Leveraging Open-Source Tools to Analyse Ground-Based Forest LiDAR Data in South Australian Forests

Abstract

This paper investigates the application of open-source software and methods for forest LiDAR analysis, with a focus on enhancing forest inventory metrics in the radiata pine forests of South Australia’s Green Triangle region. A semi-systematic survey identified 22 relevant open-source tools, evaluated for their capabilities in inventory metric extraction and practicality for implementation in industrial workflows. Ground truth data from radiata pine forests across multiple development stages provided the basis for validating the tools’ precision, accuracy, and practicality. Results showed that stratified tool selection, optimized for each forest development stage, achieved high accuracy for inventory, achieving stem detection rates up to 99.1% and errors as low as 0.94 m for height and 1.18 cm for diameter at breast height (DBH) in specific cases. Additionally, we provide scripts to support future research, discuss the limitations of our approach, and propose solutions to address these gaps in future implementations. Our findings highlight the utility of open-source tools to optimize forest inventory workflows through stratified, modular approaches.

1. Introduction

A forest inventory is a description of the quantity and quality of trees and other organisms that live in the forest and the characteristics of the land on which the forest grows. It serves as the source of information for decision-making in forest management [1]. Forest inventory also refers to the techniques for collecting, evaluating and presenting specified information on forest areas [2]. Since it is impractical to measure every tree in a forest, forest inventories rely on sampling techniques to estimate population-level characteristics from a subset of measured plots [2]. Most forest inventory systems comprise of a field survey component involving on-ground measurement and recording of selected forest and tree metrics, within spatially defined sampling units (field plots). As field measurements are expensive, only those metrics that are of interest at a point in time are included in the captured data.

Ground-based Light Detection and Ranging (LiDAR) technologies such as Terrestrial Laser Scanning (TLS), Mobile Laser Scanning (MLS) and Personal Laser Scanning (PLS) have generated significant interest among forest practitioners because of their ability to generate data that capture the 3D structure of the forest and individual trees with a level of accuracy that is adequate for forest inventory applications. A modern TLS device can quickly capture high-resolution point clouds of field plots. These LiDAR point clouds represent a high-fidelity depiction of an area of forest frozen at the point in time of capture. As forest information needs evolve or analytical tools become more powerful, these point clouds can be re-processed to extract improved or new metrics. The more recent Simultaneous Location and Mapping (SLAM) based PLS offer additional benefits, such as reduced data capture times [3] and relatively simple field procedures.

The extraction of forest metrics from high-density 3D point clouds is computationally intensive and requires efficient software. The availability of software tools is a prerequisite for the operational adoption of LiDAR-based forest assessment by forest managers. Over the years, researchers have developed a multitude of alternative methods with distinct analytical steps organized in logical processing sequences. Liang et al. [4] compared eighteen published methods for their performance in extracting forest metrics from TLS datasets captured in forests differing in terms of species and structural complexity. They assessed forests of various ‘complexities’, encompassing different species, ages, and management activities across homogeneous and heterogeneous forests. This benchmark study demonstrated that different methods have different strengths and weaknesses regarding the detection of individual trees and the subsequent measurement of tree attributes (e.g., diameter, height, and biomass), depending on forest stand conditions and scanning technique.

Several developers have made their methods (i.e., algorithms and processing pipelines) available to the general public as open-source software. These software tools are typically implemented around a specific LiDAR analysis method developed by the author of the open-sourced software. Given Liang et al.’s findings, it is expected that these software tools will inherit the strengths and weaknesses of the embedded methods and algorithms. They will perform better in some forest types than others. It also follows that there is merit in combining multiple tools in a toolchain fashion to leverage the strengths of the various tools, as evidenced by several other works: Hartley et al. [5] combined modules of CloudCompare [6], LidR [7], SimpleForest [8] and TreeLS [9] to perform phenotypical assessment of radiata pine forest in New Zealand. Rocha et al. [10] utilized different R packages to produce different inventory metrics. TLS2Trees [11] is a TLS tree extraction implementation that leverages the semantic segmentation from another tool, specifically the Forest Structural Complexity Tool [12]. Donager et al. [13] developed Spanner, an R package that implements a processing pipeline in tandem with the additional R package lidR [7], some TreeLS functions, as well as custom-made components. These studies demonstrate that the availability of effective software tools can enable the development of new processing pipelines and applications.

Open-source tools are valuable for research and development, allowing researchers to access, modify, and combine code and algorithms for new applications. Although some catalogs of open-source forestry toolkits exist (e.g., Atkins et al. [14] on forestry-specific R packages), gaps remain in evaluating their suitability for industrial use, PLS data, cross-platform implementation, and specific forest domains. While Liang et al. [4] evaluate the accuracy of the methods they assess, forest growers may find it challenging to determine the applicability of these methods to their forests without those specific forests being directly assessed in the study. Some industrial forest managers may conduct internal assessments of such tools for their specific use cases; however, there is often little incentive to publish their findings publicly.

One such forest domain with a growing interest in the industrial applications of such tool kits is the Green Triangle region of South Australia and South Western Victoria [15]. This area is characterized by extensive forest coverage and significant modern and historical forestry activity across a number of local forest growers. The Green Triangle industrial forests, especially those around Mt Gambier, are predominantly monoculture radiata pine. Therefore, local forest growers have specific regional requirements and are primarily interested in tools that can be applied with high specificity to their forests. Tools that provide higher precision through tailoring to different development stages are advantageous, as production forests vary in structure, density and complexity as operations (such as tree thinning) are conducted. Radiata pine in the Green Triangle region, therefore, serves as an effective test case for assessing new methods, due to both the abundance of local forest growers and the global prevalence of industrial monoculture pine forestry, for which findings have a degree of generalizability.

Existing benchmarking literature is extensive, but there remains a notable gap regarding open-source, cross-platform LiDAR analysis software for forestry. This study aims to address that gap by evaluating the performance and applicability of open-source forest point cloud tools for forest inventory, providing both a critical review of tools and new methodologies for performance assessment across key metrics. Our analysis is focused on a single domain, specifically South Australian radiata pine. To validate our findings, we use ground-truth data captured in situ to evaluate the tools across multiple forest development stages.

In this paper, we review open-source forest point cloud tools (up until 2024) to identify their strengths, limitations, and gaps relative to expert-defined requirements. These findings aim to inform future industrial implementation efforts. The methodology and scripts for assessment of tool performance across three key forest inventory metrics are as follows: (i) stem segmentation, (ii) diameter at breast height (DBH) and (iii) height measurements, which can be generalised to similar forests in other regions. The results for the validation of tool performance on PLS datasets were captured in a succession of development stages in a South Australian radiata pine forest (Pinus radiata D. Don), using conventional field measurements as benchmark data. We suggested a stratified implementation, discussion, limitations, and future work.

2. Survey of Open Source Tools

Our methodology for conducting a survey of open-source forestry tools is split into three parts: first, defining the functional requirements with industry experts; second, systematically identifying open-source forestry-applicable tools that meet the data requirements and lastly, verifying the characteristics of the identified tools.

2.1. Functional Requirements

To ensure our data assessment criteria were relevant to industry, we consulted six forest management experts, two of whom are co-authors of this paper. These discussions identified key forest inventory metrics that can be extracted from LiDAR scans and secondary attributes relevant to practical implementation.

2.1.1. Primary Tool Functionality

We categorized essential software functionality into two groups: data processing features and forest metric extraction. Data processing capabilities include file handling, cropping, subsampling, normalization, and automatic ground removal. Attention was paid to tree segmentation and leaf/wood segmentation, being the processes of labeling each point in a point cloud as belonging to a specific tree or as leaf/wood matter, respectively. The process of tree segmentation is critical, as it provides a stem count, and segmenting each tree is typically necessary before conducting biometric analysis. For forest metric extraction, we focused on the following:

• Stand-level metrics: Stem count, canopy, cover, and biomass estimation.
• Individual tree metrics: DBH, height, volume, stem angle, stem taper, and creation of quantitative structure models (QSMs).
• Branch and crown attributes: Branch volume, length, angle, count, and crown shape, volume and diameter.
• Feature detection: Identification of specific tree features, such as whorls, double leaders, and knots.

