion-induced errors, and computational demands [37,40]. Advanced data processing approaches, including adaptive filtering [41] and machine learning-based segmentation [42,43,44], improve accuracy but face limitations in heterogeneous forest environments and reliance on extensive training data.

To overcome these challenges, this research contributes a large-scale dataset of radiata pine forests in Western Australia (WA), specifically tailored for tree stem analysis. These data were acquired using a handheld LiDAR scanner under rigorous field protocols, which involved standardized methods and consistent data collection practices to ensure accuracy and reliability. The dataset includes 40 plots across four age classes (0–37 years). DBH and tree height were measured at each plot using a diameter tape and a Haglöf Vertex. Destructive sampling was performed on five plots, with 3 trees per plot (15 trees total) felled for detailed post-felling measurements, including section diameters and lengths. These measurements provided detailed diameter–height profiles and stem curvature metrics, enabling robust model calibration and validation. The dataset is characterized by high spatial resolution, sub-meter positional accuracy, and extensive stem attribute coverage, rendering it an invaluable resource for developing, benchmarking, and refining LiDAR-based stem modeling methods.

This study also presents the Forest Stem Extraction and Modeling (FoSEM) framework, a novel LiDAR-driven methodology designed for the high-precision extraction and characterization of individual tree stems from dense 3D point cloud data (as shown in Figure 1). Leveraging the high-resolution data captured by the handheld device, FoSEM integrates advanced processing techniques to address the limitations of existing approaches. The framework begins by segmenting the ground and normalizing terrain heights to mitigate topographical distortions, ensuring consistent reference heights. Topographical distortions refer to variations in elevation caused by uneven terrain or slopes, which can introduce inaccuracies when analyzing tree structures; therefore, mitigating these distortions is crucial to establish a reliable and uniform baseline for height measurements. It then applies K-means clustering at a predetermined elevation (5.6 m), determined using statistical methods to delineate core stem regions, followed by circle-fitting techniques to extract optimal cross-sectional geometry. Finally, the integration of these refined cross-sections with robust cylinder-fitting algorithms enables the selective extraction of true stem points and the construction of precise 3D models, providing a robust solution to the complexities of tree stem modeling in challenging forest environments.

Experimental evaluations conducted across forest plots of varying complexity—ranging from low-density stands to dense, irregularly spaced canopies—demonstrate the effectiveness of FoSEM. The framework demonstrates robust performance in quantifying critical stem parameters, achieving a Root-Mean-Square Error (RMSE) of 1.19 cm for DBH and 1.00 m for tree height, with relative RMSE (rRMSE) values of 4.67% and 4.24%, respectively, indicating reliable overall accuracy. Comparative analyses against existing state-of-the-art methods highlight FoSEM’s strengths in densely stocked, unthinned stands and in predicting diameters higher up the stem, demonstrating competitive performance despite variations in DBH and tree height accuracy across different methods. These results highlight FoSEM’s reliability and accuracy, establishing it as a robust tool for forest resource management and a valuable asset for informed decision-making in commercial forestry investments.

2. Related Work

Accurate tree stem measurement is critical for timber volume estimation, forest health assessment, and management. Methods range from traditional manual measurements to remote sensing (satellite and UAV), and advanced 3D point cloud processing. This section provides a review of these approaches, highlighting their contributions and challenges in the context of forested environments.

2.1. Traditional Ground-Based Manual Measurements

Traditional ground-based manual measurements have been fundamental to forest inventory and management, providing precise and reliable data on parameters like diameter at breast height (DBH), tree height, and canopy spread using tools such as calipers and clinometers [14,15,18,19,20,21]. These methods, exemplified by the accurate DBH measurements demonstrated in [21], are crucial for applications such as timber volume estimation. While these methods are effective for basic parameters such as DBH and height, they become particularly labor-intensive and time-consuming when measuring more complex attributes like stem curvature, often requiring invasive procedures that may disturb the environment [14,18]. In contrast, our approach provides a non-invasive, efficient solution that streamlines data collection and minimizes environmental impact, particularly in dense or challenging terrains.

2.2. Satellite-Based Remote Sensing

Satellite-based remote sensing and UAV-based scanning are two critical methods in forest monitoring, each with distinct advantages and limitations. Satellite-based remote sensing relies on sensors mounted on satellites to capture imagery and spectral data, offering extensive spatial coverage and enabling periodic monitoring of vast forest landscapes [22,23,24,25,27,28]. This approach is particularly suitable for macro-scale studies, such as estimating canopy height and timber volume [25] or mapping forest structures and health indicators [28]. However, the coarse spatial resolution of many satellite sensors can hinder the accurate detection of individual tree stems in dense forests with overlapping canopies [45,46]. Atmospheric conditions like cloud cover and haze further challenge data acquisition reliability [27]. In contrast, UAV-based scanning provides higher spatial resolution and greater flexibility, utilizing sensors such as high-resolution cameras and LiDAR systems to capture detailed data along customizable flight paths [22,23,24]. This method allows real-time data collection and rapid deployment for specific needs, such as analyzing canopy structures or estimating DBH [24]. Despite its strengths, UAV-based scanning faces operational constraints, including limited battery life, weather sensitivity, and the necessity of maintaining a direct line of sight in large or obstructed environments [17,23]. Additionally, its predominantly vertical perspectives may not adequately capture the horizontal geometry of tree stems, limiting its effectiveness in certain applications [47].

