1. Introduction
Global temperature increases caused by increased carbon emissions [
1,
2], forest carbon stock, and sequestration capacity have received increasing attention [
3,
4,
5]. Previous studies demonstrated that accurate measurement of diameter at breast height (DBH) is beneficial to forest biomass systems, thereby reducing the uncertainty in carbon stock estimation [
6,
7]. Therefore, it is essential to accurately measure the DBH. Although traditional DBH determination methods, such as field measurement, can accurately retrieve DBH [
8], they are limited in practical applications since they are labor-intensive and time-consuming.
The light detection and ranging (LiDAR) technology has been actively studied for DBH measurement. Especially, terrestrial laser scanning (TLS) has been widely used for trunk point extraction and DBH measurement [
9]. Ye et al. [
10] proposed a robust least squares ellipse fitting method under the elliptic assumption for DBH estimation. Liu et al. [
11] used octree segmentation, connected component labeling, and random Hough transform (RHT) to identify trunks and estimate DBH from TLS data. Jan et al. [
12] used the density-based spatial clustering of applications with noise (DBSCAN) and random sample consistency (RANSAC) for TLS-based DBH estimation. Although the TLS-based methods showed their unique advantage in measuring DBH, the TLS also has several limitations. First, the acquisition of TLS point clouds in forestry requires multi-station scanning and registration of multi-station data, which is time-consuming and labor-intensive to collect and preprocess large-scale data. Additionally, TLS has a spatial limitation to application because it needs to be manually carried to reach the area of interest, often requires challenging field access, and often requires a field team [
13]. Second, the high density of TLS point clouds and alignment of multi-station data lead to data redundancy and also require high hardware requirements.
Compared with TLS, airborne laser scanning (ALS) and unmanned aerial vehicle (UAV) LiDAR have the advantages of fast acquisition speed and smaller data volume, allowing to collect point cloud data over a large area [
14,
15]. However, ALS and UAV LiDAR use a top-to-bottom scanning mode to acquire point cloud data, leading to the sparse distribution of trunk point clouds, which is a great challenge for trunk point extraction and direct DBH measurement from ALS or UAV LiDAR data. Most previous studies indirectly extracted the DBH from ALS or UAV LiDAR data through the DBH estimation models. The DBH estimation models can be divided into parametric and non-parametric models. The parametric DBH models were based on multiple regression with individual tree parameters (tree height, crown width, and density) [
16,
17,
18,
19]. In non-parametric models, the machine learning algorithms, such as random forest [
15,
20], support vector machine [
21], and artificial neural network [
22], were first used to extract individual tree features and then estimate DBH. However, both parametric and non-parametric models require field DBH measurement for model development, limiting their practical applicability. Therefore, the direct measurement of DBH is essential. To our best knowledge, several previous studies attempted to directly measure DBH from ALS or UAV LiDAR data [
23,
24,
25,
26,
27]. These existing studies have successfully estimated the DBH from extremely high-density LiDAR point data, and proven that tilted scanning can increase the trunk point cloud density so as to achieve accurate measurement of DBH [
28]. However, the feasibility of their method in relatively low-density point clouds has not been demonstrated. Additionally, these studies did not assess the effect of scanning degree on trunk point extraction and DBH measurement.
This paper aims to explore the potential of relatively low-density UAV LiDAR data for trunk point extraction and direct DBH measurement as well as assess the effects of scanning angles and scanning modes. The main contributions of this paper are as follows: (1) The trunk point cloud of each tree from original UAV LiDAR data is extracted through the multiscale cylindrical detection method. (2) An effective direct method is proposed to measure the DBH from the trunk point cloud based on the multiscale ring fitting. (3) The effects of scanning angles and modes on trunk point extraction and DBH measurement are analyzed. This paper is the first attempt to analyze the influence of scanning angles and modes on trunk point extraction and direct DBH measurement, and it will provide scientific guidance for the planning of UAV routes and LiDAR scanning design.
3. Methods
This paper investigates the potential of UAV LiDAR data for trunk point extraction and direct DBH measurement. Additionally, the influence of scanning angles and modes on their accuracies is analyzed.