Among these, stem segmentation, DBH, and height measurements were prioritized due to their significance in traditional forest inventory practices.

2.1.2. Secondary Tool Functionality

We also noted a set of secondary tool attributes to assess each tool’s viability for implementation within an industry setting. These included details on the implementation of the tool software, such as the operating system, software platform, publisher, and license. We also assessed the tools’ capacity for automation, as tools that could not be automated for batched processing would create bottlenecks in a large-scale industrial setting. From our discussions, we identified that although fully operationalized industrial implementations would typically be implemented on server farms, preliminary setup, experimentation, calibration, and reviewing were more likely to happen on local machines, such as personal laptops. This structured evaluation ensured that the selected tools were both functionally robust and practically deployable in industrial forestry applications.

2.2. Semi-Systematic Review

We conducted a semi-systematic review, a method that combines systematic search principles with flexibility in refining search strategies [16], to identify open-source tools for processing forest LiDAR data. Specifically, we were interested in open-source tools that process raw point cloud data from a LiDAR scanner at the scale of either a single tree or a multi-tree plot. We focused on tools designed for terrestrial LiDAR to leverage the higher resolution afforded by such scanners.

Our search strategy involved an exhaustive literature search with Scopus as a baseline, employing specific keywords related to forest LiDAR and open-source tools. This was followed by iterative GitHub and Google Scholar searches to expand our findings. We limited our search to only tools that had an open-source version available prior to January 2024.

2.2.1. Literature Search

We started with the following search string on Scopus:

“Terrestrial laser scanning” OR “LiDAR” AND “Forest Inventory” OR “Forest Measurement” OR “Segmentation” AND “Open Source”

Our survey process is illustrated in Figure 1. This search yielded 96 relevant documents, which we refined to 33 based on titles and abstracts. Of these 33, we found 26 specific to forestry applications. These results were refined to 14 documents specific to terrestrial LiDAR applications, from which we extracted an initial list of 12 open-source tools.

Some papers were discarded because their subjects, though related to forestry, focused instead on non-terrestrial LiDAR or photogrammetry-adjacent processes. Common themes in such areas were species detection, biomass calculation, sample location determination, generation of synthetic training data for AI, creation of digital terrain models, and road detection.

Subsequent searches on GitHub and Google Scholar, using keywords such as LiDAR, Forestry, Inventory, Processing, identified ten additional tools that met our criteria, bringing the total to 22 tools. This included some general-purpose tools noted for their utility in preprocessing LiDAR data, and some ALS-centric tools that could be implemented into TLS workflows.

2.2.2. Attribute Testing

We analyzed the tools with an open coding approach. Each tool was assessed on our research hardware, designed to be similar to a standard industry research setup, specifically a Dell Precision 7760 Laptop (Win 10, 32 GB RAM, 11th Gen Intel Core i7-11850H, RTX 3080). We focused our testing on three primary attributes: stem segmentation, DBH and height. Wherever a tool claimed a feature, it was added to the review. However, we distinguished between features we could reproduce and those we could not. Some tools were challenging to operate on our hardware, while others were not tested in depth because they did not produce our primary testing attributes listed above. These tools are still included in the survey for comprehensiveness. We collated feature lists and short descriptions of each of the 22 identified tools.

After testing each tool, we contacted the developers (where possible) and asked them to review and confirm our feature lists and descriptions. Of the 22 tools, 14 developers responded to our request with feedback, which we took into consideration when editing our tool inventory. Developers were not obligated to respond to our request. The resulting tool inventory is presented in

Table 1. Overview of assessed tools for the systematic review. Confirmed features are marked with a checkmark. Tools with unvalidated outputs are highlighted in grey.

Overview	Tool Details						Data Features							Stand Metrics			Stem Metrics						Branch Metrics				Crown Metrics				Detection
Name	Platform	Test Version	Licence	Publisher	Automation	Use Case	File Operations	Point Cloud Subsampling	Scan Cropping	Ground Extraction	Tree Segmentation	Leaf/Wood Segmentation	Outputted Data Format	Stem Count	Canopy Cover	Biomass	DBH	Height	Volume	Angle	Taper	QSM Construction	Volume	Length	Angle	Count	Statistics	Shape	Volume	Diameter	Whorl	Double Leader	Knot
AdQSM	GLSL	1.7	“Open Tool”	Fan et al.	No	Individual Trees							.csv/.txt				✓	✓	✓			✓	✓	✓	✓	✓	✓	✓	✓	✓
alphashape3d	R	1.3.1	GNU GPL-2	Lafarge, Pateiro-Lopez	Yes	Crown Data							R Dataframe															✓	✓	✓
CloudCompare	C++	2.13.1	GNU GPL-2	Independent software	Yes	All Data	✓	✓	✓	✓	✓	✓	Most Filetypes				✓	✓			✓
FORTLS	R	1.4.0	GNU GPL -3	Molina-Valero et al.	Yes	All Scans			✓		✓		R Dataframe /.txt/.csv	✓			✓	✓	✓
Forest3D	Python	2021	MIT License	Windrim and Bryson	Yes	Uniform Stands				✓	✓	✓	.csv/.ply	✓			✓	✓
FSCT	Python	2022	GNU GPL -3	Sean Krisanski	Yes	All Scans	✓	✓	✓	✓	✓	✓	.las/.csv /.html/.png	✓	✓		✓	✓	✓	✓	✓						✓
ITSMe	R	1.0.0	MIT License	Terryn, L	Yes	Individual Trees							R Dataframe				✓	✓	✓				✓	✓	✓		✓			✓
LAStools	C++	2.0.2	Partially Open Source	rapidlasso	Yes	All Data	✓	✓	✓				Most Filetypes
lidR	R	4.1.1	GNU GPL -3	Roussel et al.	Yes	All Scans	✓	✓	✓	✓	✓		Most Filetypes	✓				✓									✓	✓
PointCloud Tree Modelling (PCTM)	Python	0.2	GNU GPL -3	Amsterdam Intellegence	Yes	Individual Trees							.csv				✓	✓		✓	✓						✓	✓	✓	✓
SimpleForest	Computree	5.0	GNU GPL	Jan Hackenberg	Yes	All Scans	✓	✓	✓	✓	✓	✓	CompuTree Data	✓								✓
Spanner	R	1.0.1	GNU GPL -3	Sánchez Meador et al.	Yes	All Scans					✓		R Dataframes + lidR LAS Object				✓
TreeLS	R	2.0.5	GNU GPL -3	De Conto et al.	Yes	Uniform Stands					✓		R Dataframes + lidR LAS Objects	✓			✓	✓		✓	✓
TreeQSM	MATLAB	2.4.1	GNU GPL -3	Raumonen et al.	Yes	Individual Trees							MATLAB Table				✓	✓	✓	✓	✓	✓	✓	✓	✓	✓	✓	✓	✓	✓
WhorlDetector	R / Python	2023	GNU GPL -3	SmartForest	Yes	Individual Trees							.csv																		✓
3D Forest	C++	0.52	GNU GPL-3	3D Forest Team	Yes	All Scans		✓		✓	✓		.txt/.ply/.pcd /shapefile	✓	✓		✓	✓	✓	✓	✓	✓	✓	✓			✓	✓	✓
GBSeparation	Python	1.0	“Open Source”	Tian Zhilin	Yes	Individual Trees						✓	.txt
LayerStacking	R	2017	MIT License	Elias Ayrey	Yes	All Scan Data					✓		R Dataframe
LeWoS	MATLAB	4.1	MIT License	Wang, Di	Yes	All Scan Data						✓	MATLAB Table
SSSC	MATLAB	2020	GNU GPL-3	Wang, Di	Yes	All Scans					✓	✓	MATLAB Table
TLS2Trees	Python	2023	GNU GPL-3	Phil Wilkes	Yes	All Scan Data				✓	✓	✓	.las/.csv/.ply	✓				✓									✓
TREESEG	Ubuntu	0.2.2	MIT License	Burt A et al.	Yes	All Scans				✓		✓	.pcd/.txt

. Specific descriptions of each tool are provided in Appendix A.

