Optimizing UAV-LiDAR Point Density for Eucalyptus Height Estimation in Agroforestry

Abstract

The demand for forest materials necessitates advancements in forest management and inventory practices. We explore the integration of Unmanned Aerial Vehicles (UAVs) equipped with LiDAR sensors as a cost-effective alternative for precise forest monitoring. It evaluates the impact of varying point cloud densities on the accuracy of individual tree height estimation in Eucalyptus benthamii within Crop–Livestock–Forestry systems (15.9 ha and 357 individuals·ha⁻¹). We use a DJI M600 Pro UAV with a Velodyne 32c Ultra Puck LiDAR sensor at the Center for Technological Innovation in Agriculture (NITA) in Brazil. The resulting point clouds were processed to generate Digital Terrain Models and Canopy Height Models at densities ranging from 5 to 2000 points per square meter (pts·m⁻²). Statistical analyses, including Pearson correlation, root mean square error, and bias, were conducted to compare UAV-LiDAR-derived heights with field measurements. We found that reduced point densities, particularly around 100 pts·m⁻², maintained high accuracy in height estimation (RMSE = 17.129%, BIAS = −7.889%), with more than 90% in trees’ detection. UAV-LiDAR systems with optimized point cloud densities offer a viable solution for forest monitoring. 100 pts·m⁻² is an optimal density, promoting faster data collection, lower battery consumption, and reduced computational costs on trees’ height estimates.

1. Introduction

The increasing demand for wood materials has necessitated significant advancements in traditional forest management (rational use of forest resources) and inventory practices, driven by the need to acquire accurate and comprehensive forest information promptly and cost-effectively [1]. As global forestry operations strive for efficiency, accuracy, and cost-effectiveness, the integration of cutting-edge technologies becomes imperative [2,3]. Recent developments have propelled forest management towards the 4.0 frontier, leveraging sophisticated tools and advanced technology for comprehensive forest information acquisition. This includes phytosanitary monitoring and mortality inventory of the stand, necessitating the collection of highly detailed information remotely [2,3,4,5,6].

Integrating data from traditional forest inventory with the demands of geospatial mapping can pose additional challenges for precision forestry management. Consequently, employing technologies like remote sensing and Light Detection and Ranging (LiDAR), when precisely calibrated, can be more seamlessly integrated with data gathered by contemporary harvesting equipment. This integration provides a dataset that better supports spatiotemporal analysis and enhances the efficiency of large-scale (>1000 ha) plantation operations [7,8].

One of the significant advantages of these techniques is their spatial resolution, which allows for periodic measurement of the entire forest (census), unlike conventional techniques that rely on sampling units [9,10]. Among these tools, active remote sensing, particularly LiDAR, has been consistently applied to measure large forest areas (>1000 ha) [10,11]. This technology facilitates the measurement of various forest metrics (e.g., tree height, volume, basal area, and tree diameter) and the three-dimensional representation of both individual trees and entire groups [4,11,12,13,14,15].

Despite its versatility, ALS remains an expensive technology, limiting its application in large areas and continuous forest inventories that operate on a multi-temporal scale [16]. Thus, the use of unmanned aerial vehicles (UAVs) equipped with cameras offers a cost-effective alternative for periodic forest monitoring [17,18,19,20,21]. However, the UAV + camera system does not match the accuracy of ALS, as passive sensors fail to capture precise information from the ground and within the forest [22,23,24,25]. To address this issue and reduce data acquisition costs, LiDAR sensors have recently been integrated with UAVs [12,26,27].

The pioneering project that combined LiDAR sensors with UAVs was conducted by [28], which achieved point clouds with densities ranging from 100 to 1500 points·m⁻², utilized to identify individual trees. Since then, numerous studies have explored the potential of UAV-LiDAR, including the measurement of tree diameter at breast height [4]. Therefore, it is essential to assess the impact of point density on the accuracy of forest metrics, particularly for individual trees, to determine the density thresholds that can be reduced without compromising accuracy [26,29,30]. Also, few studies have systematically quantified the minimum viable LiDAR density in tropical agroforestry systems.

ALS systems seldom achieve high densities (>10 pts·m⁻²); hence, it is critical to evaluate the impact of point density on UAV-LiDAR systems, which typically result in point clouds with extremely high densities (>800 pts·m⁻²) [4]. This evaluation will aid in defining operational costs for data acquisition and ensuring accuracy in forest measurements, thereby promoting the development of Forest Management and Inventory 4.0 [4,10,31].