Ground-based methods, including Terrestrial Laser Scanning (TLS) and Handheld Mobile Laser Scanning (HMLS), complement remote sensing techniques by offering high-resolution 3D data for detailed forest analysis. TLS uses stationary setups to generate precise measurements of DBH, stem height, and volume, creating accurate 3D models for studying tree growth and structural dynamics [48,49]. However, its reliance on line-of-sight access, labor-intensive setup, and high equipment costs limit its practicality for large-scale surveys, particularly in dense or rugged terrains [40]. HMLS systems, such as backpack-mounted or handheld LiDAR devices, provide greater mobility and adaptability, enabling dynamic data collection across challenging environments like steep slopes and dense underbrush [35,36]. These systems facilitate rapid deployment and comprehensive surveys, as demonstrated in rugged landscapes [36]. Nevertheless, HMLS is susceptible to motion-induced errors, requiring sophisticated data processing and alignment techniques to ensure accuracy [37]. Like TLS, the high cost of equipment and logistical challenges in remote deployments, coupled with battery life limitations and physical strain on operators, pose significant barriers to its widespread adoption [35,36]. Both TLS and HMLS are invaluable for detailed analyses, but their limitations underscore the need for integrating multiple methods to address the diverse demands of forest monitoring.

2.3. LiDAR 3D Point Cloud Data Processing Methods

Effective processing of high-resolution LiDAR point clouds is essential for accurate tree stem modeling and parameter extraction. This involves tackling challenges such as noise, occlusions, and complex vegetation structures through advanced methodologies. Adaptive filtering techniques dynamically adjust filtering parameters based on local point density and vegetation characteristics, improving tree segmentation accuracy by mitigating noise and irrelevant points [41,50,51,52]. For instance, adaptive morphological filters have been used to distinguish ground and non-ground points in dense forests, enhancing segmentation precision [41]. Machine learning (ML) and deep learning (DL) algorithms, such as convolutional neural networks (CNNs) and Random Forests, have further automated tree detection and parameter estimation [42,43,44]. These approaches leverage geometric attributes of point clouds for tasks like DBH estimation and stem segmentation, achieving high accuracy but requiring extensive training datasets and computational resources [53,54,55,56]. Hybrid methods combine traditional geometric algorithms with ML techniques to improve accuracy and efficiency in data processing. For example, K-means clustering integrated with CNN-based classification enhances tree segmentation in dense forests, though these methods can introduce complexity and computational demands [57,58].

Advanced geometric and statistical methods complement these approaches by focusing on precise stem modelings [59,60]. Techniques like Voronoi tessellation, RANSAC-based segmentation, and surface fitting algorithms have improved segmentation and geometric representation in challenging forest environments [61,62,63]. Voronoi tessellation partitions point clouds into regions associated with individual trees, while RANSAC algorithms robustly extract stem geometries even in noisy datasets [63,64]. Surface fitting methods, including least squares fitting and spline interpolation, enable detailed modeling of stem curvature and tapering [65,66]. Despite their effectiveness, these methods are influenced by input data quality and occlusions, often requiring significant computational resources for large-scale surveys [65,66,67]. Together, these diverse methodologies highlight the potential of LiDAR data processing while underscoring the need for adaptable and efficient solutions tailored to specific forest conditions.

In summary, traditional ground-based methods are precise but impractical for large-scale surveys, while remote sensing offers scalability but struggles with resolution in dense forests. LiDAR techniques like TLS and HMLS capture detailed data but face logistical and computational challenges. Integrating HMLS mobility with advanced algorithms is key to overcoming noise, occlusions, and segmentation issues for efficient, scalable tree stem modeling in complex forests.

3. Materials and Methods

This section of the paper details the study sites, Hovermap data acquisition, field data and ground-truth generation, and the methodologies employed.

3.1. Study Sites

This study was conducted in two Pinus radiata plantations in WA, encompassing a total of 40 plots: site 1 with 18 plots and site 2 with 22 plots, as illustrated in Figure 2a. The plots exhibited varying stem densities, ranging from 100 to 1250 stems per hectare. Each plot had a radius of 11.28 m, ensuring uniform sampling areas and providing a standardized framework for data collection and analysis, as shown in Figure 2b. Both study sites offer optimal growing conditions for Pinus radiata, characterized by a Mediterranean climate with mild, wet winters and hot, dry summers, as well as well-drained, fertile soils. These favorable conditions, combined with effective forest management practices, make these plantations commercially significant. They produce high-quality, fast-growing timber, primarily for construction purposes, including veneer and engineered wood products. Furthermore, the proximity of these sites to processing facilities and transportation infrastructure enhances their economic value.

3.2. Field Data and Ground-Truth Generation

Accurate and detailed ground-truth data are essential for validating LiDAR-based tree stem modeling methods. To achieve this, a total of 40 circular plots were selected, distributed across four age classes: (i) 0–8 years, (ii) 12–18 years, (iii) 27–34 years, and (iv) 35–37 years. At each plot, diameter at breast height (DBH) and tree heights were measured for all standing trees. DBH was recorded using a diameter tape, while heights were measured with a Haglöf Vertex. Among the 40 plots, 5 were selected for destructive sampling, with 3 trees per plot (a total of 15 trees) felled for detailed post-felling measurements. Three trees, aged 8, 18, and 34 years, were selected from the fifteen felled specimens to derive performance metrics for the model developed in this study. Measurements of DBH and height taken before felling, along with section diameters and lengths recorded post-felling, were used for model validation.

The destructive sampling process involved cutting selected trees as close to ground level as safely possible using a harvesting machine. The trunks were segmented into manageable sections for precise measurement. Diameters were recorded at consistent intervals: every 2 m between 3 m and 15 m height, every 3 m above that, and at specific points (0.65 m, 1.3 m, and 2 m) near the base. These measurements extended beyond the merchantable portion of the stem to the apex of the crown. A diameter tape or calipers and a measuring staff ensured accuracy and repeatability. The total tree height was recorded as the length from the stump cut to the crown apex along the stem’s longitudinal axis. For each segment, the position along the stem was noted, enabling the construction of a detailed diameter–height profile for the entire trunk. This detailed dataset documented not only DBH but also stem height, curvature, and diameter changes at various heights (as summarized in

Table 1. Descriptive statistics of 40 Pinus radiata plots: This table provides a summary of tree characteristics, including diameter at breast height (DBH) and tree height, across different age classes of Pinus radiata. The columns detail the number of plots, tree count range, mean tree count, DBH range and mean (mm), and height range and mean (m).