2.3. Tool Survey Analysis

With the exception of LAStools, which is only partially open-source, every other assessed tool had some form of open-source license. Most tools (14) use some version of the GNU GPL (General Public License) to share their software, typically GPLv3 or GPLv2. Another subset of tools (5) use the MIT License. Two tools (specifically AdQSM and GBSeparation) simply state that they are "open-source" without referring to a formal open-source format. A caveat exists in that though the identified MATLAB tools are open source, MATLAB itself is not, and a license is required for the platform’s use; these tools are still included as they meet our criteria, but operators should be aware of this limitation.

Most tools (8) were based on R libraries, as the platform is well-suited to point cloud analysis. Python was the second closest choice (6), with regular implementation of machine learning libraries for measurement and assessment. The other common platform was MATLAB (3), utilized with the shared benefits of ease of statistical analysis and data libraries. The remaining tools were implemented in either Ubuntu, C++ or GLSL.

Automation varied across the assessed tools. R, Python, and MATLAB-based tools were generally easy to automate, with some (e.g., FSCT) offering built-in batch processing functions. Others, such as LAStools and CloudCompare, provided command-line automation. The primary exception was AdQSM, which prioritized visualization over batch processing.

We found only one tool, WhorlDetector, capable of specific feature detection. Notably, no tool appeared to offer the detection of knots or double leaders. We identified no tools that provided functionality for calculating biomass, which is understandable given the complexities of predicting biomass without considering contextual factors. There may be no direct estimation of biomass, but biomass may be approximated based on determined diameter, height, and volume. We anticipate that the ease of QSM generation among these software suites will make implementing a double leader detector relatively straightforward. QSMs, provided they are correct, offer detailed insights into key structural features such as branching patterns and multi-leader tree configurations.

3. Functional Assessment of Tools

In this section, we narrow our focus to the tools identified in the survey that are most relevant to our target domain. As the introduction outlines, our primary interest is in tools applicable to radiata pine forestry in the Green Triangle region. To assess these tools, we collected scans and ground-truth measurements to create a validation dataset, which will be used to evaluate their performance.

We captured validation data across different stand development stages to examine the performance of tools as a function of age and thinning state. We distinguished three development stages: post-first thinning, age less than 15 (labelled T1), post-second thinning, age between 15 and 25 (T2), and post-third thinning, age greater than 25 (T3). Un-thinned stands (T0) were omitted from this study.

Tools selected for analysis are those capable of performing stem segmentation (stem counts), diameter at breast height (DBH) measurement, or height measurement, as these metrics are primary descriptors in most forest inventories (as discussed in Section 2.1). Our assessment is divided into two sections: an assessment of stem segmentation tools and an assessment of stem metric extraction tools.

3.1. Materials: Validation Dataset

To effectively validate the tools, we utilized validation data from local forests in the Mount Gambier area, part of the Green Triangle region. A team of professional foresters captured LiDAR scans and inventory metrics for a predetermined list of forest plots. We selected a total of 27 plots from a mix of development stages: Seven post first thinning (T1), ten post second thinning (T2) and ten post third thinning (T3).

Figure 2 shows the locations of the surveyed forest plots. Each plot had a LiDAR scan captured using either a Hovermap H100 or a Hovermap ST-X (https://emesent.com/hovermap-series/, accessed on 14 August 2024) mounted on a backpack. The datasets generated by the Hovermap contain the coordinates of the points in a local coordinate system, as well as the intensity and range (distance of a point to a scanner) data. This produced .laz files, which averaged 101 million points, 935 MB. All plots were conventionally measured. Trees within the circular or rectangular plots were numbered. The location, DBH, and height (for every other tree) were recorded. Heights were measured with a Haglof Vertex (https://haglofsweden.com/project/vertex-5/, accessed on 14 August 2024), and DBH was measured at 1.3 m from the ground, over-bark, using a measurement tape to the nearest 1 mm. The locations of the trees were recorded using the scan’s local coordinate systems, to enable subsequent matching of trees measured in the field and trees detected by the various tools. Overall the validation dataset comprised of 924 tree locations, 924 DBH references and 465 height references for the validation. A summary of these measurements is provided in

Table 2. Overview of average field measurements.

Development Stage	Stocking (stems/ha)	DBH (cm)	Height (m)
T1	702	25.4	24.6
T2	449	34.7	29.1
T3	244	44.7	33.4

.

LiDAR Data Preprocessing

We implemented a preprocessing step to optimize our results and enable the use of tools with more significant memory requirements. Our target hardware did not have the memory to process the full scans with some tools at source resolution, necessitating a subsampling step. For our hardware, a 15 mm subsampling step was determined to be sufficient to prevent memory allocation errors and was utilized for all tools.

Through testing, we found that most tools provided outputs with higher accuracy when they were run on noise-reduced point clouds. We tested several approaches to noise reduction and ultimately found that we obtained the best results using a basic intensity filter to retain only intensity values between 5 and 255. Removing the lowest intensity points in this manner seemed to remove noise most effectively, likely due to long range and low confidence points correlating with lower reflectivity, as well as foliage generating weaker reflections than stems and branches.

We utilized the CloudCompare command line interface (version 2.13.1) to run an automatic preprocessing step (our implementation script is provided as Supplementary Material) on each scan, which applied the subsampling and intensity filter steps, and converted the point clouds to .las format. After processing, the converted files had a mean size of 1.71 GB (SD = 0.60 GB), while the scans contained a mean of 57 million points (SD = 20 million points), reducing the average scan by 44 million points.

3.2. Methodology

Our assessment methodology was divided into three phases: tool selection, tool calibration, and tool assessment. These phases were conducted separately for the stem segmentation tools and stem metric extraction tools. We first selected, calibrated and assessed tools suitable for stem segmentation using the full point cloud scans from the validation dataset. We then selected and assessed tools suitable for stem metric extraction using the segmented stems from the previous step. These processes are described in the following sections.

3.2.1. Initial Tool Selection

Our tool selection involved identifying all suitable tools from those listed in Table 1. We excluded any tool that did not support automation. We also excluded any tool we could not get to work on available software and hardware platforms. For stem segmentation tools, we included every tool in Table 1 with a verified tree segmentation feature, totaling three tools. For stem metric extraction tools, we included every tool in Table 1 with either a verified DBH or verified height feature, totaling five tools.

3.2.2. Tool Calibration

Each tool was assessed, first to determine whether it could process every scan in our validation dataset and, second, to calibrate settings to achieve the best possible results. We calibrated the stem extraction tools by testing them on small subsets of validation data and comparing detected tree positions to the ground-truth field maps. After evaluating each tool’s performance with default settings, we systematically adjusted parameters to optimize results. To calibrate the stem metric extraction tools, we gauged how realistic the results were and how consistent they were between repeated runs. There were typically fewer settings to modify for these tools. For some tools, we adjusted the detection thresholds and segmentation parameters to maximize accuracy. We provide the fully modified code in our Supplementary Material for these instances. Any tool for which we could not reproduce consistent, meaningful results was considered unsuccessfully calibrated and excluded from subsequent analysis.