In this context, this study hypothesizes that there exists an optimal density of UAV-LiDAR point clouds that provide accuracy comparable to the maximum cloud density. Therefore, the objective of this work was to evaluate the performance of different densities of UAV-LiDAR point clouds in estimating the individual height of Eucalyptus benthamii Maiden et Cambage in Crop–Livestock–Forestry systems.

2. Materials and Methods

2.1. Study Area, Forest Inventory and UAV-LIDAR

The study was conducted at the Center for Technological Innovation in Agriculture (NITA) project, located at the Canguiri experimental station in the municipality of Pinhais, Paraná, Brazil (Figure 1), with central coordinates 25°23′30″ S and 49°07′30″ W. The region’s climate is classified as humid subtropical-Cfb according to the Köppen classification, with an average annual precipitation of 1550 mm, an average temperature of 17 °C, and an altitude of 935 m [32]. The predominant soils in the region are Cambisols, Oxisols, Organosols, and Gleisols [33]. NITA comprises 35 ha; however, in this study, only 15.79 ha were evaluated. The assessed area includes three management systems containing the forestry component: Livestock-Forest (PF), Crop-Forest (LF), and Crop-Livestock-Forest (LPF). Approximately 5.27 ha for each system.

In these three systems, the forest component comprises the species Eucalyptus benthamii Maiden et Cambage, planted along a contour line in 2012, with a simple row arrangement, and an initial spacing of 14 m between rows and 2 m between plants (357 individuals·ha⁻¹) [34]. It is worth noting that the stand was thinned after 44 months, with the intensity of 44% [34].

A comprehensive inventory of the area was conducted in September 2019, measuring the circumference at 1.30 m from the ground (CBH), which was converted to diameter at breast height (DBH) after dividing it by pi, and the total height (h) of all the trees (census) using a centimeter tape and a SWE Haglöf Vertex IV^® hypsometer (resolution of 0.1 m − error range = ±1–2 m). Additionally, the geographic position of the trees (1869 trees) was recorded using a BRA Garmin 62CSX GPS (accuracy ± 1 m).

The UAV-LiDAR data were collected in September 2019 using the GatorEye system, which consists of a USA DJI M600 Pro UAV, USA Velodyne 32c Ultra Puck LiDAR sensor, USA STIM300 inertial system, L1/L2 receiver, and SSD. This system features 32 lasers with a range of up to 220 m, and produces almost 1.2 million points per second [4,35]. The flight was conducted at a speed of 8 m/s, with 15 m between flight lines (90% overlap) and a flight height of 45 m, generating 120.66 million points with a maximum density of 2000 points.m⁻². Approximately 25 flights were operated.

2.2. Comparison Between Heights Derived from the Point Cloud and Field Measures

Pre-processing of the point cloud was performed using LASTOOLS software Version 1.8, where the cloud was merged and cropped to the study area using the LASMERGE and LASCLIP functions [36]. Subsequently, the data were processed in the R programming language [37] using the lidR [38] and rLiDAR [39] packages (Figure 2). We use an Avell notebook with 64 GB of RAM, an Intel Core i9 processor, 36 MB of cache, and an NVIDIA GeForce RTX 4080 graphics card with 8 GB (NVIDIA, Santa Clara, CA, USA).

The point cloud was homogenized into nine densities, commonly found in the literature: 2000, 1500, 1000, 500, 250, 100, 50, 25, and 5 points·m⁻² using the ‘lasfilterdecimate’ function by spatially uniform selection, ensuring the robustness of the database [38]. The points were classified and separated according to terrain and surface to generate the Digital Terrain Model (DTM) and normalize the cloud [38]. The functions ‘lasground’, ‘lasnormalize’, ‘grid_terrain’, and ‘grid_canopy’ were used.

Subsequently, the Digital Surface Model (DSM) and Canopy Height Model (CHM) were generated, applying a Gaussian-type smoothing filter with sigma 0.5 by the function ‘CHMsmoothing’ to remove possible noises and trees with a height below those measured in the field [40]. The normalized clouds and Smoothed CHMs were combined to derive the trees’ crown area, treetop and assess individual tree heights with the ‘FindTressCHM’ function. So, we use a spatial join function to join by position the heights measured in the field and derived by UAV-Lidar

The algorithm use the maximum height of the trees was then extracted based on the highest pixel present in the tree crown areas on the CHM and the normalized cloud, which also facilitated the counting of individuals [39].