Species	Age-Class (Year)	#Plots	#Trees		DBH (mm)		Height (m)
			Range	Mean	Range	Mean	Range	Mean
Pinus radiata	0–8	2	45–50	48	56–200	131.63	9.2–16.4	13.55
Pinus radiata	12–18	14	19–26	24	86–386	219.24	9.2–28.9	21.9
Pinus radiata	27–34	14	9–22	16	131–542	359.1	19.1–42.7	30.94
Pinus radiata	35–37	10	6–16	8	300–724	501.33	24–42.7	34.36

). These high-precision ground-truth data served as empirical benchmarks for validating the LiDAR-derived tree models.

By comparing the measured diameters and heights of the felled trees to the estimates derived from the LiDAR point clouds, the accuracy, robustness, and consistency of the proposed methodology were assessed. Figure 3 illustrates the destructive sampling and measurement process, highlighting the attention to detail during manual data collection and the nature of the ground-truth dataset. These data underpinned the subsequent validation and refinement of the LiDAR-based stem modeling approach, ensuring a rigorous evaluation of its performance.

3.3. Hovermap Data Collection

The Emesent Hovermap was selected as the primary data acquisition tool for this study due to its advanced SLAM (Simultaneous Localization and Mapping) capabilities, which operate independently of GNSS signal availability. This feature enables the generation of coherent, high-quality point clouds in GNSS-denied environments, such as dense forests [29,31]. The Hovermap employs a Velodyne Puck-LITE (VLP-16) laser scanner (Velodyne LiDAR, Inc., San Jose, CA, USA), providing full 360° spherical coverage and capturing up to 300,000 points per second [29,31]. Its versatile deployment options, including handheld and backpack-mounted configurations, facilitate efficient navigation in complex terrains (as shown in Figure 4).

For this study, the handheld configuration was employed, allowing precise data acquisition while navigating through the plots. Data collection began at the center of each of the 40 circular plots, with the operator walking in a circular trajectory. After completing one circle, the radius was incrementally increased, and the process was repeated until the entire plot was covered. This approach ensured complete coverage of each plot. Data collection was conducted under calm weather conditions to minimize the influence of dynamic elements, such as swaying treetops, on data quality. Figure 2b illustrates the trajectory generated during the data collection process. Over 8000 frames of LiDAR data were recorded across all plots, resulting in a dataset exceeding 50 GB in size. Each 26 m × 26 m plot was fully covered, with point cloud densities averaging over 58,000 points per square meter. Additionally, precise differential GPS (dGPS) locations for each plot center ensured sub-meter spatial accuracy. Ground Control Points (GCPs), clearly visible in the point clouds, were used to maintain precise orientation and alignment, ensuring spatial consistency across all datasets (as shown in Figure 2c).

These features, combined with rigorous post-processing steps such as noise reduction, down-sampling, and the SLAM algorithm’s “loop closure” process, enhanced the trajectory’s accuracy by revisiting previously scanned areas to improve local consistency. This data collection methodology ensured the reliability of the data for modeling and analysis. Previous studies have further validated the effectiveness of handheld SLAM-based LiDAR systems in forest environments, highlighting their suitability for capturing detailed structural information beneath dense forest canopies [29,31,32,34].

3.4. Methods

This part delineates the complete methodological framework utilized for the extraction and analysis of tree stems and calculating the tree stem features from LiDAR-derived 3D point cloud data. Building upon the preprocessed datasets described in the previous section, the methodology is structured into three primary components: (1) data preprocessing, (2) tree extraction and segmentation, (3) stem point extraction, and (4) feature extraction and results validation (as shown in Figure 5).

3.4.1. Data Preprocessing

The raw LiDAR data collected using the Hovermap underwent a preprocessing pipeline to prepare the datasets for subsequent tree stem modeling. This preprocessing workflow was meticulously designed to enhance data quality, reduce computational demands, and ensure the accuracy of the extracted forest metrics.

Initially, the raw point cloud data were down-sampled using a voxel grid filter. This process involved partitioning the point cloud into a three-dimensional grid of voxels and replacing all points within each voxel with a single representative point, typically the centroid [68]. By reducing the density of the point cloud, the voxel grid filter significantly decreased the data volume, thereby optimizing processing time and computational resources while maintaining the structural integrity of the forest canopy and tree stems [69].

Following down-sampling, ground segmentation was performed using the Simple Morphological Filter (SMRF) algorithm. The SMRF effectively separates ground points from non-ground points by analyzing local elevation variations within the point cloud [41]. The algorithm applies morphological operations to construct a minimum elevation surface map, facilitating the accurate distinction of terrain points from vegetation and structural elements. Specifically, the SMRF can be mathematically represented as:

$$Z_{\mathrm{filtered}}(x,y)=\underset{(i,j)\in \mathcal{W}}{min}\{Z(x+i,y+j)+\alpha ·h(i,j)\}$$

(1)

where $Z_{\mathrm{filtered}}(x,y)$ denotes the filtered ground elevation at coordinates $(x,y)$, $\mathcal{W}$ represents the structuring element defining the neighborhood window, $\alpha$ is a scaling factor controlling the influence of elevation offsets, and $h(i,j)$ is a height offset function applied within the window. This operation effectively captures the ground surface by minimizing the elevation values within the defined window, accounting for local terrain variability.

After removing ground points identified by the SMRF algorithm, the dataset was confined to vegetation and stem points, facilitating the focused analysis of the tree structure. Next, to standardize vertical metrics across differing terrains, elevation normalization was applied. Specifically, for each point with measured elevation $Z(x,y)$, the corresponding ground elevation $Z_g(x,y)$ was subtracted to yield a normalized height $h(x,y)=Z(x,y)-Z_g(x,y)$. By providing a consistent reference frame that accounted for terrain variability, this normalization ensured that tree height and canopy attributes accurately reflected true vertical vegetation distribution across all plots. Subsequently, the processed point clouds were indexed and organized according to their geographic coordinates and assigned plot identifiers, thereby streamlining data retrieval and ensuring spatial coherence. This meticulous preparation of each experimental plot’s dataset fostered reliable and efficient downstream modeling of tree stems.