3.2.3. Candidate Tools

After selection criteria, calibration and testing, our list of suitable tools for stem segmentation included the following: FSCT, Forest3D and TreeLS.

After selection criteria, calibration and testing, our list of suitable tools for stem metric extraction included the following: PCTM, FSCT, ITSMe, TreeLS and TreeQSM.

3.2.4. Tool Assessment

Our tool validation was divided between processes for stem segmentation tools and processes for stem metric extraction tools. These two processes are detailed as follows:

Stem Detection Tool Assessment

In order to assess the outputs of the stem segmentation tools, each of the 27 scans from Section 3.1 was processed with the identified tools from Section 3.2.3, with their most effective calibration settings. To validate these outputs, an R script (provided in Supplementary Material) was used to compare the positions of detected trees in the scans to the measured tree locations in the validation dataset.

This analysis calculated the number of correct detections in the form of the number of true positives, false positives and false negatives. Hit rate (the proportion of correctly identified trees out of the total) and error rate (the proportion of incorrect identifications out of the total) were calculated after the analysis. The ground-truth data tree positions were used to generate a field measurement area polygon, representing the area within which a tree in the field would be recorded. Detected trees outside this polygon were removed from the analysis as they would be outside the ground-truth field measurement area. Detected trees that matched to the nearest ground-truth validation point within a maximum distance threshold (1 m) were tagged as true positives, while unmatched ground-truth trees were tagged as false negatives, and unmatched detected trees were tagged as false positives, as illustrated in Figure 3.

As a result of detecting multiple trees from a single stem, tools FSCT and Forest3D necessitated a merging step that combined trees based on proximity. Given that production plantations are planted a consistent, known distance apart, it was possible to combine tree detections within a small threshold distance to reduce detection noise. We applied this merging step, combining scan outputs within a maximum threshold distance (1 m), and setting the new location based on the output with the largest DBH (as our experimentation showed that this was most effective for correctly spacing trees).

The final calculated rates of false positives, false negatives, and true positives were analyzed using a Kruskal–Wallis test on a per-tool basis and for each development stage. Where significance was found, we employed a Dunn test for post-hoc comparisons.

Stem Metric Extraction Tool Assessment

We were concerned with two specific inventory metrics: DBH and height. We first needed to extract point clouds for each tree to validate these metrics. We elected to use the stem segmentation tool, which performed best in the prior stem detection assessment. We used a filtering script to isolate the individual tree point clouds and save them as single tree scans. These single tree scans were matched to their real-world twin tree and assigned ground-truth height and DBH measurements from the validation data.

Each of the inventory generation tools then processed these tree scans, and the resulting extracted inventory metrics were compared against the field measurements to assess their accuracy. We calculated the Root Mean Squared Error (RMSE) and Mean Bias Error (MBE) [5] of these measures as follows:

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{{\sum }_{i=1}^n{({\widehat{y}}_i-y_i)}^2}{n}}}$$

$$\mathrm{MBE}={\displaystyle \frac{1}{n}}\sum _{i=1}^n{y}_i-{\widehat{y}}_i$$

where $y_i$ represents field measurements, ${\widehat{y}}_i$ represents predicted measurements from point clouds, $\widehat{y}$ is the average of the observed values, and n represents the sample size. To assess the variance between height and DBH measurements across each tool, we applied Levene’s test to check for homogeneity of variance, followed by Welch’s ANOVA to determine significance. Where significance was determined, we employed a Games–Howell test for post-hoc pairwise comparisons.

4. Results

The results of our processing, review, and subsequent statistical analysis are presented in this section. All the statistical analysis was conducted in RStudio [17]. Dunn tests were conducted using the FSA package [18], and Games–Howell tests were conducted using the rstatix package [19].

4.1. Stem Segmentation

Results of stem segmentation across all plots for each tool are shown in

Table 3. Mean ($\overline{x}$) and Std. Dev ($\sigma$) of detection ratios for each tool across the different development stages.

Tool Name	Development Stage	True Positive		False Negative		False Positive
		$\overline{\mathit{x}}$	$\mathit{\sigma }$	$\overline{\mathit{x}}$	$\mathit{\sigma }$	$\overline{\mathit{x}}$	$\mathit{\sigma }$
FSCT
	All	0.904	0.174	0.096	0.174	0.141	0.189
	T1	0.808	0.079	0.192	0.079	0.107	0.105
	T2	0.884	0.259	0.116	0.259	0.033	0.034
	T3	0.991	0.018	0.009	0.018	0.274	0.247
Forest3D
	All	0.939	0.074	0.061	0.074	0.109	0.210
	T1	0.856	0.095	0.144	0.095	0.072	0.080
	T2	0.955	0.038	0.045	0.038	0.022	0.027
	T3	0.981	0.027	0.019	0.027	0.221	0.315
TreeLS
	All	0.953	0.081	0.047	0.081	0.141	0.307
	T1	0.987	0.018	0.013	0.018	0.097	0.063
	T2	0.959	0.073	0.041	0.073	0.009	0.030
	T3	0.925	0.109	0.075	0.109	0.302	0.469

and illustrated in Figure 4. Information about processing times and calculated stocking estimations compared to field stockings is detailed in

Table 4. Summary of average stem segmentation results.

Tool	Development Stage	Processing Time (m)	Stocking (per ha)
			Field	Tool
FSCT
	T1	424	702	644
	T2	231	449	413
	T3	182	244	326
Forest3D
	T1	110	702	659
	T2	67	449	439
	T3	44	244	307
TreeLS
	T1	19	702	763
	T2	11	449	436
	T3	7	244	314

. The outcomes show an average of over 90% successful detections (true positives) across each tool. TreeLS had the highest ratio of true positives to false negatives, detecting an average of 95.3% of stems within the validation area of each scan. Notably, Forest3D had the lowest rate of false positives after the merging step (10.9%) compared to FSCT and TreeLS (14.1% and 14.1%, respectively). For T1 and T2 scans, TreeLS was most effective, detecting 98.7% and 95.9% of stems, respectively. However, for T3, FSCT had the highest successful detection ratio at 99.1%.

The Kruskal–Wallis test yielded significant variation among the tools at T1 ($x^2$ = 14.841, p = <0.001, df = 2). A post-hoc Dunn test showed that TreeLS differed significantly from Forest3D ($Z=2.595$, $p=\mathbf{0}.\mathbf{012}$) and FSCT ($Z=3.763$, $p<\mathbf{0}.\mathbf{001}$) There was no statistically significant difference among tools for T2 and T3, and no variation in false positive rates across any stage.

4.2. Stem Metric Extraction

As we determined TreeLS to have the most effective stem segmentation overall, we utilized the segmented stems from this tool for the inventory analysis step. Inventory analysis processing times are detailed in

Table 5. Summary of approximate per tree processing time for stem metric analysis for each toolkit.

Toolkit	Analysis Time (/tree)
FSCT	148 s
TreeLS	N/A
ITSMe	2 s
PCTM	86 s
TreeQSM	30 s

, but TreeLS is omitted as it generated inventory metrics during its stem segmentation.

4.2.1. Height Assessment

We found relatively consistent height results overall for each of the assessed tools, as evidenced in Figure 5, but we also found statistically significant differences between them. Our Levene’s test indicated significant differences in variances among the tools (F(4, 129) = 3.1605, p = 0.016), violating the assumption of equal variances. Our subsequent Welch’s ANOVA indicated a significant effect of tool type on RMSE (F(4, 63.643) = 934.44, p < 0.001), indicating that RMSE varied significantly across the tools.

Post-hoc pairwise analysis is shown in

Table 6. Pairwise comparisons of RMSE of height results between tools for all development stages.

