Tree height is one of the key dendrometric parameters for indirectly estimating the timber volume or aboveground biomass of a forest. Field measurement is time-consuming and labor-intensive, while unmanned aerial vehicle (UAV)-borne LiDAR is a more efficient tool for acquiring tree heights of large-area forests. Although individual tree heights extracted from point cloud data are of high accuracy, they are still affected by some weather and environment factors. In this study, taking a planted M. glyptostroboides (Metasequoia glyptostroboides Hu & W.C. Cheng) stand as the study object, we preliminarily assessed the effects of various illumination conditions (solar altitude angle and cloud cover) on tree height extraction using UAV LiDAR. The eight point clouds of the target stand were scanned at four time points (sunrise, noon, sunset, and night) in two consecutive days (sunny and overcast), respectively. The point clouds were first classified into ground points and aboveground vegetation points, which accordingly produced digital elevation model (DEM) and digital surface model (DSM). Then, the canopy height model (CHM) was obtained by subtracting DEM from DSM. Subsequently, individual trees were segmented based on the seed points identified by local maxima filtering. Finally, the individual tree heights of sample trees were separately extracted and assessed against the in situ measured values. As results, the R2 and RMSEs of tree heights obtained in the overcast daytime were commonly better than those in the sunny daytime; the R2 and RMSEs at night were superior among all time points, while those at noon were poorest. These indicated that the accuracy of individual tree height extraction had an inverse correlation with the intensity of illumination. To obtain more accurate tree heights for forestry applications, it is best to acquire point cloud data using UAV LiDAR at night, or at least not at noon when the illumination is generally strongest.
Keywords: individual tree height; UAV LiDAR; point cloud; illumination condition; point segmentation; solar altitude angle
Tree height is one of the key dendrometric parameters for indirectly estimating the timber volume or aboveground biomass of a forest [[
As an active remote sensing technology, LiDAR emits laser pulses and receives their returns reflected by object surfaces [[
In recent years, with the development of unmanned aerial vehicles (UAVs) and lightweight sensors, UAV-borne LiDAR has gradually matured [[
Although individual tree heights acquired by UAV LiDAR are of high accuracy [[
In this paper, we conduct a pilot study on the effects of UAV-borne LiDAR for individual tree height extraction under various illumination conditions, including solar altitude angle and cloud cover. Eight point clouds of the target stand were scanned at four time points (sunrise, noon, sunset, and night) over two consecutive days (sunny and overcast). Our specific objectives were as follows: (
The study site is located at the lower reach of the Ma'anxi Creek in the northwestern part of Beibei District (Figure 1a), Chongqing Municipality, China. The creek begins at the Longtanzi Reservoir and ends at the Jialing River, with a length of approximately 4 km. This region possesses a subtropical monsoon climate, which is characterized by low wind, high humidity, and abundant rainfall in the hot season. Many tree species such as Camphora officinarum, Neosinocalamus affinis, and M. glyptostroboides (Metasequoia glyptostroboides Hu & W.C. Cheng) grow along the Ma'anxi creek. The total area of the study site (Figure 1c) was approximately 0.26 ha with an average northwest slope of 16.47°. M. glyptostroboides is the predominant species of the planted stand with a relatively uniform spacing and tree age, but there is almost no regular management. The trunk of M. glyptostroboides is tall and straight, with a steeple-shaped crown when viewed from the side and a rounded crown when viewed from the top [[
As shown in Figure 2a, the UAV-borne LiDAR system used in the point scanning was a Feima D2000 (FEIMA Robotics, Shenzhen, China) equipped with a Feima D-LiDAR2000 (FEIMA Robotics, Shenzhen, China). The Feima D2000 is a four-rotor drone with a payload of up to 700 g, and is positioned using an onboard RTK (real-time kinematic) GNSS (Global Navigation Satellite System). The D-LiDAR2000 is a triple-echo laser scanner capable of detecting three returns from a single laser pulse, including those from the ground, tree trunks, and vegetation canopies. The detailed technical parameters of this LiDAR are listed in Table 1.