Finally, the preprocessed point clouds were organized and indexed based on geographic coordinates and plot identifiers. This organization allowed for the efficient access and management of the datasets during the tree stem modeling phase. Each experimental plot’s data were meticulously prepared, ensuring spatial consistency and integrity across all datasets.

3.4.2. Tree Segmentation and Extraction

The initial phase of tree extraction involved segmenting the preprocessed point cloud data to isolate individual trees, including crowns and stems. This process is crucial for accurately modeling tree structures and was divided into several detailed sub-steps:

To accurately identify core stem regions, the point cloud was first segmented into horizontal height bands ranging from 5.6 m to 5.9 m above ground level. This specific height band was chosen as it typically corresponds to the lower sections of tree stems, providing a stable basis for clustering [70]. The number of clusters in K-means was set adaptively based on the local density of the height band. Within that height band, an enhanced K-means clustering algorithm was employed to group points belonging to individual tree stems [71]. The objective function for this enhanced K-means clustering was defined as:

$$\mathcal{J}=\sum _{i=1}^k\sum _{\mathbf{x}\in S_i}w\left(\mathbf{x}\right)·\parallel \mathbf{x}-{\mu }_i{\parallel }^2+\lambda \sum _{i=1}^k{\parallel {\mu }_i-{\mu }_0\parallel }^2$$

(2)

where $\mathcal{J}$ represents the total clustering objective function; k is the number of clusters; $S_i$ denotes the ith cluster; $\mathbf{x}$ is a point within cluster $S_i$; ${\mu }_i$ is the centroid of cluster $S_i$; ${\mu }_0$ is the global mean of all centroids; $w\left(\mathbf{x}\right)$ is a weight assigned to point $\mathbf{x}$, reflecting its significance based on spatial density or other relevant criteria; $\lambda$ is a regularization parameter that controls the influence of the centroid dispersion term. The objective function $\mathcal{J}$ comprises two primary components: (1) Weighted Intra-cluster Variance: ${\sum }_{i=1}^k{\sum }_{\mathbf{x}\in S_i}w\left(\mathbf{x}\right)·{\parallel \mathbf{x}-{\mu }_i\parallel }^2$. This term aims to minimize the weighted sum of squared distances between each point $\mathbf{x}$ and its respective cluster centroid ${\mu }_i$. The weights $w\left(\mathbf{x}\right)$ can be adjusted based on factors such as point density or confidence levels derived from prior segmentation steps, enhancing the clustering performance in areas with varying point distributions. (2) Centroid Dispersion Regularization: $\lambda {\sum }_{i=1}^k{\parallel {\mu }_i-{\mu }_0\parallel }^2$. This regularization term penalizes the dispersion of cluster centroids from the global mean ${\mu }_0$, promoting a more balanced distribution of clusters across the height band. The parameter $\lambda$ controls the strength of this penalty, allowing for flexibility in balancing intra-cluster compactness and inter-cluster separation. The inclusion of weighting and regularization in the K-means objective function facilitates a more nuanced clustering, particularly in environments with heterogeneous point distributions and varying tree stem characteristics [72]. By assigning appropriate weights and controlling centroid dispersion, the algorithm effectively groups spatially proximate points within the specified height band, thereby isolating individual tree stems from the surrounding vegetation with higher precision [71].

Post clustering, segments containing fewer than 20 points were discarded. This thresholding was implemented to eliminate noise and spurious detection, which are often the result of sensor inaccuracies, transient objects, or small underbrush that do not represent significant tree stems. Retaining only segments with 20 or more points ensured that the subsequent analysis focused on meaningful tree structures, thereby enhancing the reliability of the extraction process.

For each valid tree segment, the Bird’s Eye View (BEV) height compression was employed to transform the 3D point cloud into a 2D plane parallel to the ground [73]. By discarding the vertical dimension and optionally applying rotation, scaling, and translation, this projection reduces the complexity of geometric analysis, enabling more efficient cross-sectional modeling of tree stems [74,75]. Specifically, let $(x,y,z)$ represent the original 3D coordinates of a point. The BEV projection maps this point to a 2D coordinate system $(X_{BEV},Y_{BEV})$ through a combined linear transformation and translation:

$$\left[\begin{array}{c}{X}_{BEV}\\ Y_{BEV}\end{array}\right]=\underset{\mathrm{Rotation}\mathrm{and}\mathrm{scaling}}{\underbrace{\left[\begin{array}{ccc}scos\left(\varphi \right)& -ssin\left(\varphi \right)& 0\\ ssin\left(\varphi \right)& scos\left(\varphi \right)& 0\end{array}\right]}}\left[\begin{array}{c}x\\ y\\ z\end{array}\right]+\underset{\mathrm{Translation}}{\underbrace{\left[\begin{array}{c}\Delta X\\ \Delta Y\end{array}\right]}}$$

(3)

Here, s is a scaling factor to adjust the spatial resolution, $\varphi$ is a rotation angle to align the coordinate system with the forest plot orientation, and $(\Delta X,\Delta Y)$ is a translation vector to position the projected data in a convenient 2D reference frame. By selecting $s=1$, $\varphi =0$, and $(\Delta X,\Delta Y)=(0,0)$, the transformation reduces to a simple orthographic projection that discards the vertical dimension, retaining only the x and y coordinates of the original 3D points. In practice, these parameters were chosen to best represent the tree stem structure, facilitating subsequent 2D geometric fitting for the stem cross-sectional analysis.