The precision and accuracy of the tree heights derived from the UAV-LiDAR and those measured in the field were evaluated using statistical metrics: Pearson correlation coefficient (r)—Equation (1), root mean square error (RMSE)—Equation (2), bias—Equation (3), graphical analysis, and paired t-test [4,25,31,40]. The t-test was applied to test the following hypotheses: H0 = the heights derived by the UAV-LiDAR system are statistically similar to the heights measured in the field, H1 = the heights derived by the UAV-LiDAR system are statistically different from the heights measured in the field. To evaluate the performance of heights derived from the point cloud by height class, bias and RMSE metrics were applied.

$$r={\displaystyle \frac{{\sum }_{i=1}^n\left(yi-\overline{y}\right)\left(\widehat{yi}-\overline{\widehat{y}}\right)}{\sqrt{{\sum }_{i=1}^n{\left(yi-\overline{y}\right)}^2}{\sum }_{i=1}^n{\left(\widehat{yi}-\overline{\widehat{y}}\right)}^2}}$$

(1)

$$RMSE (\%)={\displaystyle \frac{100}{\overline{y}}}\sqrt{{\displaystyle \frac{{\sum }_{i=1}^n{(y_i-{\widehat{y}}_i)}^2}{n}}}$$

(2)

$$Bias (\%)={\displaystyle \frac{100}{\overline{y}}} {\displaystyle \frac{{\sum }_{i=1}^n(y_i-{\widehat{y}}_i)}{n}}$$

(3)

where $y_i$ is the field-based tree height of the ith tree; $\overline{y}$ is mean of field-based tree heights; ${\widehat{y}}_i$ is UAV-Lidar-derived tree height of the ith tree; and n is the sample size (1869 trees).

The Tukey test at a 5% significance level was used to determine whether there were differences between the heights derived at different point densities, enabling the definition of an ideal density. Additionally, to assess the influence of heights derived from point clouds, these measurements were incorporated into the individual tree volume formula, allowing for the inference of their impact on the forest’s volumetric stock, using a shape factor of 0.45 (v = DBH × h × ff) [41,42]. We obtained wood stock by the product of the trees’ number (n) and those mean individual volume (Vmi). The volume by area (V), we found by the Quotient of wood stock and the field area.

3. Results

3.1. Evaluation of Height Estimated on Forest Inventory and UAV-Lidar

The minimum height measured in the forest inventory was 11.10 m, while the maximum was 26.20 m. The heights derived by LiDAR presented values similar to the forest inventory (

Table 1. Descriptive statistics of individual heights across different point cloud densities in the NITA Project, Pinhais, Paraná, Brazil.

Scenarios		Height (m)			Statistical Metric
		Min	$\overline{\mathit{X}}$ ± SD	Max	SE	CV (%)
Field		11.10	19.92 ± 2.65	26.20	0.0630	17.75
Point cloud density (pts/m2)	2000	11.76	21.39 ± 3.66	28.06	0.1048	17.47
	1500	11.76	21.62 ± 3.82	27.97	0.1047	17.67
	1000	11.35	21.64 ± 3.99	28.30	0.1044	18.43
	500	11.18	21.65 ± 4.02	28.30	0.1041	18.54
	250	11.25	21.55 ± 4.15	28.38	0.1044	19.26
	100	11.10	21.44 ± 4.17	28.44	0.1041	19.44
	50	10.95	21.12 ± 4.40	28.38	0.1077	20.85
	25	10.91	20.90 ± 4.45	28.30	0.1088	21.30
	5	11.23	20.12 ± 4.16	27.92	0.0886	20.65

Min is the minimal height, $\overline{X}$ is the mean of heights, Max is the maximum height, SD is the standard deviation, SE’s the standard error, and CV is the coefficient of variation.

). The standard deviations (SD) of the derived heights ranged between 3.66 and 4.45, with standard errors (SE) between 0.088 and 0.017. The highest coefficient of variation (CV) was evident for the density of 25 pts·m⁻², while 2000 and 1500 pts·m⁻² obtained CV similar to the field data (Table 1). In general, all densities underestimated heights in the classes with class centers at 17.5 m and 22.5 m, while they overestimated the classes at 7.5 m, 12.5 m, and 27.5 m, except for 2000 pts·m⁻², which underestimated the frequency of individuals in the 7.5 m class.