Contrast	Mean Difference	SE	df	p Value	95% CI
FSCT-ITSMe	$-2.150$	$0.259$	$40.6$	$<\mathbf{0.001}$	$[-3.190,-1.110]$
FSCT-PCTM	$-2.180$	$0.260$	$41.0$	$<\mathbf{0.001}$	$[-3.230,-1.130]$
FSCT-TreeLS	$-2.260$	$0.261$	$41.6$	$<\mathbf{0.001}$	$[-3.310,-1.200]$
FSCT-TreeQSM	$-2.280$	$0.253$	$38.5$	$$<\mathbf{0.001}$$	$[-3.310,-1.260]$
ITSMe-PCTM	$-0.031$	$0.187$	$52.0$	$1.000$	$[-0.778,0.716]$
ITSMe-TreeLS	$-0.107$	$0.189$	$51.9$	$0.994$	$[-0.864,0.649]$
ITSMe-TreeQSM	$-0.131$	$0.178$	$51.6$	$0.985$	$[-0.842,0.580]$
PCTM-TreeLS	$-0.077$	$0.191$	$52.0$	$0.999$	$[-0.839,0.685]$
PCTM-TreeQSM	$-0.100$	$0.179$	$51.5$	$0.995$	$[-0.817,0.617]$
TreeLS-TreeQSM	$-0.024$	$0.182$	$51.2$	$1.000$	$[-0.751,0.703]$

. FSCT had a significantly higher absolute error compared to all other tools (all p < 0.001), and was unable to produce results for a proportion of trees across age classes. We observed no significant difference between PCTM, ITSME, TreeLS and TreeQSM. These tools performed very consistently for height measurement, achieving an $R^2$ of approximately 0.91. FSCT, by contrast, appeared to underestimate tree height, but did so consistently ($R^2$ = 0.79). Among the best performing tools, TreeQSM had the lowest RMSE ($1.32$ $\mathrm{m}$), while TreeLS had the smallest MBE ( $0.89$ $\mathrm{m}$).

Analysis of effectiveness on individual development stages yielded similar results, as evidenced in Figure 6 and detailed in Appendix B 

Table A1. Pairwise comparisons of height RMSE across each thinning.

Thin	Contrast	Mean Difference	SE	df	p Value	95% CI
T1	FSCT-ITSMe	$-1.642$	$0.345$	$8.3$	$0.054$	$[-3.312,0.027]$
	FSCT-PCTM	$-1.733$	$0.342$	$8.1$	$\mathbf{0.041}$	$[-3.399,-0.067]$
	FSCT-TreeLS	$-1.847$	$0.339$	$7.9$	$\mathbf{0.030}$	$[-3.509,-0.184]$
	FSCT-TreeQSM	$-1.947$	$0.320$	$6.4$	$\mathbf{0.025}$	$[-3.608,-0.285]$
	ITSMe-PCTM	$-0.091$	$0.196$	$12.0$	$0.997$	$[-0.972,0.791]$
	ITSMe-TreeLS	$-0.204$	$0.190$	$11.9$	$0.937$	$[-1.061,0.653]$
	ITSMe-TreeQSM	$-0.304$	$0.153$	$8.0$	$0.642$	$[-1.052,0.444]$
	PCTM-TreeLS	$-0.114$	$0.185$	$12.0$	$0.992$	$[-0.948,0.721]$
	PCTM-TreeQSM	$-0.213$	$0.147$	$8.2$	$0.838$	$[-0.929,0.502]$
	TreeLS-TreeQSM	$-0.100$	$0.140$	$8.5$	$0.985$	$[-0.773,0.573]$
T2	FSCT-ITSMe	$-2.224$	$0.497$	$13.8$	$\mathbf{0.046}$	$[-4.419,-0.028]$
	FSCT-PCTM	$-2.279$	$0.498$	$13.9$	$\mathbf{0.041}$	$[-4.479,-0.080]$
	FSCT-TreeLS	$-2.473$	$0.490$	$13.3$	$\mathbf{0.023}$	$[-4.646,-0.300]$
	FSCT-TreeQSM	$-2.433$	$0.472$	$12.0$	$\mathbf{0.023}$	$[-4.557,-0.308]$
	ITSMe-PCTM	$-0.056$	$0.379$	$18.0$	$1.000$	$[-1.677,1.566]$
	ITSMe-TreeLS	$-0.249$	$0.368$	$17.9$	$0.988$	$[-1.821,1.323]$
	ITSMe-TreeQSM	$-0.209$	$0.343$	$17.2$	$0.992$	$[-1.683,1.265]$
	PCTM-TreeLS	$-0.193$	$0.369$	$17.9$	$0.996$	$[-1.773,1.387]$
	PCTM-TreeQSM	$-0.153$	$0.345$	$17.2$	$0.998$	$[-1.636,1.330]$
	TreeLS-TreeQSM	$0.040$	$0.332$	$17.6$	$1.000$	$[-1.384,1.464]$
T2	FSCT-ITSMe	$-2.453$	$0.428$	$14.7$	$\mathbf{0.008}$	$[-4.328,-0.579]$
	FSCT-PCTM	$-2.417$	$0.430$	$14.8$	$\mathbf{0.009}$	$[-4.297,-0.536]$
	FSCT-TreeLS	$-2.351$	$0.445$	$15.9$	$\mathbf{0.013}$	$[-4.281,-0.422]$
	FSCT-TreeQSM	$-2.385$	$0.431$	$14.9$	$\mathbf{0.010}$	$[-4.268,-0.501]$
	ITSMe-PCTM	$0.037$	$0.313$	$18.0$	$1.000$	$[-1.302,1.376]$
	ITSMe-TreeLS	$0.102$	$0.333$	$17.7$	$0.999$	$[-1.326,1.530]$
	ITSMe-TreeQSM	$0.068$	$0.314$	$18.0$	$1.000$	$[-1.276,1.413]$
	PCTM-TreeLS	$0.066$	$0.336$	$17.8$	$1.000$	$[-1.372,1.503]$
	PCTM-TreeQSM	$0.032$	$0.317$	$18.0$	$1.000$	$[-1.323,1.387]$
	TreeLS-TreeQSM	$-0.034$	$0.337$	$17.8$	$1.000$	$[-1.476,1.408]$

. T1 plots showed high correlation, despite the complexity of T1 scan data. FSCT was the only statistical outlier with an $R^2$ of 0.28, consistently underestimating tree height. The difference between FSCT and the other tools—PCTM, TreeLS and TreeQSM—was statistically significant. ITSMe was the only tool that was not significantly different to FSCT (p = 0.054) TreeQSM showed marginally higher $R^2$ than the other tools (0.78), in addition to the lowest RMSE (0.94 m) and the smallest MBE (0.31 m).

T2 Plots had the highest overall correlation. FSCT was again the only statistical outlier (p < 0.001 contrasted against all other tools). TreeLS and TreeQSM had the highest correlation ($R^2$ = 0.81, 0.81). TreeLS had the smallest RMSE (1.53 m) whilst TreeQSM had the smallest MBE (0.22 m).

T3 saw the lowest correlation. FSCT was once more the only statistical outlier (p < 0.001 when contrasted against all other tools). TreeQSM, TreeLS, PCTM and ITSMe were roughly equivalent, with $R^2$ of ∼0.36 and RMSE of ∼ $1.7$ $\mathrm{m}$ for each. TreeQSM saw the highest $R^2$ and error (0.37 and $1.28$ $\mathrm{m}$, respectively), though this was close to parity with PCTM, ITSME and TreeLS. In this instance, ITSMe had the highest accuracy, scoring RMSE of $1.69$ $\mathrm{m}$ and MBE of only $0.04$ $\mathrm{m}$.