Under natural conditions, the variation in light intensity during a day is generally caused by the different altitude angles of the sun at different times. Additionally, at the same solar altitude angle, different cloud covers have different blocking effects on sunlight, which can also cause changes in illumination intensity. In this study, to assess the effects of different illumination conditions (solar altitude angle and cloud cover) on tree height extraction, there were eight UAV-based scanning operations for the study site carried out at four time points (sunrise, noon, sunset, and night) over two consecutive days (sunny and overcast). During the two days, the other weather conditions in our experiments were relatively consistent, including no rainfall, low wind speed (no tree shaking), insensible change in humidity, and so forth. The specific information of the eight UAV-borne LiDAR scanning measurements is listed in Table 2.
As indicated in Figure 1b, a cross-flight route was adopted in the eight UAV LiDAR scanning operations. To ensure consistent and uniform laser scanning across the entire study site, the side overlap was set as 60% and the flight speed was 5 m/s. Furthermore, the relative flight altitude was set as 90 m, guaranteeing a point density of approximately 2000 points/m
Subsequently, the raw point cloud data underwent pipeline processing, including utilizing Inertial Explorer 8.70 (https://novatel.com/products/, accessed on 2 January 2024) for waypoint post-processing and the Data Processing Toolbox of Feima UAVmanager for point cloud resolving (https://
In order to facilitate the identification and localization of the sample trees, a DJI Phantom 4 pro (DJ-Innovations, Shenzhen, China) (Figure 2b) equipped with a Sony Exmor R CMOS sensor (Sony Group, Beijing, China) was used to capture RGB images of the study site on 10 September 2023. The flight altitude for the drone was set as 90 m, guaranteeing a ground sampling distance (GSD) of approximately 0.05 m. The heading overlap and lateral overlap were set to 90% and 70%, respectively. Based on the overlapping RGB images, the orthoimage (Figure 2e) of the study site was produced using the Pix4Dmapper Version 4.5.6 (https://
To improve the accuracy of in situ measured tree heights, only those trees with fewer occlusions and easy measurement in the study site were selected as samples. As a result, there were 32 M. glyptostroboides purposely selected as sample trees. As indicated in Figure 2e, the spatial locations and number identifiers of those sample trees were separately labeled. The height of each sample tree was measured using a RTS-882 total station (Figure 2c) from multiple directions or angles [[
To assess the illumination effects, the individual heights of sample trees were separately extracted from the eight point clouds scanned by UAV-borne LiDAR. The specific procedures are as follows. Firstly, the ground and aboveground tree points were separated via classifying the raw point clouds. Secondly, the digital elevation model (DEM) and digital surface model (DSM) were produced by rasterizing the ground points and non-ground points, respectively. Thirdly, the canopy height model (CHM) was acquired by subtracting the DEM from the corresponding DSM. Fourthly, the seed points representing the highest points of individual trees were generated by using Gaussian filtering. Based on the seed points, the CHM was segmented into individual trees. Fifthly, the crown boundaries of individual trees were extracted, and the tree heights were determined as the highest points falling into the boundaries. Finally, the accuracies of the extracted individual heights of sample trees under different illumination conditions were assessed against the measured values in terms of various metrics. The detailed workflow of this study is illustrated in Figure 3.
