1. Introduction
Plantation forests play a critical role in meeting global wood supply demands, while providing raw materials for timber and paper industries [
1,
2,
3]. Beyond their economic value, they contribute significantly to carbon sequestration by acting as carbon sinks, thereby helping to mitigate climate change [
4,
5]. Many countries rely on fast-growing species such as radiata pine (
Pinus radiata D. Don.) to maintain sustainable forestry practices that deliver both economic and environmental benefits [
6]. In New Zealand, radiata pine accounts for over 90% of the plantation forestry sector, and these forests are crucial for the country’s economy, export industry, and carbon offset strategies [
7]. Consequently, accurate and efficient monitoring and management of these plantations are essential to maintain productivity, resilience, and long-term sustainability in a changing climate.
Effective management of plantation forests requires frequent and precise forest inventories to assess growth rates, timber yield, and overall forest health [
8,
9,
10]. Traditional inventory methods typically rely on ground-based measurements, where tree attributes such as diameter at breast height (DBH), tree height, and stem volume are collected manually. These measurements are often obtained at the stand level, using plot averages to estimate tree attributes across a plantation. However, while these methods are widely used, they have several drawbacks, including high labour costs, time-intensive data collection, sampling bias, and limited spatial coverage [
11]. The need for a more efficient, scalable, and precise inventory system has led to the increased adoption of remote sensing technologies.
Recent advancements in remote sensing technologies, particularly Light Detection and Ranging (LiDAR), have transformed forest inventory practices by enabling high-resolution, three-dimensional (3D) mapping of forest structures [
12,
13]. Airborne and terrestrial LiDAR systems have been extensively used to estimate tree-level attributes with high accuracy [
9,
14,
15]. More recently, unmanned aerial vehicle (UAV)-mounted LiDAR has emerged as a cost-effective and highly flexible alternative for detailed forest inventory, especially for tree-level measurements in plantation forests [
10,
16,
17,
18]. UAV-LiDAR systems offer high spatial resolution, reduced occlusion effects, and improved cost efficiency, making them ideal for precise and repeatable forest measurements. With an increasing number of applications globally and the proliferation of consumer-grade UAV-LiDAR systems, the accessibility and affordability of this technology have significantly improved [
10].
In operational inventory applications, LiDAR is most often utilised with an area-based approach (ABA) using LiDAR acquired from fixed-wing aircraft (ALS). The ABA leverages features derived from LiDAR surface models and canopy height point clouds to predict key stand attributes, using either parametric methods such as mixed-effects models or non-parametric techniques like the nearest neighbour imputation [
19,
20,
21]. While ABA facilitates the description of the spatial variation across a stand, usually at the resolution of a plot, it does not provide information on stand density or the distributions of heights and diameters [
8,
22]. This approach is well suited to characterising large forests but not as cost effective as the use of UAV-derived LiDAR for small- to moderate-sized stands [
23]. In contrast, the prediction of individual tree dimensions using UAV-LiDAR provides near complete censuses of stands and can provide detailed distributions of tree characteristics such as volume, height, and diameter. These outputs are especially valuable when repeat acquisitions are available, as they can be interpolated within growth simulators to determine likely volume production, grade outturn, and stand value [
22].
Despite the growing adoption of UAV-LiDAR using the ABA for tree inventory, significant knowledge gaps remain around application of this methodology. The accuracy of LiDAR-based predictions using the ABA has been shown to be highly dependent on sampling density, flight parameters, and data processing techniques [
11]. While the number of calibration plots can often be substantially reduced without unduly impacting the accuracy of LiDAR-derived forest estimates, changes in the predictive accuracy vary by forest type, site conditions, prediction method, and target attribute [
24,
25,
26]. This underscores the need for more strategic sampling designs, particularly if the aim is to reduce field effort while implementing UAV-LiDAR applications.