After projecting the 3D points into the BEV plane, the next step was to determine the optimal circle that best represented the stem’s cross-sectional geometry. In practice, a robust weighted least squares approach was employed to minimize discrepancies between the observed points and the fitted circle, accommodating irregularities in point density and outliers. Let $(x_i,y_i)$ be the coordinates of the BEV-projected points and $(a,b,r)$ be the circle parameters, where $(a,b)$ denotes the center and r the radius of the circle. Let $w_i$ be a weight associated with each point, reflecting its reliability (based on local point density or proximity to other high-confidence points), and let $\lambda$ be a regularization parameter that stabilizes the solution by penalizing large deviations from an initial radius estimate $r_0$ [76,77]. The objective function is given by:

$$\underset{a,b,r}{min}\sum _{i=1}^n{w}_i{(\sqrt{{(x_i-a)}^2+{(y_i-b)}^2}-r)}^2+\lambda {(r-r_0)}^2$$

(4)

Here, $\mathcal{J}$, defined as ${min}_{a,b,r}{\sum }_{i=1}^n{w}_i{\left(\sqrt{{(x_i-a)}^2+{(y_i-b)}^2}-r\right)}^2$ ensures that the circle closely fits the BEV points while accommodating varying confidence levels through $w_i$. $\lambda {(r-r_0)}^2$, controlled by $\lambda$, encourages the radius to remain near a chosen reference $r_0$, thus preventing overfitting and improving numerical stability. By iteratively solving this optimization problem, a precise and stable circle fit is achieved, providing an accurate estimate of the stem’s diameter for subsequent stem modeling and analysis [77].

To account for stem variability such as bending, inclination, and canopy expansion, the fitted radius r was increased by 50%, yielding an adjusted radius $r^{\prime }=1.5r$. This enlargement ensured full coverage of the stem region, mitigating the risk of excluding points due to geometric distortions or natural stem tapering. The rationale for this adjustment was based on the understanding that tree stems are not perfectly cylindrical and often exhibit variations in thickness and curvature.

Using the adjusted radius $r^{\prime }$, a cylindrical volume was defined around each tree segment’s centroid. All points within this cylinder were then classified as part of the respective tree stem. This segmentation effectively isolated the stem from surrounding vegetation and structural elements, enabling accurate modeling of the tree’s 3D structure.

3.4.3. Stem Point Extraction

The extraction of precise stem points was critical for accurate feature measurement and subsequent analysis. This process involved iterative cylinder fitting to model the tree stem’s 3D structure, ensuring robustness against occlusions and stem variability.

For each segmented tree, cylinder fitting was performed starting from the lower 5% height of the stem and progressing upwards. The Maximum Likelihood Estimation Sample Consensus (MLESAC) algorithm was utilized to robustly fit cylindrical models to the point cloud data. MLESAC enhances the reliability of cylinder fitting by maximizing the likelihood of inliers while minimizing the influence of outliers, effectively distinguishing the stem from branches and other structural elements, with thresholds for inlier distances set at 0.1 m to balance sensitivity and specificity [78]. The cylinder fitting objective is formulated as:

$$\mathcal{L}\left(\theta \right)=\sum _{i=1}^{n}log\left({\displaystyle \frac{1}{\sqrt{2\pi }\sigma }}exp\left(-{\displaystyle \frac{{(d_i-d_0)}^2}{2{\sigma }^2}}\right)\right)$$

(5)

where $\theta$ represents the cylinder parameters (axis direction, radius, and position), $d_i$ is the distance of point i from the cylinder surface, $d_0$ is the nominal distance threshold, and $\sigma$ is the standard deviation.

The iterative nature of this process allowed for the sequential refinement of the stem model, starting from the base and moving upwards, ensuring that each segment of the stem was accurately captured [79].

Each fitted cylinder was validated by comparing the radius of the current segment with that of the preceding one. A decrease in radius typically indicates normal stem tapering, whereas an increase suggests the presence of branches erroneously identified as stem segments. In cases where an increase in radius was detected, the algorithm discarded the upper segment to prevent misclassification. This validation step was crucial for maintaining the integrity of the stem model, ensuring that only true stem points were retained.

This iterative process continued until the entire stem height was accurately modeled, resulting in a precise 3D representation of each tree’s stem.

3.4.4. Feature Extraction and Results Validation

The final phase of the methodology focused on extracting quantitative features from the accurately modeled tree stems and validating these features against ground-truth measurements to ensure accuracy and reliability.

From the 3D stem models, the diameter at breast height (DBH) was calculated at 1.3 m above ground level. This measurement was derived from the fitted cylindrical geometry of the stem, providing a standardized metric for comparing tree sizes. Additionally, tree height was determined by identifying the highest point within each stem segment, offering a complete understanding of each tree’s vertical structure. These parameters are fundamental for estimating forestry commercial timber volume storage and value assessment, which are critical components in assessing forest health and productivity.

Each stem segment’s center point and radius were meticulously recorded, enabling a detailed analysis of stem tapering and curvature. By examining the variation in radii across different segments of the stem, it is possible to assess the structural integrity of the tree. Consistent tapering indicates healthy growth patterns, while irregular variations may signal structural anomalies such as excessive bending or the presence of defects. This granular level of detail allows for the identification of subtle changes in tree morphology, which are essential for monitoring tree health and growth dynamics.

To further evaluate the structural integrity of the tree stems, deviation values were calculated for each segment, quantifying the stem’s curvature in both the x and y directions. Specifically, the deviations were defined as the angles between consecutive stem segments projected onto the respective axes. For each segment i, the x deviation angle (${\theta }_{x,i}$) and y deviation angle (${\theta }_{y,i}$) were calculated using the following formulas:

$${\theta }_{x,i}=arctan\left({\displaystyle \frac{\Delta x_i}{\Delta s_i}}\right),{\theta }_{y,i}=arctan\left({\displaystyle \frac{\Delta y_i}{\Delta s_i}}\right)$$

(6)

where $\Delta x_i=x_i-x_{i-1}$ is the change in the x-coordinate between segment i and segment $i-1$, $\Delta y_i=y_i-y_{i-1}$ is the change in the y-coordinate between segment i and segment $i-1$, and $\Delta s_i$ is the horizontal distance between segments $i-1$ and i, calculated as $\Delta s_i=\sqrt{{(\Delta x_i)}^2+{(\Delta y_i)}^2}$.