4.2.2. DBH Assessment

The DBH assessment score showed a considerable variation between tools, as evidenced in Figure 7, with statistical analysis showing significant differences. Our Levene’s test showed significant differences between variances among the tools (F(4, 129) = 6.0023, p < 0.001)) violating the assumption of equal variances. As with height, we again utilized Welch’s ANOVA to determine whether absolute error differed significantly among the tools. Welch’s ANOVA indicated a significant effect of tool type on the RMSE (F(4, 62.485) = 9.2977, p < 0.001), suggesting that the RMSE varied significantly across the tools.

Post-hoc pairwise comparisons (illustrated in

Table 7. Pairwise comparisons of RMSE of DBH results between tools for all development stages.

Contrast	Mean Difference	SE	df	p Value	95% CI [Lower, Upper]
FSCT-ITSMe	$10.600$	$1.870$	$29.8$	$\mathbf{0.003}$	$[2.940,18.300]$
FSCT-PCTM	$2.520$	$1.340$	$33.8$	$0.677$	$[-2.960,8.000]$
FSCT-TreeLS	$-4.830$	$0.940$	$42.9$	$\mathbf{0.006}$	$[-8.620,-1.050]$
FSCT-TreeQSM	$-1.690$	$0.826$	$47.5$	$0.602$	$[-5.000,1.620]$
ITSMe-PCTM	$-8.090$	$2.200$	$46.3$	$0.086$	$[-16.900,0.723]$
ITSMe-TreeLS	$-15.400$	$1.970$	$35.9$	$<\mathbf{0.001}$	$[-23.500,-7.430]$
ITSMe-TreeQSM	$-12.300$	$1.920$	$32.9$	$<\mathbf{0.001}$	$[-20.100,-4.460]$
PCTM-TreeLS	$-7.350$	$1.490$	$44.3$	$<\mathbf{0.001}$	$[-13.300,-1.380]$
PCTM-TreeQSM	$-4.210$	$1.420$	$39.6$	$0.240$	$[-9.930,1.510]$
TreeLS-TreeQSM	$3.140$	$1.040$	$50.3$	$0.221$	$[-1.020,7.310]$

) revealed that there were no significant differences between PCTM, FSCT and TreeQSM. However, TreeLS and ITSMe varied significantly from most other tools. ITSME had the highest RMSE (22.47 cm) and highest bias (10.35 cm), while TreeLS showed had the lowest RMSE (8.74 cm) and lowest bias (4.58 cm). FSCT had the highest $R^2$ across all development stages (0.38).

The efficacy of different tools across development stages provided more precise insights into their strengths and weaknesses, as outlined in Figure 8 and detailed in Appendix B 

Table A2. Pairwise comparisons of DBH RMSE across each thinning.

Thin	Contrast	Mean Difference	SE	df	p Value	95% CI
T1	FSCT-ITSMe	$10.873$	$2.206$	$7.6$	$0.051$	$[-0.042,21.788]$
	FSCT-PCTM	$11.937$	$2.086$	$7.8$	0.024	$[1.687,22.186]$
	FSCT-TreeLS	$4.162$	$2.044$	$7.9$	$0.623$	$[-5.853,14.177]$
	FSCT-TreeQSM	$3.830$	$1.672$	$9.0$	$0.521$	$[-4.127,11.787]$
	ITSMe-PCTM	$1.064$	$2.837$	$12.0$	$0.999$	$[-11.733,13.861]$
	ITSMe-TreeLS	$-6.711$	$2.806$	$11.9$	$0.474$	$[-19.376,5.953]$
	ITSMe-TreeQSM	$-7.043$	$2.548$	$10.9$	$0.347$	$[-18.718,4.632]$
	PCTM-TreeLS	$-7.775$	$2.713$	$12.0$	$0.311$	$[-20.004,4.454]$
	PCTM-TreeQSM	$-8.107$	$2.445$	$11.2$	$0.201$	$[-19.248,3.034]$
	TreeLS-TreeQSM	$-0.332$	$2.409$	$11.4$	$1.000$	$[-11.289,10.626]$
T2	FSCT-ITSMe	$11.236$	$4.035$	$9.8$	$0.346$	$[-7.624,30.095]$
	FSCT-PCTM	$1.838$	$2.541$	$11.1$	$0.984$	$[-9.766,13.442]$
	FSCT-TreeLS	$-7.418$	$0.891$	$10.4$	$<\mathbf{0.001}$	$[-11.531,-3.305]$
	FSCT-TreeQSM	$-2.326$	$1.263$	$16.9$	$0.694$	$[-7.763,3.112]$
	ITSMe-PCTM	$-9.398$	$4.622$	$14.9$	$0.615$	$[-29.608,10.813]$
	ITSMe-TreeLS	$-18.654$	$3.963$	$9.1$	$0.052$	$[-37.443,0.136]$
	ITSMe-TreeQSM	$-13.561$	$4.062$	$10.0$	$0.203$	$[-32.453,5.331]$
	PCTM-TreeLS	$-9.256$	$2.424$	$9.3$	$0.128$	$[-20.695,2.183]$
	PCTM-TreeQSM	$-4.164$	$2.584$	$11.8$	$0.783$	$[-15.849,7.521]$
	TreeLS-TreeQSM	$5.093$	$1.007$	$11.1$	$\mathbf{0.028}$	$[0.494,9.691]$
T2	FSCT-ITSMe	$9.786$	$2.928$	$10.9$	$0.197$	$[-3.621,23.193]$
	FSCT-PCTM	$-3.407$	$1.171$	$17.2$	$0.282$	$[-8.441,1.627]$
	FSCT-TreeLS	$-8.565$	$0.940$	$10.0$	$<\mathbf{0.001}$	$[-12.941,-4.190]$
	FSCT-TreeQSM	$-4.936$	$1.117$	$16.1$	$\mathbf{0.044}$	$[-9.771,-0.101]$
	ITSMe-PCTM	$-13.193$	$2.876$	$10.2$	$0.053$	$[-26.518,0.131]$
	ITSMe-TreeLS	$-18.351$	$2.790$	$9.1$	$\mathbf{0.008}$	$[-31.583,-5.119]$
	ITSMe-TreeQSM	$-14.722$	$2.854$	$10.0$	$\mathbf{0.029}$	$[-28.018,-1.426]$
	PCTM-TreeLS	$-5.158$	$0.762$	$10.5$	$\mathbf{0.005}$	$[-8.669,-1.646]$
	PCTM-TreeQSM	$-1.528$	$0.972$	$17.7$	$0.798$	$[-5.692,2.635]$
	TreeLS-TreeQSM	$3.629$	$0.675$	$11.0$	$\mathbf{0.020}$	$[0.540,6.719]$

. T1 plots had the least accuracy for validation overall. The only statistically significant difference was detected between FSCT and PCTM (p = 0.024), with the former performing most effectively. FSCT had the lowest RMSE (7.23 cm) and smallest bias (2.29 cm) by a considerable margin, though correlations were low across the board at this development stage.

T2 saw higher accuracy across all tools, but TreeLS was the most accurate by a considerable margin, producing an $R^2$ of 0.82 with a very low RMSE of 2.75 cm. Interestingly, despite its higher RMSE and lower $R^2$, TreeQSM had a lower MBE by a factor of 2. We only observed statistically significant differences between TreeLS and both FSCT and TreeQSM for this age class.

Accuracy was highest for all tools at T3, with the exception of FSCT. TreeLS was again the most accurate, with an $R^2$ of 0.9 and an RMSE of only 1.18 cm. PCTM, similar to TreeQSM at T2, showed the lowest MBE with only −0.61 cm, compared to TreeLS’s of −0.91 cm. We observed a statistically significant difference between TreeLS and each other tool, as it performed exceptionally well with T3 scans.