The density with the lowest error in the total number of individuals was 25 pts·m⁻², followed by 1000 pts·m⁻². The densities 1500, 1000, 100 and 25 pts·m⁻² obtained a difference of less than 140 trees identified on the crop system. However, all the densities are overestimated on the 7.5, 12.5 m and 27.5 m classes, as well as underestimate in the 17.5 m, and 22.5 m classes (

Table 2. Number of Eucalyptus benthamii identified across different point cloud densities in the NITA Project, Pinhais, Paraná, Brazil.

Height’s Center Class (m)	Number of Individuals (n)
	Field	Point Cloud Density (pts·m−2)
		2000	1500	1000	500	250	100	50	25	5
7.5	13	4	18	20	20	20	23	21	32	60
12.5	87	208	206	214	207	211	225	248	281	438
17.5	788	395	401	404	391	386	400	392	394	646
22.5	974	623	632	642	642	670	691	703	708	1004
27.5	7	468	478	475	452	428	405	381	342	231
Total	1869	1698	1735	1755	1712	1715	1744	1745	1757	2379
Difference (n)		171	134	114	157	154	125	124	112	−510
Difference (%)		9.15	7.17	6.10	8.40	8.24	6.69	6.63	5.99	−27.29

). In general, the UAV-LiDAR was able to identify more than 90% of individuals from Eucalyptus benthamii.

It was found that 2000 pts·m⁻² represented the shape of the tree and the relief in detail, so that when reducing the density of points, information on the shape of the stem, crown and height of the tree is lost, as well as returns that favor the representation of the relief (Figure 3).

In addition to the representation of the tree, the density of points caused few differences in the digital terrain model (DTM) (Figure 4a), with discrepancies of up to 0.06% between the elevation of the highest density (2000 pts·m⁻²) and the others. Between 2000 and 1500 pts·m⁻², the differences remain close to 0%, while 1000, 500, 250, and 100 pts·m⁻² are close to 0.003%, which are located southwest of the study area. For 50 pts·m⁻², 25 pts·m⁻², and 5 pts·m⁻², the differences are around −0.007%, also located in the southwest of the project as the other densities.

In the digital canopy model, the density of 2000 pts·m⁻² showed heights of up to 27.5 m (Figure 4b). For densities down to 25 pts·m⁻², there is a predominance of differences on height close to the −11.9% error class, with some pixels showing greater differences distributed throughout the area. On another hand, 5 pts·m⁻² showed predominant differences of −52.2% relative to the density of 2000 pts·m⁻² (Figure 4b).

3.2. Comparison Between Heights Derived from the Point Cloud and Field Measures

The greatest correlation between the heights measured in the field inventory (FI) and those derived by the UAV-LiDAR was evident for the density of 2000 pts·m⁻²; as density decreased, so did the correlation. The lowest RMSE was 14.55% for the density of 2000 pts·m⁻². As density decreased, RMSE increased. All densities tended to overestimate heights (Table 1 and Figure 5). Despite this, at all densities, the derived and measured heights were statistically similar (

Table 3. Pearson correlation coefficient (r), root mean square error (RMSE), bias and t-test probability value for each point density of the Eucalyptus benthamii height derivation from the NITA project, Pinhais, Paraná, Brazil.

Point Cloud Density (pts·m−2)	r	RMSE (m)	RMSE (%)	Bias (m)	Bias (%)	t (p-Value)
2000	0.792	2.904	14.547	−1.675	−8.389	0.839
1500	0.787	2.922	14.660	−1.681	−8.433	0.874
1000	0.757	3.141	15.763	−1.713	−8.598	0.891
500	0.736	3.241	16.270	−1.737	−8.720	0.849
250	0.690	3.440	17.315	−1.673	−8.418	0.846
100	0.690	3.405	17.129	−1.568	−7.889	0.869
50	0.617	3.693	18.609	−1.277	−6.433	0.868
25	0.598	3.726	18.791	−1.073	−5.412	0.878
10	0.578	3.724	18.811	−0.484	−2.447	0.297
5	0.579	3.426	17.314	−0.334	−1.688	0.839

).

Overall, the residuals showed trends on the dispersion, concentrating between −3 and 3 deviations (Figure 5a), particularly at the lowest densities (50 pts·m⁻², 25 pts·m⁻², and 5 pts·m⁻²). The dispersion of heights denotes the accuracy of densities 2000 pts·m⁻², 1500 pts·m⁻², and 1000 pts·m⁻², as well as biased estimates for densities 50 pts·m⁻², 25 pts·m⁻², and 5 pts·m⁻², tending to overestimate heights (Figure 5b).