Compared with the ABA inventory method, predictions of key individual tree metrics, such as diameter and volume, from LiDAR are less well researched. LiDAR-derived height is often used as a predictor of tree diameter because of the strong relationship and the robustness of LiDAR height measurements [
27]. However, this relationship can break down in older stands, where competition effects lead to a wide range of diameters and volumes for a given height [
28,
29]. Thus, a number of studies have extended the set of predictor variables to include metrics describing the tree crown and the canopy structure [
10,
30]. In addition to ordinary least squares, commonly used parametric modelling approaches include both linear and non-linear mixed effects modelling, which are able to accommodate random variation between plots and stands [
31,
32]. Non-parametric methods have also been used such as random forest and
k-most similar neighbour [
33,
34]. Compared with parametric approaches, these non-parametric methods involve fewer assumptions and can better accommodate non-linear relationships and the collinearity between predictor variables often present in LiDAR data.
Although it is critical information for developing economically feasible operational workflows, few studies have examined how individual tree prediction from LiDAR is impacted by reducing the number of measured trees. Several studies have explored how many trees are required for calibrating mixed effects models at a site different to the region over which they were fitted [
32,
35,
36,
37]. However, little research has investigated the impact of reducing tree numbers for non-parametric methods such as random forest. Two practical scenarios are especially relevant, as follows: first, how many tree measurements are needed to accurately train a random forest model at a new site; and second, whether a generalised model trained across many sites can be used to accurately predict tree dimensions at a new site without any measurements. Given that machine learning methods such as random forest require considerable data, it is likely that more calibration measurements are required than for parametric approaches.
Addressing this research gap is crucial to reducing inventory costs, improving operational efficiency and fully realizing the benefits of LiDAR technology. A better understanding of the trade-offs between field effort and model accuracy will support the transition to single-tree inventory-based workflows. Ultimately, the goal of this research is to develop a generalised model that can accurately predict tree dimensions from UAV-LiDAR using minimal field measurements. This is particularly relevant for countries like New Zealand where the forestry sector is economically important and where plantation forests have a relatively homogenous structure [
7].
This study evaluates UAV-LiDAR-derived metrics for predicting tree-level diameter and volume in New Zealand radiata pine plantations across a national trial series. To achieve this goal, we used high-quality UAV-LiDAR datasets from 20 diverse sites—varying in age, stand density, and site quality—with matching field-measured tree attributes. From these data, 61 LiDAR-derived metrics were developed to characterise horizontal and vertical tree canopy structure. Using these LiDAR-derived metrics, random forest models were used to determine (i) the accuracy with which tree diameter and volume could be predicted for each site, (ii) how prediction accuracy varies with the number of measured trees in the training dataset, and (iii) using leave-one-site-out cross-validation, the extent to which tree dimensions can be predicted at an unmeasured site using a generalised model.
4. Discussion
The individual tree segmentation was highly accurate (F1 score = 0.96) with few false positives and false negatives. Detection accuracy depends on many factors, including tree species, forest complexity, point density, algorithm selection, and data quality [
49,
50,
51]. The use of UAV-LiDAR sensors in this study, which provided very dense and accurately located point clouds, contributed to the high detection rate. High accuracy was also achieved through the systematic establishment of the layout and strict planting spacing at each site, which allowed the average planting spacing to be used as the moving window size for the detection of tree peaks. Additionally, the even-aged nature of the stands allowed for the use of a single height threshold in the tree-peak detection algorithm. For mature sites, false positives were primarily regeneration, leaning, and windthrown trees that were caught on neighbouring canopies. Within young stands, other vegetation—such as regeneration and weeds—and non-vegetative objects, such as harvest residue also contributed to these false positives. False negatives in mature stands were mainly due to suppressed trees that did not reach the canopy layer, trees with broken tops, and leaning or windthrown trees obscured by canopy closure. In young stands, significantly shorter trees that were replanted after a survival assessment and plantings in skid areas were missed because they were below the height threshold.
The height model was very accurate. Although height measurements were collected from only a subset of the full dataset, there were over 7000 observations, representing 20% of the entire dataset, which provided a robust representation of the overall height distribution. Height was predicted from the zmax, with an
R2 of 0.99 and rRMSE of 8.33%, and the predictions were relatively unbiased. These results are consistent with previous research using UAV-LiDAR to predict individual tree heights, which have reported
R2 values ranging from 0.65 to 0.96 and rRMSEs from 3.70 to 9.03% [
52,
53,
54,
55,
56,
57,
58,
59,
60,
61]. The high value of
R2 found here is likely attributable to the wide range in height values in this study, as
R2 tends to increase when a variable covers a wide range [
62].