Accurate classification of ground points and aboveground forest points from the raw points clouds is the prerequisite for extracting individual heights of M. glyptostroboides. Before classification, the point denoising should be first performed. The points outside the target scanning area have significant errors and need to be removed in advance. Subsequently, the noise points that are caused by the device or external factors are filtered out from the remaining points. In this study, the improved progressive TIN densification algorithm [[
The CHM is the surface model which reflects the distance between the vegetation surface and the ground [[
(
The DSM generally includes the elevation information of buildings, trees, and other objects above the ground, and is produced from the classified vegetation points. The DEM reflects the true undulations of the ground, and is produced from the classified ground points. Since both the DSM and DEM are in raster format, their productions involve the step of rasterization. As the cell size impacts the extraction of the crown boundaries and seed points for individual tree segmentation, the rasterized DSM and DEM should have the same spatial resolution, which were determined by repeated experiments [[
The k-means clustering algorithm is used to segment individual trees from the point cloud [[
The tree height commonly refers to the distance from the top of the tree crown to the ground. The crown boundaries of segmented individual trees are the convex hull polygons generated by projecting the belonging points onto the ground. Then, the individual tree heights are determined as the highest points falling into the boundaries.
The paired samples t-test was employed to assess the statistical significance of differences between every two groups of field-measured and eight LiDAR-derived tree heights. This test is particularly suitable for datasets involving measurements taken under different conditions [[
The individual tree heights of sample trees extracted from point clouds acquired under varying illumination conditions were subjected to comparison with in situ measured values. The metrics, including R
(
(
(
(
where
The eight point clouds scanned by UAV LiDAR were separately classified into the ground and aboveground vegetation points using the software LiDAR360 V6.0 (https://
Using the classified point clouds in the sunny night as an example, the process of generating the CHM of the study site is illustrated in Figure 5. Several candidate spatial resolutions (0.1 m, 0.3 m, and 0.5 m) were sequentially tested, and 0.3 m was determined as the optimum one. Then, the ground points were interpolated and rasterized into an DEM image with the optimum spatial resolution (Figure 5a). Meanwhile, the DSM in Figure 5b was generated by interpolating and rasterizing aboveground vegetation points. It shared the same spatial resolution with the DEM. Based on the DEM and DSM, the resulting CHM is indicated in Figure 5c. Notably, the maximum height recorded for an individual tree in the study site was 21.871 m. With greater growth space and more abundant sunlight, the trees located on the edges grew much lusher and had higher heights.
Through empiric experiments, the optimal parameters of Gaussian filtering to those CHMs were determined to be a sigma value of 11 and a convolution kernel of 9. As shown in Figure 6, more than 200 seed points were initially generated from the CHM acquired during the sunny night. Subsequently, through on-site verification and manual editing, the number of retained seed points in the CHM was reduced to 195. This process was crucial for eliminating seed points incorrectly extracted due to factors such as shading or fluctuating terrain. Thus, each retained seed point corresponded to the top of an individual tree that truly existed.
Based on the retained seed points, the individual trees in the sunny night were segmented. As observed in Figure 7a, the points belonging to the individual trees were rendered in various colors for the purpose of easy distinguishment. Then, as indicated in Figure 7b, the corresponding crown boundaries of individual trees were obtained.
The crown boundaries were subsequently superimposed on the CHM to extract individual tree heights. In the same way, the heights of 195 trees under different illumination conditions were extracted and are shown in Figure 8, exhibiting similar height range and spatial distribution on the whole. The statistics, including average tree height and standard deviation of tree heights under different illumination conditions, are listed in Table 4. The average tree heights were all around 13 m, with internal trees typically below 13.5 m in height and edge trees above 13.5 m. Several trees as high as 17 m were also located in the edge area. The standard deviations of tree heights were all around 2.50 m.