In general, all densities exhibited higher RMSE and Bias for the height classes 7.5 m and 27.5 m, while in the central classes these statistical measures presented moderate values. The highest densities (2000 pts·m⁻², 1500 pts·m⁻², and 1000 pts·m⁻²) obtained lower values for the RMSE and Bias metrics in the height classes (Figure 6).

Using the Tukey test at a 95% probability, densities from 2000 pts·m⁻² to 100 pts·m⁻² resulted in similar heights, which differed from the others. The densities of 50 pts·m⁻² and 25 pts·m⁻² are also similar, while 5 pts·m⁻² differs from the others (Figure 7). The derived heights, despite overestimating the field heights, have a similar distribution, with a median close to the center and a negatively skewed distribution.

The heights derived by the UAV-LiDAR tended to overestimate the average mean individual volume (Vmi), except for the density of 5 pts·m⁻², which underestimated the Vmi. The differences varied between 1.21% for 5 pts·m⁻² and −9.38% for 2000 pts·m⁻². Despite this, due to the number of individuals accounted for by each point cloud density, there is low variation between the volume stocks calculated with the height derived from the UAV-LiDAR and the field data, with the largest divergence (−25.75%) obtained by the density of 5 pts·m⁻², which overestimated the number of individuals in the area. Stocks with densities of 2000 pts·m⁻² and 100 pts·m⁻² are the closest to the stock quantified with field data (

Table 4. Mean individual volume (Vmi), volume of the area (V), total number of trees (n), and height inference of Eucalyptus benthamii in the NITA project, Pinhais, Paraná, Brazil.

Scenario		Stands Variables				Difference (%)
		Vmi (m3·tree−1)	n	V (m3·ha−1)	Stock (m3)	Vmi	Stock
Field		0.6963	1869	82.42	1301.38	-	-
Point cloud density (pts·m−2)	2000	0.7616	1698	81.90	1293.24	−9.38	0.63
	1500	0.7595	1735	83.45	1317.70	−9.07	−1.25
	1000	0.7546	1755	83.87	1324.32	−8.37	−1.76
	500	0.7551	1712	81.87	1292.72	−8.44	0.67
	250	0.7517	1715	81.64	1289.16	−7.96	0.94
	100	0.7452	1744	82.31	1299.62	−7.02	0.14
	50	0.7388	1745	81.64	1289.16	−6.10	0.94
	25	0.7228	1757	80.43	1269.99	−3.81	2.41
	5	0.6879	2379	103.64	1636.48	1.21	−25.75

).

4. Discussion

4.1. Forest Inventory and UAV-Lidar

The reduction in point density caused greater divergences in the representation of the tree profile, especially at the top. However, despite the smaller number of points, there are no major differences in the elevation model. As a result, the reduction in point density directly influences the digital canopy model, causing differences in the height derived for each density [13,15,43].

There were no changes in the point cloud that would distort the relief classification. Therefore, at all densities, the Digital Terrain Model (DTM) is of high quality, with a low incidence of errors, which favors reducing the density of the point cloud without compromising the precision and accuracy of the derived heights [30]. Additionally, the cloud with the lowest density in our study (5 pts·m⁻²) is equivalent to the maximum density obtained in the literature, with densities ranging from 5.7 pts·m⁻² to 37.5 pts·m⁻² [30,44,45]. Although this density is considered low for the conditions of this study, it was effective for identifying forest species and measuring height and biomass in other research [28,44,45,46,47].

The DTMs derived from various point densities were consistently of high quality, capturing essential topographical features accurately even at reduced densities. While the reduction in point density affected the digital canopy model, the overall accuracy of height estimation remained high when compared with other studies [4,8,25,30,31].

4.2. Comparison Between Heights Derived from the Point Cloud and Field Measures

This method of data acquisition has obtained accurate heights with a strong correlation up to densities of 100 pts·m⁻² to 2000 pts·m⁻² and moderate for all densities lower than 100 pts·m⁻². In studies with lower point densities (<100 pts·m⁻²), correlations remain between moderate and strong [25,31,48], which explains why low densities present a correlation between 0.5 and 0.7.