The predictions of the tree diameters using all of the measured trees exhibited a wide range in accuracy but were generally unbiased at all sites. The values of
R2 ranged from 0.33 to 0.90, while the rRMSE ranged from 6.2 to 13.4%. This entire range was fairly similar to the combined range for the previous UAV-LiDAR predictions of the tree diameters (
R2 in the range of 0.45–0.84; rRMSE in the range from 9.9 to 25.1%) [
54,
58,
59,
63,
64,
65,
66]. The similarity in these ranges highlights the importance of site and stand conditions on model accuracy and demonstrates the wide variability covered within this study.
Compared with the predictions for the diameter, the volume estimates had slightly higher mean
R2 values (0.746 vs. 0.713) but markedly lower accuracy, as indicated by the higher rRMSE values (19.57 vs. 9.70%). The higher rRMSE values for the volume predictions are consistent with previous research [
65,
67] and likely reflects error compounding, as the volume is determined from both the diameter and height. In contrast,
R2 values were comparable between the diameter and volume, as the relative range for the volumes was wider than that for the diameters (
Table 2), and
R2 generally increases with a greater range [
62]. There was wide variation in the accuracy of the volume predictions across the 20 sites, but the average accuracy (rRMSE = 19.57%) was consistent with previous estimates of LiDAR-predicted volumes from UAVs [
59,
65] and fixed-wing aircraft [
67].
The predictions of the diameter and volume made for each site were most sensitive to metrics describing the tree size and structure. The area and volume of the 3D convexhull represent the spatial extent and overall size of the tree crown. As larger trees have bigger crowns, these metrics are effective proxies for tree size. The number of points (n) and the total intensity of all points (itot), respectively, measure the total number of returns and the total reflected laser energy from the tree. Both of these features depend on the foliage density, crown size, and branching structure, all of which are linked to diameter and volume [
68].
The analyses showed a gradual interchange in the importance of different LiDAR-derived predictors of diameter and volume with increasing stand age. In younger stands (<6 years), which are more openly grown with minimal crown competition, LiDAR height-related metrics were typically the most important, reflecting the strong allometric relationship between tree height, diameter, and volume in early growth stages [
9,
69,
70,
71]. As stands develop further (approximately 6–11 years), simple LiDAR height metrics become less effective at capturing the variability in tree size, as crown closure occurs and trees start competing [
72]. During this phase, trees begin to differentiate more in crown development, and those with broader and deeper crowns tend to accumulate a greater stem diameter and volume than similarly tall but narrower-crowned individuals. Accordingly, 3D crown shape metrics—such as the convexhull area and volume—often emerge as among the most important predictors during this stage [
73]. In stands older than ca. 11 years, where crown closure, layering, and competition are more pronounced, the total intensity of LiDAR returns (itot) frequently becomes the dominant predictor. The variable itot effectively serves as a proxy for crown biomass and leaf area index (LAI) [
74,
75,
76], since it captures canopy density and structural complexity (e.g., foliage quantity and branch structure) in the upper canopy [
77]. Consequently, attributes related to leaf area and crown complexity, such as itot, play an increasingly influential role in predicting tree diameter and volume in later growth stages [
77].
The variations in the R2 for diameter and volume across sites were attributable to similar variables. Generally, the model’s R2 value was highest in younger, shorter, and smaller trees, which had low values for the LiDAR’s maximum height and standard deviation of the LiDAR’s height distribution. The model’s R2 values were also high at sites with low itot and lower values for the 2D and 3D convexhull area and volume. The model accuracy was highest in stands aged ≤ 6 years, which is the age at which the optimal balance between relatively high R2 values and the lowest rRMSE occur. The mean values for the R2 and rRMSE for these stands were, respectively, 0.801 and 9.2% for the diameter and 0.808 and 19.3% for the volume. The high accuracy at this age most likely results from better laser penetration of the canopy and the homogenous stand structure, prior to significant crown competition, when height percentiles and the canopy area are still closely linked to the tree diameter and volume. When compared with the other sites, the R2 values for the nine-year-old stand at site 12 were low for both the tree diameter and stem volume. This site was heavily infested with the woody weed gorse (Ulex europaeus) and had a denser understorey, which likely impeded LiDAR penetration. The reduced canopy penetration may have affected the distribution of UAV-LiDAR points within individual tree segments, contributing to the lower R2 values in the tree attribute predictions.