The results of the paired-samples t-test between every two groups of field-measured and eight LiDAR-derived tree heights are shown in Figure 9. The number 1 indicates the rejection of the null hypothesis, meaning that there was a significant difference (p < 0.05) existing between the two group of tree heights, while 0 was the opposite (p > 0.05). In Figure 9, it can be seen that there were significant differences in most pairs of tree height groups excluding self-comparisons, including field measurement–sunny night, field measurement–overcast night, sunny sunrise–overcast sunrise, sunny sunset–overcast sunset, and so forth. For the sunny day condition, there were more significant differences in pair groups that included time points of noon or night than those that included sunrise or sunset. For the overcast day, more non-significant differences were achieved due to slight differences in illumination intensity at different time points except overcast night. For the comparison of the two days, more significant differences in pair groups were attained. Also, there were some pair groups that did not show significant differences, such as field measurement–sunny sunset, field measurement–sunny sunset, sunny sunrise–sunny sunset, sunny noon–overcast noon, and sunny night–overcast night. The reason was possibly that the illumination intensities at the two time points were very close. For instance, due to the lack of solar illumination on both sunny and overcast nights, the impacts of illuminations on tree height extraction were minimized, and the accuracies were very close.
By matching the tree positions in Figure 2e and Figure 7b, it was confirmed that all 32 sample trees were correctly segmented from the point cloud acquired during the sunny night. Similarly, in the point clouds obtained at the other seven time points, all sample trees were properly segmented as well. Thus, the accuracies of the extracted tree heights of those sample trees (Figure 8) under varying illumination conditions were separately assessed against the corresponding in situ measured values listed in Table 3. As shown in Figure 10, the R
The slight differences in these accuracy metrics under different illumination conditions can be more clearly observed in Figure 11. As shown in Figure 11a, the R
Although individual tree heights extracted from point cloud data were of high accuracy, they were still affected by illumination conditions. To obtain more accurate tree heights for forestry applications, it is best to acquire point clouds using UAV-borne LiDAR at night. If such a task has to be carried out during the day, it is better not to do so at noon when the illumination is generally strongest.
Solar radiation has negative impacts on the accuracies of individual tree height extraction by forming solar background noise (SBN) which restricts the signal-to-noise ratio (SNR) of the detector of LiDAR. Commonly, the manufacturers of LiDARs install a narrow band-pass filter [[
Although the laser pulse/returns and the sunlight are two independent light sources, direct interference between them does exist during propagation [[
In addition, solar radiation might indirectly affect the intensity of laser pulse/returns by influencing environmental factors such as air humidity. Generally, the moisture in the air has a certain absorption effect on laser pulses in the near-infrared band [[
In a word, the SBN caused by solar radiation was the dominant factor affecting the height measurement of individual trees using UAV LiDAR in our experiments. The direct interference with laser pulse/returns made by sunlight during transmission was negligible. The impact of air humidity depends on the base value and the magnitude of the variation induced by solar radiation. In our experiments, the air humidity was relatively low and consistent and not the major factor causing the differences in tree height measurement.
Besides illumination conditions, there are still other factors affecting the ranging accuracy of a LiDAR mounted on the UAV, including the technical parameters of the sensor, settings in flight operation, characteristics of forest stand, and so forth.
As one of the most important technical parameters of a LiDAR, the laser return number determines how many returns a laser scanner can detect from a single laser pulse. Each laser return is reflected by different object surfaces in a stand, including the top of the tree crown, branches in the middle layer, understory vegetation, and the ground. Therefore, the number of laser returns is theoretically proportional to the complexity of detectable forest structures [[
Flight altitude is one of the most important settings for UAV flight operation. With the increment of flight altitude, the lengths of propagation paths of a laser pulse and their returns accordingly increase, implying an aggravated attenuation by the atmosphere. At the same time, it also adds the negative impact of solar radiation on laser pulses and their returns, ultimately affecting the ranging accuracy. As listed in Table 1, the ranging accuracy of ±2 cm is only ensured at the flight altitude of 50 m. In our study, to avoid colliding with a nearby building that was 80 m high, the flight altitude of the UAV was set to 90 m. This also reduced the ranging accuracy to a certain extent. For a large forest distributed in a mountainous area, it is beneficial to maintain relative flight altitude via adopting a terrain-following mode, thereby ensuring consistency in ranging accuracy [[
The characteristics of the target forest stand also have a significant impact on the accuracy of individual tree height extraction [[
In fact, the factors affecting the range accuracy of UAV LiDAR are more than those studied and discussed in this paper. In a natural environment, it is exceedingly challenging to achieve complete consistency in experimental conditions. While the effects of a certain factor on individual tree height extraction are studied, the relative consistency of other factors should be maintained.