By analyzing these deviation angles, the methodology assesses the degree of stem curvature, providing insights into the tree’s stability and resilience. High deviation values in either the x or y direction may indicate significant bending or irregular growth patterns, which could be symptomatic of environmental stressors or genetic factors affecting the tree’s development.

4. Results

This section outlines the experimental procedures and results validating FoSEM for tree stem extraction and feature analysis using LiDAR-derived 3D point cloud data. It is organized into three main subsections: (1) Experimental Dataset and Forest Plot Parameters, (2) Results and Validation, and (3) Tree Stem Flexure Analysis. These subsections detail the implementation process, dataset characteristics, and FoSEM’s effectiveness through a comparative analysis with ground-truth measurements.

4.1. Experimental Dataset and Forest Plot Parameters

The evaluation data from the 40 circular plots distributed across two regions in WA were utilized. For a detailed analysis, three representative plots were selected and categorized according to their structural complexity: low, medium, and high, using factors such as tree density, spatial distribution, branch presence, and overall canopy structure. These plots, summarized in

Table 2. Comparison of point cloud characteristics and tree measurement trends across three test plots. The table summarizes tree characteristics, including diameter at breast height (DBH) and tree height, along with point cloud metrics such as the number of points, density, and scan duration, ensuring full coverage of each field plot.

#Plot	Trees	Age (Year)	DBH (mm)		Height (m)		Points (mil)	Density (m−2)	Duration (min)
			Range	Mean	Range	Mean
25	15	34	310–529	412.41	28.8–42.7	36.11	29.2	43,171	7
31	24	18	155–305	241.19	10.3–26.5	22.95	50.7	75,142	8
42	50	8	60–200	135.28	9.6–15.8	13.83	39.5	58,443	10

, were used to evaluate the effectiveness of the proposed model. Additionally, three felled trees aged 8, 18, and 34 years were chosen from the categorized plots to derive performance metrics for the model. Each plot is detailed below.

Plot 25 (low complexity): Located at site 1, Plot 25 contained 15 trees aged approximately 34 years, with DBH ranging from 310 mm to 529 mm (average 412.41 mm) and heights from 28.8 m to 42.7 m (average 36.11 m), as shown in Table 2. The trees were uniformly distributed with minimal overlap or occlusion, exhibiting straight stems with consistent tapering and minimal branching. The absence of significant underbrush further reduced noise, allowing for efficient and accurate tree stem modeling.

Plot 31 (medium complexity): Located at site 2, Plot 31 comprised 24 trees aged around 18 years, with DBH ranging from 155 mm to 305 mm (average 241.19 mm) and heights from 10.3 m to 26.5 m (average 22.95 m), as shown in Table 2. This plot featured moderate tree density, leading to increased spatial overlap and occlusions in the LiDAR data. Moderate branching, pronounced stem curvature, and underbrush contributed to segmentation challenges, requiring advanced algorithms for accurate modeling. Despite these complexities, the relatively organized spatial distribution supported effective segmentation with appropriate preprocessing.

Plot 42 (high complexity): Located at site 2, Plot 42 contained 50 trees aged around 8 years, with DBH ranging from 60 mm to 200 mm (average 135.28 mm) and heights between 9.6 m and 15.8 m (average 13.83 m), as shown in Table 2. This plot was highly complex due to its dense tree distribution, significant spatial overlap, and frequent occlusions in the LiDAR data. Extensive branching, irregular stem curvatures, and dense underbrush add to the noise, complicating stem isolation. The thin stems of younger trees were further obscured by foliage, posing challenges for segmentation. These conditions made Plot 42 an ideal test case for evaluating the robustness of FoSEM in high-complexity environments.

4.2. Results and Validation

The effectiveness of the proposed framework was evaluated by comparing the LiDAR-derived measurements with ground-truth data obtained through manual sampling. The comparison focused on key tree parameters, including diameter at breast height (DBH), tree height, and stem segment characteristics. The results are categorized based on the complexity of the plots: low, medium, and high complexity. Detailed comparisons for sample trees in each plot are presented in

Table 3. Tree measurements and section details for Plot 25, Tree 6. This table presents the true and predicted values for DBH, tree height, stem length, and diameter at different sections of the stem. The “Diameter (mm)” column shows the true value followed by the predicted value in an underlined format for easy comparison.

#Plot	#Tree	DBH (mm)	Tree Height (m)	Stem Length (m)	Section Length (m)	Diameter (mm)
					0.12	387/391.87
					0.9	326/335.42
					1.3	307/307.84
					2.0	298/287.17
					3.1	271/276.68
					5.1	246/249.23
					7.0	232/238.01
25	6	317/307.84	30.4/31.14	30.37/29.53	8.9	222/228.16
					11.0	220/226.79
					13.0	187/186.91
					14.9	179/178.73
					17.9	165/171.37
					21.1	133/139.16
					24.0	98/94.09
					27.0	79/88.67
					30.0	15/19.28

,

Table 4. Tree measurements and section details for Plot 31, Tree 4. This table presents the true and predicted values for DBH, tree height, stem length, and diameter at different sections of the stem. The “Diameter (mm)” column shows the true value followed by the predicted value in an underlined format for easy comparison.

#Plot	#Tree	DBH (mm)	Tree Height (m)	Stem Length (m)	Section Length (m)	Diameter (mm)
					0.12	348/350.99
					0.65	307/323.71
					1.30	300/315.27
					2.00	281/287.42
					3.00	266/273.95
					5.00	234/244.37
					7.00	219/226.23
31	4	303/315.27	26.1/27.07	27.4/26.9	9.00	207/211.86
					11.00	188/194.79
					13.00	171/179.28
					15.20	151/159.92
					18.00	120/128.47
					21.00	88/97.31
					24.00	51/59.46
					27.00	10/17.03

, and

Table 5. Tree measurements and section details for Plot 42, Tree 35. This table presents the true and predicted values for DBH, tree height, stem length, and diameter at different sections of the stem. The “Diameter (mm)” column shows the true value followed by the predicted value in an underlined format for easy comparison.