5. Discussion

The discussion of our results is divided between stem segmentation, individual tree metrics, suggested implementation, limitations, and future work. There is a considerable breadth of both methodologies for inventory generation and open-source implementation of approaches, along with the state-of-the-art of open-source inventory tools and their potential for the future of forest inventory assessment.

5.1. Stem Segmentation

Tree detection performance varied across development stages. Post first thinning (T1) plots were the most challenging to segment due to higher tree density and the presence of branches and foliage low on the stem. In contrast, post-third thinning (T3) plots, characterized by larger, widely spaced trees and self-pruned stems, were expected to facilitate more accurate segmentation. However, our results indicate that while segmentation accuracy was highest in T1 plots, T3 plots exhibited an increased prevalence of false positives, primarily attributed to the presence of pine wildlings and undergrowth. Although these false positives often corresponded to correctly detected vegetation, they are not of interest to a wood-volume-oriented forest inventory. In standard field procedures, including those used in this study, such vegetation is not enumerated.

As expected, we found that the development stage had a significant impact on tool performance: In our analysis we found that TreeLS had the highest ratio of true positives to false negatives, with over 95% overall, and accuracy being highest at T1. The TreeLS functions used in this study use a K-Nearest Neighbors (KNN) [20] tree mapping algorithm to detect tree regions. The Hough transformation algorithm [21] is used for stem denoising, and Iterated Reweighted Least Squares (IRLS) [22] for stem modeling. This approach is robust, as Hough transformations effectively preserve stem features at greater canopy heights [9], improving the accuracy of the subsequent stem modeling phase by maintaining a more complete point cloud dataset. TreeLS performed worst at T3, which is likely due to the calibration conducted in Section 3.2.2, favoring younger stands. FSCT and Forest3D demonstrated more accurate detection than TreeLS at T3, with FSCT scoring slightly higher despite having more false positives. These differences can be attributed to the underlying segmentation approaches. Forest3D (using the non-machine learning pipeline) employs layer-based circle fitting, non-maxima suppression, and rule-based clustering. These approaches provide performance consistent across thinning stages. FSCT utilizes a PointNet++ [23] based deep learning approach for stem detection. The toolkit struggled to detect T1 stems, likely due to not being trained on sufficiently similar small pine trees [12]. We expect that training a model on such a dataset would significantly improve results. These findings suggest that an optimal segmentation strategy would involve using TreeLS for T1 and T2 and FSCT for T3, yielding an overall segmentation accuracy of approximately 97.9%. A unique calibration for each tool on the relevant development stages would likely further improve results.

Improving the removal of false positives is critical to inventory assessment. In a production environment, failure to eliminate false positives would unacceptably bias inventory results. However, it often is possible to detect false positives because they tend to have combinations of attributes, either of the segmented tree point cloud, or the estimated tree metrics, that are atypical and specific. Detected false positives can then be removed in a post-processing step. This is an area of future research. While our validation approach excluded false positives from final measurements, refining segmentation algorithms to eliminate these errors is critical for automated inventory assessments.

Beyond segmentation accuracy, processing time is a key consideration for practical implementation. TreeLS consistently exhibited the shortest processing times across all development stages, completing T3 scans in an average of seven minutes, whereas FSCT required 182 min on the same hardware. While these performance gaps may be reduced through software optimization, tailored calibration, and pre-processing, it is unlikely they can be entirely eliminated due to the differences in computational costs of the algorithms underpinning each of the tools. The observed differences highlight an order-of-magnitude disparity in processing performance. Although computational hardware capabilities will continue to improve, the demand to process increasingly large datasets is likely to grow correspondingly, maintaining a comparable order-of-magnitude difference. Therefore, faster tools such as TreeLS may remain preferable in operational contexts where computational resources or processing time are limited.

5.2. Individual Tree Metrics

We found encouraging results across all metrics and a significant variation in performance across development stages.

5.2.1. Height Assessment

Height assessment was relatively consistent across all tools but varied across development stages. TreeQSM produced the best results for T1, while TreeQSM and TreeLS provided similar results for T2. ITSMe and TreeQSM were the most accurate at T3, with ITSMe yielding slightly lower RMSE and lower relative bias. Based on these results, we can determine that a stratified approach—using TreeQSM for T1 and T2 and ITSMe for T3—appears to be the most effective strategy under current calibration conditions.

Approaches to height measurements are largely consistent between tools, with most estimating height as the vertical distance between the highest and lowest detected points in the tree point cloud. However, approaches to tree segmentation vary, impacting the lowest and highest points for each tree. FSCT and TreeLS incorporate statistical filtering to eliminate noise. ITSMe utilizes a DTM or calculates the median ground height when a DTM is unavailable. This may be responsible for its higher accuracy at T3, as it can better account for ground height uncertainty in terrain with complex understory vegetation. TreeQSM differs from the other methods, deriving height from the dimensions of its fitted cylinders within the QSM model, which may explain its consistently high performance.

The margin of error for height measurements obtained from the field is $\approx 2$ $\mathrm{m}$, a value comparable to the RMSE of our LiDAR-derived estimates across all development stages. This suggests that the automated extraction methods achieve at least the same level of accuracy as traditional field measurements.

Despite these strong results, T3 remains more challenging to assess due to the inherent limitations in terrestrial LiDAR scanning. The reduced density of canopy returns at greater heights—caused by occlusion and limited vertical coverage—introduces uncertainty in height measurements. Airborne LiDAR, or a fusion of TLS and ALS data, could mitigate this issue by providing a more complete representation of tree crowns. FSCT’s tendency to underestimate height is likely a consequence of its feature detection methodology, which incorporates a leaf/wood segmentation step. This segmentation performs best with a complete scan, and as such, the sparser canopy data are likely biasing the leaf/wood segmentation towards the ground. With more complete canopy scans, FSCT and QSMs are expected to perform much better overall, providing more detail than a simple height measurement.

5.2.2. DBH Assessment

DBH estimation exhibited greater variability than height measurement, with performance differences attributable to tool-specific methodologies and the structural complexity of different development stages. TreeLS produced the most accurate overall results, with the highest $R^2$, lowest RMSE and lowest MBE. However, we saw better results for individual development stages: FSCT provided the best results for T1 plots, producing results with an RMSE of 7.23 cm and a small negative bias while TreeLS performed best at T2 and T3, attaining RMSE values of 2.75 cm and 1.81 cm, respectively.

The complexity of T1 scans, characterized by dense branching and understory vegetation, influenced DBH estimates by occasionally including non-target vegetation, leading to overestimation. FSCT calculates DBH using a mean fitted cylinder diameter, derived from RANSAC circle fitting during its stem segmentation step. This likely explains its superior performance at T1, as it effectively filters out smaller shrubs and understory interference. In contrast, ITSMe and PCTM apply least squares circle fitting algorithms, while TreeQSM determines DBH from the closest cylinder in its QSM model, neither of which performed as accurately. The IRLS approach utilized by TreeLS demonstrated the most consistent accuracy across all tree development stages.

5.2.3. Comparison to Related Work

These results align with prior research on LiDAR-based forest inventory while demonstrating improvements attributable to more recent scanning technologies. Liang et al. [4] evaluated correctness (analogous to our true positive rate) and completeness (a measure of how much of the reference tree was successfully detected) using data from a Leica HDS6100. They reported “best efforts” yielding correctness and completeness rates as high as 93.6% and 90.4%, respectively. Kükenbrink et al. [24] performed a similar evaluation, calculating Tree Detection Rates (TDR) using data from a FARO Focus 3D S120 and a Leica BLK360. They, in turn, reported a TDR of 94.6% (leaf-off) and 82% (leaf-on) Our correctness rates of 98.7% (T1), 95.5% (T2), and 99.1% (T3) surpass these benchmarks. While these results are all for terrestrial LiDAR, our increase in accuracy is partially attributable to the advantageous layout of production forests and the affordances of MLS LiDAR.