Although we preliminarily evaluated the effects of various illumination conditions (solar altitude angle and cloud cover) on tree height extraction using UAV LiDAR, there were still some limitations in our study. For instance, without measuring illumination intensities at those time points, only the qualitative relationship between light intensity and accuracy of tree height extraction was obtained. For the sake of safety, the flight altitude of the UAV exceeded the optimal ranging accuracy scope of the LiDAR. The insufficient number and distribution of field-measured sample trees might result in a lack of statistical independence. The magnitudes of the negative effects of the SBN on ranging accuracy were not determined.
Nevertheless, some measures could be taken to further improve the reliability of the studies on this topic in the future, such as choosing a larger area of planted forest as the study object, and ensuring no high mountains or buildings are present in the surrounding area. In this way, enough sample trees can be selected to measure heights in situ for verification and statistical analysis, and the UAV can conduct point scanning at the altitude within the optimal ranging scope of the LiDAR. The photometer is used to measure the illumination intensity while LiDAR scanning so that the quantitative relationship between illumination intensity and the accuracy of extracted tree heights can be established.
Although tree heights extracted from point cloud data are of high accuracy, they are still affected by weather and environmental factors. In this study, taking a planted coniferous stand of M. glyptostroboides as an example, we preliminarily assessed the effects of various illumination conditions (solar altitude angle and cloud cover) on individual tree height extraction using UAV-borne LiDAR. The major conclusions we drew were as follows:
- Individual tree heights extracted from the eight point clouds scanned by UAV LiDAR under various illumination conditions were of high accuracy. There were statistically significant differences in most pairs of tree height groups. The R
2 s were all above 0.95; the Biases had absolute values ranging from 0.05 to 0.3 m; the RMSEs were all less than 0.6 m; and the rRMSEs were all below 4%. - The accuracy of individual tree height extraction had an inverse correlation with the intensity of illumination. The R
2 and RMSEs of tree heights obtained in the overcast day were commonly better than those in the sunny day; the R2 and RMSEs at night were superior among all time points, while those at noon were poorest. - To obtain more accurate tree heights for forestry applications, it is best to acquire point cloud using UAV LiDAR at night. If such a task has to be carried out during the day, it is better not to do so at noon.
Nevertheless, our study was just a first attempt in this topic, and the above conclusions were achieved under the conditions of our simple experiments, including tree species, site conditions, LiDAR sensor, point processing procedure, and so forth. There are still other factors affecting the accuracy of individual tree height extraction, such as the interferences of understory vegetation, the different growth patterns of various tree species, and the differences in the anti-interfering functionality in various laser devices. Moreover, only qualitative divisions were made for lighting conditions in terms of solar altitude angles and cloud cover in this study, and the quantitative relationship between lighting intensity and the accuracy of tree height extraction is required for further study. In practical applications, one could choose the appropriate illumination conditions for LiDAR data collection depending on the purpose and accuracy requirements. In future studies, more rigorous and diverse experiments for tree height extraction using UAV-borne LiDAR need to be carried out. The further reduction in the potential errors of field-measured tree heights will be beneficial for the statistical significance tests. The effects of other factors such as different laser return numbers, various flight altitudes, and different tree species and stand structures on the accuracy of individual tree height extraction should be further evaluated.
Graph: Figure 1 (a) The location of the study site; (b) the cross-flight trajectory of UAV-borne LiDAR; (c) the point cloud rendered in elevation (top view).
Graph: Figure 2 (a) Feima D2000 equipped with D-LiDAR2000; (b) DJI Phantom 4 Pro with FC6310; (c) RTS-882 total station; (d) Huace RTK GNSS; (e) locations of sample trees for in situ measurement.