#Plot	#Tree	DBH (mm)	Tree Height (m)	Stem Length (m)	Section Length (m)	Diameter (mm)
					0.1	180/186.02
					0.8	147/162.28
					1.3	140/159.87
					2.0	132/133.74
42	35	146/159.87	14.3/15.53	14.7/14.26	3.0	126/126.35
					5.0	114/112.84
					7.2	95/95.42
					9.0	76/73.85
					11.0	54/50.28
					13.0	30/26.03

for Plot 25, Plot 31, and Plot 42, respectively.

Table 3, Table 4, and Table 5 illustrate the accuracy and reliability of the FoSEM framework under varying levels of forest complexity. In Plot 25 (low complexity), the predicted DBH closely aligned with the true value, with a predicted DBH of 307.84 mm compared to 317 mm, while tree height and stem length predictions were similarly accurate at 31.14 m and 29.53 m, respectively, compared to true values of 30.4 m and 30.37 m. Diameter predictions across stem sections further validated the framework’s robustness in simpler environments. In Plot 31 (medium complexity), the predicted DBH was 315.27 mm, closely matching the true value of 303 mm, while tree height and stem length predictions were 27.07 m and 26.9 m, respectively, compared to true values of 26.1 m and 27.4 m. These results highlight the framework’s adaptability and precision in moderately complex environments. For Plot 42 (High Complexity), the framework maintained strong agreement between true and predicted values despite the plot’s intricate structure, with a predicted DBH of 159.87 mm compared to 146 mm, a tree height of 15.53 m against 14.3 m, and a stem length of 14.26 m versus 14.7 m. These findings demonstrate FoSEM’s robustness and accuracy in extracting key tree features across varying forest complexities, even under challenging conditions.

As shown in

Table 6. Comparison of FoSEM with state-of-the-art methods for DBH and height accuracy. The table categorizes results into two main sections: DBH (RMSE, rRMSE) and height (RMSE, rRMSE), providing a clear and concise performance comparison.

Method	Tree Type	DBH (cm)		Height (m)
		RMSE	rRMSE (%)	RMSE	rRMSE (%)
LiDar Based Individual Tree Detection (ITD) Approach [80]	Castor aralia	7.39	-	-	-
LiDar Based Individual Tree Detection (ITD) Approach [80]	Japanese Oak	11.87	-	-	-
LiDar Based Individual Tree Detection (ITD) Approach [80]	Monarch Birch	7.05	-	-	-
Improved Area-Based Approach [81]	Chinese fir	1.71	12.91%	-	-
Improved Area-Based Approach [81]	Eucalypt	2.51	16.44%	-	-
Improved Area-Based Approach [81]	Pooled	3.34	22.76%	-	-
Multiscale Cylindrical Detection and Multiscale Ring Fitting [83]	Broad-leaved Artificial Poplar	1.6	-	-	-
Octree Segmentation, Connected Component Labeling, and Random Hough Transform [82]	Pinus yunnanensis	1.17	-	0.54	-
Octree Segmentation, Connected Component Labeling, and Random Hough Transform [82]	Pinus densata	1.28	-	0.57	-
Octree Segmentation, Connected Component Labeling, and Random Hough Transform [82]	Picea and Abies fabri	-	-	1.28	-
Octree Segmentation, Connected Component Labeling, and Random Hough Transform [82]	Quercus semecarpifolia	1.22	-	1.23	-
Forest Stem Extraction and Modeling (FoSEM)	Pinus radiata	1.19	4.67%	1.00	4.24%

, FoSEM demonstrated higher accuracy and reliability in predicting DBH and height for Pinus radiata, outperforming state-of-the-art methods. It achieved the lowest RMSE and rRMSE values, with an RMSE of 1.19 mm and an rRMSE of 4.67% for DBH, and 1.00 m and 4.24% for height, highlighting its superior precision. The rRMSE metric ensures a fair comparison across forest and tree types by normalizing RMSE relative to observed values, enabling consistent evaluation under varying conditions. In comparison, alternative methods like the LiDAR-Based Individual Tree Detection (ITD) approach [80] and the Improved Area-Based Approach from Zhang et al. [81] achieved competitive but less precise results, with RMSE values for DBH often exceeding 7.00 mm and rRMSE reaching 16.44%. Similarly, the Octree Segmentation and Random Hough Transform method from Liu et al. [82] performed well for other species, such as Pinus yunnanensis, with RMSE values of 1.17 mm for DBH and 0.54 m for height, but lacked detailed rRMSE metrics for direct comparison. These results, combined with FoSEM’s demonstrated accuracy in varying complexity levels across Plots 25, 31, and 42, underscore its potential as a robust and reliable tool for tree stem modeling and analysis in diverse forest environments.

4.3. Tree Stem Flexure Analysis

To further assess each tree’s structural condition, we evaluated stem flexure by examining angular deviations between consecutive stem segments. These deviations capture how each segment shifts horizontally relative to the previous one, providing a practical measure of bending or leaning. As summarized in

Table 7. Tree curvature details for Plot 25, Tree 6 (a), Plot 31, Tree 4 (b), and Plot 42, Tree 35 (c). This table presents the section lengths and predicted deviation angles along the x-axis and y-axis for each section of the stem.