The number of measured trees could be significantly reduced without a major reduction in model accuracy. Reducing the number of training trees led to only slight increases in the rRMSEs and small decreases in the
R2 until the sample size dropped below 300, after which the model performance declined more noticeably. However, even with only 100 calibration trees, model accuracy was relatively similar to predictions using the complete dataset. Moreover, rMBE remained low across all sample sizes, indicating minimal bias even with reduced training data. This is an important finding because it suggests that accurate and unbiased predictions of diameter and volume can be achieved with as few as 100–300 trees, significantly reducing field data collection efforts without greatly compromising model performance. Given the high cost of tree measurement, there would be significant cost savings associated with these reductions, and this previously noted advantage of area-based LiDAR prediction [
21] is also a significant advantage for tree-level estimation using LiDAR.
The model predictions for the individual sites without tree measurements were generally less accurate and far more biased. We are unaware of any comparable research that investigates the accuracy of random forest using leave-one-site-out cross-validation, and this does highlight a potential limitation of this modelling approach. The R2 was reasonable for many sites and, for instance, averaged 0.827 for predictions of the diameter in stands that were ≤5 years. However, the bias was generally quite high across the age range. This suggests that while a reasonable ranking of diameter can be obtained for young sites, which may be useful for genetic evaluation of this important trait, calibration data would generally be required to obtain an accurate estimate for inventory purposes. The limitations of having insufficient data in the training set using leave-one-site-out cross-validation were most evident for predictions made at site 20. As this was the oldest site with the largest trees, the generalised model did not robustly predict tree dimensions at this site, using the mainly smaller tree dimensions within the training set. These findings emphasise the need for comprehensive coverage of tree sizes and site conditions within the training data to support the development of robust, generalisable models.
The presented results strongly suggest that individual tree delineation and characterisation using UAV-LiDAR would be feasible for inventory purposes particularly when calibration data are available. Traditional plot-based inventory is time consuming and costly and often has high error, as it only subsamples the stand [
21]. Stand inventories using LiDAR, which are typically acquired from fixed-wing aircraft, are commonly conducted using the area-based approach (ABA), which predicts key forest metrics at the plot level. However, the ABA is less cost-effective for small- to moderate-sized stands than UAV-acquired data [
23] as the ABA does not predict individual tree metrics. In contrast, predictions of individual tree dimensions using UAV-LiDAR can provide an almost complete distribution of stand dimensions [
8,
22], which can be used within growth models to determine tree dimensions, grade outturn, and stand value at harvest [
22]. A particularly useful application of the method described here by forestry companies may be in improving yield estimates for mid-rotation stands. Within New Zealand, yield tables for these stands are typically developed from silvicultural quality-control plot measurements, with growth projected forward to harvest age. The approach described in this study would overcome the variability inherent in plot sampling and provide an accurate and unbiased tree list to seed the growth model.
Further research is required to mitigate remaining impediments to widespread industry adoption of this method. Although UAV-LiDAR sensors are now a cost-effective option for inventory, data are often time-consuming to process and the conversion of raw point clouds into individual tree metrics is often a very involved multi-step process [
10]. Increases in computing power and the rapid development of sophisticated algorithms and pipelines to fit models and extract useful metrics are likely to largely overcome this issue. In tandem with these developments, ground truth and LiDAR data collection should be expanded and a wider range of models should be evaluated, including deep learning approaches [
78], to determine whether model generality can be improved. It would also be interesting to compare the model predictions of volume made here with those derived from photogrammetric point clouds. Further research should also evaluate whether LiDAR can be used to accurately identify malformed trees, and approaches should be developed for including these low-value trees into inventory systems, particularly at young ages before they are thinned out.
It is important to note that there are often restrictions that can hinder UAV mobilisation. These include line-of-sight constraints, regulations around UAV use, poor weather conditions, and insufficient flight endurance for the area of interest. Consequently, UAV-LiDAR acquisition will not be suitable for tree-level inventory in all situations. However, the results presented here represent a significant step forward in the establishment of the credibility of this methodology to improve inventory accuracy.