We also compared our results for individual tree parameters. Liang et al. [4] reported DBH RMSE values below 2 cm for multiple development stages, and Kükenbrink et al. [24] reported DBH extraction performance with RMSE ranging from 2% to 9%. Our T3 results are on par with both of these benchmarks. However, our estimates for younger stands exhibited greater variance than both studies, suggesting that additional refinement is necessary for high-density forest conditions. Our height estimations yielded lower RMSE values than those reported in Liang et al. [4], where the most accurate estimates reached 2.4 m. In contrast, we achieved errors of less than 1 m, a difference likely driven by advancements in LiDAR technology and the adoption of mobile terrestrial LiDAR. However, these results are still outperformed by the ALS tree measurement; Sparks and Smith [25] were able to estimate tree height from ALS data with an RMSE of $0.69$ $\mathrm{m}$, functionally identical to field measurements.

5.3. Suggested Implementation

Based on our findings, we determined that the use of a stratified toolchain approach—using the most effective tool for each development stage—would be the most effective for setting up an automated industrial implementation for our area of interest (radiata pine in the Green Triangle region). The specific tools to implement at each development stage are detailed in

Table 8. Summary of ideal tool choices for each inventory metric at each development stage.

	Stem Segmentation	DBH Tool	Height Tool
T1	TreeLS	FSCT	TreeQSM
T2	TreeLS	TreeLS	TreeQSM
T3	FSCT	TreeLS	ITSMe

.

Furthermore, we suggest utilizing CloudCompare for point cloud processing (filtering and subsampling). We tested a simple batch file-based implementation of this setup (Included in the Supplementary Material) and produced results consistent with those observed throughout the study. As this setup contains a mix of R packages (TreeLS, ITSMe), MATLAB packages (TreeQSM), C++ tools (CloudCompare) and Python packages (FSCT), we suggest implementing it using Python as a wrapper: Each component of the toolchain can be called from Python using additional tools: MATLAB via the MATLAB^® Engine API for Python [26], R via rpy2—R in Python [27], and CloudCompare using the CloudComPy Python wrapper [28].

Based on our findings, this implementation would produce the most precise inventory metrics from our scans. Our specific methodology is likely to be suitable for other similarly structured tree species in similar locales, and the stratified toolchain approach is likely to be an appropriate option for the management of other forests worldwide. It is likely to be beneficial to specialize further and set up unique calibrations for different locations and site qualities at broader management levels. While our implementation focused on a narrow set of essential metrics, a variety of additional parameters relevant to forest management could be calculated depending on the specific priorities and requirements of forest managers. However, it is important to recognise that the ability to extract detailed information from forest point clouds does not directly equate to meeting industry requirements. These are distinct challenges: one involving technical precision in data processing and the other addressing practical, operational, and decision-making needs in forestry management workflows. Bridging this gap will require continued collaboration between researchers, technologists, and forestry practitioners.

5.4. Limitations

We were unable to reproduce consistent, meaningful results with some tools, so we did not include them for validation. However, this does not suggest that these tools cannot produce meaningful results; it simply means that the setup and calibration on our specific dataset were not suited to the toolkit.

We did not assess the precision of detections; the exact centers of trees and relative offsets were only examined as far as the relative distance between plantings. Investigating the true accuracy of tree detections will become important when attempting to fuse these data with GPS/GNSS data and aligning scans over multiple timescales.

This assessment is inherently limited by its focus on open-source, offline desktop tools. Private industrial solutions will likely offer more accurate methodologies when personalized to the forest of interest. Additionally, it would be possible to obtain higher precision in measurements through the use of more powerful computing hardware. Specialized hardware with more memory would facilitate analysis with less preprocessing, minimizing the data lost through the subsampling step.

The preprocessing step was tailored to maximize results for T1 scans. However, as the latter development stages (T2 and T3) are less affected by noise, extensive preprocessing may not be necessary. Stratifying the preprocessing approach in the same manner as the toolchain will likely produce the best results overall.

Our individual tree analysis step is completed with no additional processing after the tree scans are extracted. A noise reduction step may improve these measures. Our validation dataset did not include any edge plots. It also did not include any information on crown shape, form, taper, or faults, all of which could have been used to test the precision of further tools. Our study focused exclusively on radiata pine within the Green Triangle region, so our findings are not perfectly transferable to other regions or species. Operators in other areas will need to conduct their own calibration for best results. As we only assessed production forests, our findings are not necessarily applicable to mixed-species forests, which would likely need different approaches to stem segmentation.

5.5. Future Work

Based on our findings, future research should focus on improving stem detection and reducing false positives, as well as refining diameter measurements for young trees. There is scope for improvement of metrics related to ecological assessment, such as biomass estimation and leaf area index. Improving the computational efficiency of existing algorithms presents an important area for future work, particularly to enable faster and more scalable processing of large datasets. Additionally, addressing the under-explored area of automated feature detection, including identifying double leaders, would be of significant interest in forestry management. The rapid development of new tools and advanced AI methods presents opportunities to refine stem and feature detection, warranting continued assessment and exploration. Moreover, a comprehensive evaluation of tree position detection accuracy across tools is crucial for supporting future data fusion research and integration with forest digital twins.

We note the potential utility of an additional preprocessing step on individual tree point clouds prior to stem metric extraction. Improving false positive removal through retroactive feature analysis would be advantageous for maximizing the precision of stem detection. Exploring the potential improvements in processing time and accuracy resulting from enhanced hardware and alternative preprocessing approaches is crucial. For our proposed stratified pipeline, we need to generalize our findings through testing available reference benchmark datasets (such as [29,30]), ensuring robustness and accuracy. There is potential utility for automatic calibration tools for industrial workflows. A modular toolkit that allows users to drag and drop tools and link outputs to inputs in a node-based process tree would enhance the usability and flexibility of these tools for both research and industry.

6. Conclusions

Our findings demonstrate that a stratified approach to tool selection for generating forest inventory metrics from LiDAR point cloud scans is a practical and effective methodology suitable for industry applications. This approach delivers substantially higher accuracy in biometric measurements compared to relying on single tools. Our systematic review found that a wide range of freely available open-source tools can meet the requirements of precision forestry. We also note that some feature gaps remain to be addressed, notably in fault detection.

Our calibration and experimentation showed that different tools excel at measuring different development stages. From these results, we conclude that utilizing a stratified approach would allow us to conduct forest inventory generation with the highest accuracy for radiata pine in the Green Triangle. Specifically, for stem segmentation, utilizing TreeLS for T1 and T2, and FSCT for T3 would facilitate stem detection rates of 98.7%,95.9% and 99.1%, respectively, a considerable improvement over using one tool for all development stages. For DBH, utilizing FSCT for T1 and TreeLS for T2 and T3 would facilitate measurements with a RMSE of 7.23 cm, 2.75 cm and 1.81 cm, respectively. For height, utilizing TreeQSM for T1 and T2, and ITSMe for T3 would facilitate measurements with a RMSE of 0.94 m, 1.53 m and 1.69 m, respectively. This approach leverages various platforms and programming languages, utilizing the strengths of each to obtain the best results. Furthermore, we outline directions for future work, including expanding future research and improving the tools for industry application.

Testing both stem segmentation and tree metrics in one study is inherently difficult; the results of the former are reflected in the results of the latter. The results of Height and DBH assessments depend on catching and removing false positives and accounting for false negatives from the overall stem segmentation. Our calibration is unlikely to yield the best results possible with open-source inventory tools, but should serve as a suitable reference point for future tool development.

The collaborative approach to forest assessment tools and techniques apparent across point cloud analysis is encouraging. Collaborative efforts are the most effective approach to advancing our understanding of our forests.