(a) Plot 25, Tree 6
Section Length (m)	Deviation Angle (∘)
	x-Axis	y-Axis
0.12	0.00	0.00
0.9	−1.29	2.33
1.3	−0.86	2.03
2.0	−0.49	1.82
3.1	−0.09	1.66
5.1	−2.69	1.35
7.0	−3.77	−1.01
8.9	−1.35	0.51
11.0	−0.97	0.44
13.0	−0.08	−0.22
14.9	−0.55	0.43
17.9	0.55	0.12
21.1	0.44	0.14
24.0	−0.35	0.85
27.0	1.31	−0.38
(b) Plot 31, Tree 4
Section Length (m)	Deviation Angle (∘)
	x-Axis	y-Axis
0.12	0.00	0.00
0.65	−1.29	2.33
1.30	−0.86	2.03
2.00	−0.49	1.82
3.00	−0.09	1.66
5.00	−2.69	1.35
7.00	−3.77	−1.01
9.00	−1.35	0.51
11.00	−0.97	0.44
13.00	−0.08	−0.22
15.20	−0.55	0.43
18.00	0.55	0.12
21.00	0.44	0.14
24.00	−0.35	0.85
27.00	1.31	−0.38
(c) Plot 42, Tree 35
Section Length (m)	Deviation Angle (∘)
	x-Axis	y-Axis
0.1	0.00	0.00
0.8	−2.67	−3.41
1.3	−1.05	−0.29
2.0	−1.05	−0.29
3.0	−0.68	−0.36
5.0	−0.66	−0.12
7.2	0.08	−0.31
9.0	−0.16	0.35
11.0	0.09	−0.03
13.0	0.03	0.12

, the magnitudes of these angles varied across plots of different complexity. (As shown in Figure 6), in the low-complexity plot (Plot 25), deviation angles remained small and relatively consistent, indicating that trees generally maintained upright forms with minimal curvature. By contrast, the medium-complexity plot (Plot 31) exhibited moderate angular shifts at specific heights, reflecting localized bending likely influenced by neighboring trees and mid-level stand density. The highest angles occurred in the most complex plot (Plot 42), where close spacing and competition for light contributed to more pronounced leaning or curved growth patterns. These observations highlight the advantage of capturing fine-scale geometric details beyond diameter and height measurements. In practice, larger angular deviations may reveal areas of stem instability or abnormal taper, assisting in identifying potential structural weaknesses or environmental stress effects. Such insights can aid forest managers in prioritizing silvicultural interventions, ensuring safer stands and enhancing the reliability of downstream timber assessments.

5. Discussion

While the proposed FoSEM framework demonstrated high accuracy and robustness in modeling tree stems from dense 3D point cloud data, several limitations warrant further discussion. A critical constraint lies in its reliance on high-density LiDAR data. The framework’s preprocessing and segmentation steps, such as adaptive filtering and clustering, are optimized for datasets with detailed spatial resolution. In low-density point clouds, the reduced point coverage can lead to inaccuracies in stem segmentation and parameter extraction. Sparse data amplify challenges such as increased noise, missing sections of stems, and diminished reliability of geometric fitting techniques, which are integral to FoSEM’s methodology.

Another limitation arises from its current application to radiata pine plantations, which are characterized by relatively uniform structural attributes and manageable levels of occlusion compared to mixed-species forests or tropical rainforests. Forests with vastly different structural characteristics, such as heterogeneous canopies, varying tree morphologies, and dense understories, could introduce additional complexities. For instance, highly irregular tree forms and overlapping vegetation may exacerbate occlusion and segmentation errors, requiring significant adaptation of the clustering and fitting algorithms.

Scaling the FoSEM framework to larger forested areas also presents challenges. Computational efficiency, particularly for iterative cylinder fitting and point cloud filtering, becomes a critical bottleneck when applied to extensive datasets. Moreover, variations in terrain and vegetation characteristics across regions may necessitate site-specific tuning of parameters, such as those governing ground segmentation or height normalization, further complicating large-scale implementation.

To address these limitations, future work will focus on extending the applicability of FoSEM to diverse forest types. Systematic experiments will be conducted in forests with structural characteristics significantly different from radiata pine plantations, including mixed-species temperate forests, tropical rainforests, and boreal ecosystems. These efforts aim to refine the framework’s adaptability and robustness in environments with complex canopy structures and higher variability in tree morphology. Additionally, scaling the framework to regional and national levels will require optimizing computational workflows, incorporating distributed processing techniques, and leveraging advancements in edge computing for real-time data processing. Expanding the geographic scope of experiments to include forests in other countries and climates will further validate the framework’s versatility and ensure its utility in global forestry applications.

6. Conclusions

This study introduced the Forest Stem Extraction and Modeling (FoSEM) framework, a LiDAR-based methodology designed to accurately model individual tree stems from dense and complex forest 3D point clouds. By integrating ground segmentation, height normalization, enhanced K-means clustering, circle fitting for cross-sectional geometry, and iterative MLESAC-based cylinder fitting, FoSEM effectively isolated and quantified stem attributes across varying forest conditions. Experimental evaluations conducted on 40 radiata pine plots of low, medium, and high complexity demonstrated FoSEM’s superior accuracy, achieving a Root-Mean-Square Error (RMSE) of 1.19 cm for diameter at breast height (DBH) and 1.00 m for tree height, outperforming existing methods as highlighted in Table 6. Additionally, the framework maintained consistent performance across different plot complexities, with low and medium complexity plots showing minimal deviations and high complexity plots still achieving acceptable accuracy despite increased challenges. Complementing the methodological advancements, this research also developed a large-scale dataset encompassing 40 experimental plots with diverse tree ages and densities, providing precise ground-truth measurements through destructive sampling. This dataset serves as a valuable benchmark for future studies and facilitates the reproducibility and scalability of LiDAR-based forest inventory techniques. While FoSEM has proven to be robust and efficient, future work will focus on refining segmentation algorithms, incorporating advanced machine learning techniques to further enhance accuracy, and expanding the dataset to include a broader range of forest types and conditions. Overall, FoSEM represents a significant advancement in LiDAR-driven tree stem modeling, offering a scalable and precise tool for sustainable forest management and conservation efforts.