Graph: Figure 3 The workflow of this study.
Graph: Figure 4 Classified point cloud—sunny night data as an example. (a) Ground points (side view); (b) aboveground vegetation points (side view).
Graph: Figure 5 The CHM created from point cloud—sunny night data as an example. (a) DEM; (b) DSM; (c) CHM.
Graph: Figure 6 The resulting seed points from CHM—sunny night data as an example.
Graph: Figure 7 Segmented individual trees—sunny night data as an example. (a) Point cloud of individual trees rendered in colors; (b) crown boundaries of individual trees.
Graph: Figure 8 Extracted individual tree heights under different illumination conditions. (a) Sunny sunrise; (b) sunny noon; (c) sunny sunset; (d) sunny night; (e) overcast sunrise; (f) overcast noon; (g) overcast sunset; (h) overcast night.
Graph: Figure 9 The results of the statistical significance test between every two groups of field-measured and eight LiDAR-derived tree heights using paired samples t-test; 1 indicates that the null hypothesis is rejected, while 0 is the opposite.
Graph: Figure 10 Accuracy assessment of extracted individual tree heights against in situ measured values under different illumination conditions. (a) Sunny sunrise; (b) sunny noon; (c) sunny sunset; (d) Sunny night; (e) overcast sunrise; (f) overcast noon; (g) overcast sunset; (h) overcast night.
Graph: Figure 11 Metric curves for extracted individual tree heights at four time points in two days. (a) R2; (b) Bias; (c) RMSE; (d) rRMSE.
Table 1 Technical parameters of Feima D-LiDAR2000.
Parameter Value/Description Platform Feima D-2000 Laser class class 1 Ranging mode TOF (Time of Flight) Wavelength 905 nm Return number 3 returns Return intensity 8 bit Ranging accuracy ±2 cm (50 m) Point frequency 240 kpts/s Horizontal FOV (Field of View) 70.4° Vertical FOV 4.5°/77.2°
Table 2 Specific information of the eight UAV LiDAR scanning measurements.
Description Date 10 September 2023 11 September 2023 Atmospheric condition Sunny Overcast Time points Sunrise (6:00); noon (12:00); sunset (18:00); night (21:00)
Table 3 The in situ measured heights of sample trees.
Tree ID Height (m) Tree ID Height (m) Tree ID Height (m) 1 14.869 12 18.240 23 17.004 2 14.647 13 17.766 24 17.540 3 14.654 14 13.750 25 12.164 4 16.556 15 16.463 26 16.305 5 18.961 16 15.753 27 17.035 6 11.563 17 14.839 28 12.728 7 14.295 18 15.701 29 11.981 8 10.627 19 16.753 30 12.108 9 14.619 20 19.739 31 8.652 10 8.793 21 18.522 32 14.503 11 17.476 22 15.267
Table 4 The statistics of extracted tree heights under various illumination conditions.
Atmospheric Condition Time Point Average Tree Height (m) Standard Deviation of Tree Height (m) Sunny sunrise 12.97 2.49 noon 12.80 2.56 sunset 12.93 2.44 night 13.09 2.56 Overcast sunrise 12.99 2.47 noon 12.83 2.58 sunset 13.04 2.49 night 13.10 2.53
Conceptualization, J.L.; methodology, T.L. and J.L.; validation, T.L. and W.W.; formal analysis, J.L. and T.L.; investigation, T.L., W.W. and R.J.; data curation, R.J. and W.W.; writing—original draft preparation, T.L.; writing—review and editing, J.L.; visualization, T.L.; supervision, J.L.; project administration, J.L.; funding acquisition, J.L. and R.J. All authors have read and agreed to the published version of the manuscript.
The data presented in this study are available on request from the corresponding author.
The authors declare no conflicts of interest.
By Tianxi Li; Jiayuan Lin; Wenjian Wu and Rui Jiang
Reported by Author; Author; Author; Author