The height derivations obtained in this study presented RMSE and bias similar to other studies, with RMSE values between 14% and 18% and bias ranging from −1% to 8%, indicating that this system is accurate for height estimates [26,31,49]. The greatest divergences in the height classes 7.5 m and 27.5 m may be related to the fact that height measurements in the field are subject to equipment and operator bias, which affects the accuracy in determining this variable [50,51,52,53]. Additionally, the digital devices for measuring height, such as the Vertex IV (error range = ±1–2 m), have lower performance for measuring taller trees, which may have occurred under the conditions of this study, as minimal oscillations in the view of the tree lead to large errors in estimating total height [54,55,56]. These errors could potentially be minimized if direct measurements of some trees were conducted to adjust a hypsometric equation [55,56].

The importance of height as a variable for forest management cannot be overstated, as it is used to determine the productive capacity of the site, model volume, biomass, carbon, and forest production [57,58,59,60]. However, this variable is difficult to measure and subject to various errors [50]. Therefore, the results obtained by the UAV-LiDAR sensor used in this study proved to be satisfactory for collecting tree heights more quickly and with less influence of errors [4,31,58,61,62,63].

Accurate determination of the number of individuals and their heights is essential for quantifying the forest timber volume. At densities ranging from 2000 pts·m⁻² to 50 pts·m⁻², the precise number of individuals resulted in a little difference in total wood volume compared to field data, indicating the accuracy of data derived from the UAV-LiDAR system. Higher point densities can lead to more accurate measurements [64,65]. However, in this study, differences were found up to 100 pts·m⁻², significantly reducing the number of flights and the machine operational costs during data processing.

Such a reduction is viable, as it is possible to plan the flight to acquire data with lower density, favoring an increase in speed and altitude of the sensor, consequently reducing battery consumption [12]. This procedure allows for faster processing and lower computational costs [66,67]. Consequently, reducing the density of points lowers the costs inherent to acquiring UAV-LiDAR data, as increased speed and altitude improve system autonomy, enabling flights over larger areas within a similar time interval required to generate a high-density point cloud over a smaller area. However, supplementary experiments are required to validate this in the field, with a flight plan based on lower densities.

UAV-LiDAR is still a new system and faces some obstacles [49], such as the need for an extremely stable platform for collecting LiDAR data (flight stability and wind resistance, navigation accuracy and quality (GNSS + IMU), and operational safety—return to home), where gusts of wind can create different point densities in the same location [68]. Therefore, homogenization of the point cloud is an essential processing step.

However, the UAV DJI M600 Pro, as a multi-rotor platform, reduces vibration during the flight plan, favoring the acquisition of LiDAR data [16]. Despite covering many points in the forest area, the high overlap can produce some noise, which could originate from tree branches and leaves or possibly from equipment vibrations and strong winds, generating UAV displacement and consequently altering pulse collection direction. Therefore, the high density of points and data acquisition costs make UAV-LiDAR a viable alternative for measuring forest variables (e.g., height, DBH, basal area, crown’s diameter, and area) in small areas [4,16,29,31,62,69,70,71,72,73].

Furthermore, determining the optimal point cloud density is essential for obtaining other important metrics, such as crown diameter, crown height, and crown area [72], which can be used to estimate DBH [4], basal area [70], tree volume [26,74], the forest’s wood stock, and biomass [71,74,75]. But in future studies, we recommend assessing other forest types, including native forests and pure stands, to check if this behavior is the same in denser forests with different species and more complex canopy and terrain structures [31].

5. Conclusions

This study confirms the hypothesis that an optimal density of UAV-LiDAR point clouds can provide accuracy comparable to the maximum cloud density. Lower point densities were found to yield precise measurements for estimating the height of Eucalyptus benthamii in Crop–Livestock–Forestry systems, particularly around 100 pts·m⁻², balancing accuracy with minimized data collection efforts.

The UAV-LiDAR-derived heights showed a strong correlation with field measurements at higher densities and a moderate to strong correlation even at lower densities, validating the reliability of UAV-LiDAR systems for height estimation. Accurate determination of individual tree heights is crucial for quantifying forest variables, such as volume and aboveground biomass.

By demonstrating the feasibility of using reduced point densities without compromising the tree’s height estimates, this study lays a solid foundation for optimizing UAV-LiDAR usage in forestry. The density of 100 pts·m⁻² is an optimal and cost-efficient density, promoting practical benefits such as faster data collection, lower battery consumption, and reduced computational costs. Thus, UAV-LiDAR emerges as a viable and cost-effective tool for forest monitoring and management, supporting efficient and accurate forest inventory practices.