Figure 4 depicts the flowchart of this study. First, the real and simulated UAV LiDAR data were obtained with the designed scanning angles and modes and various scanning angles and modes, respectively. Then, those real and simulated data were preprocessed, including noise removal, point cloud filtering, and normalization. Third, the trunk point cloud for each tree was extracted from UAV LiDAR data based on the multiscale cylindrical detection method. Fourth, the DBH was measured based on the sliced trunk point clouds using a multiscale ring fitting method. Lastly, we analyzed the effects of scanning angles and modes on trunk point extraction and DBH measurement. In this study, the LiDAR data were preprocessed using the PCM (PCM was developed by the team of Prof. Cheng Wang, Chinese Academy of Sciences,
http://www.lidarcas.cn/soft, accessed on 14 April 2022) software, and the other operations were implemented by python programming.
3.1. Trunk Point Extraction
We used a multiscale cylindrical detection method to extract trunk point cloud data from UAV LiDAR data, which is described in detail as follows.
(1) Point cloud preprocessing. The influence of surface topography on the subsequent extraction process is eliminated through the point cloud preprocessing. First, the point cloud data were classified into ground and non-ground points. In recent years, several filtering methods have been proposed for various natural and urban scenes [
33,
34,
35,
36,
37,
38]. Among them, the cloth simulation filtering (CSF) [
34] algorithm is used, which is suitable for plain areas with flat terrain. Second, the ground points were interpolated to generate the digital terrain model (DTM) (resolution of 0.25 m) using the inverse distance weighted (IDW) method. Lastly, the point cloud data was normalized along with the Z-coordinate, removing elevation information.
(2) Coarse extraction of the trunk point cloud data. The tree-apexes of broad-leaved forests are not obvious, and the trunk parts have fewer branches (as shown in
Figure 5a). Therefore, this study coarsely extracted the trunk point clouds using a multiscale cylinder. The normalized point cloud with an elevation range of 0.5–2 m was selected as the seed point (see
Figure 5a). Then, the multiscale cylinder was constructed using the z-direction of the unmarked seed point p (the red point in
Figure 5b) as the cylinder rotation axis and r as the radius, where the value of r is determined according to the range of DBH values in the study area. The cylindrical point set T was obtained by clipping the normalized point cloud using a multiscale cylinder. Given that the seed point cloud may contain trunk and branch points, the physical form of the tree was used to determine whether the point set T belonged to the trunk point cloud. Since the tree trunk is generally a near cylinder, the trunk point cloud was uniformly distributed in the vertical direction. Additionally, the heights of all trees in this study area were higher than 2 m. Therefore, the cylindrical point set T was regarded as the trunk point cloud in this study when the point set T was uniformly distributed in the vertical direction and its maximum height was greater than 2 m. If T was the trunk point cloud, all points in T were marked; in contrast, only point p was marked to avoid repeated calculation. The above process was repeated until all seed points were marked, as shown in
Figure 5c. The different colors in
Figure 5c represent different segmented trunk point clouds during coarse extraction.
(3) Fine extraction of the trunk point cloud. The point cloud may be over-segmented during the process of coarse extraction (as shown in
Figure 5c), which may significantly affect the DBH measurement. To better conduct the subsequent DBH measurement, we needed to merge the over-segmented trunk point clouds through fine extraction. There are two key steps to implement it. First, the center coordinates of each trunk point in the XOY plane were first calculated based on the coarsely extracted point cloud. Second, the trunk points whose centers were close (Euclidean distance is less than 1 m) were merged, as shown in
Figure 5d.
(4) Accuracy evaluation. The trunk extraction accuracy was evaluated using the ratio of the number of correctly extracted tree trunks to the ground-truth. The positions of each tree can be erroneous due to the poor signal of GPS in the forest. Thus, the Euclidean distance in the XOY plane between the extracted trunk point cloud and the coordinates measured by GPS were used for the determination of correctly extracted tree trunks.
3.2. Direct Measurement of DBH
For direct measurement of DBH, a ring fitting-based method is proposed to overcome the sparse distribution of UAV point clouds and non-circular trunk cross-section. The proposed method was conducted as follows.
(1) Trunk point cloud slices. For TLS data, only point cloud slices with a height of about 1.3 m are generally obtained to measure DBH [
39,
40]. However, our study selected the point clouds with elevations ranging from 0.5 to 2 m to measure the DBH because it can not only ensure enough trunk points, but also exclude branch points that affect the DBH measurement.
(2) Parameter setting of rings. This paper adopted a multiscale ring fitting for DBH measurement instead of RANSAC or Hough transform, which are typical methods suitable for ordered point clouds. The centers of the multiscale rings were first determined. The centers of the rings were drawn inside the outer rectangle of the sliced point cloud in the XOY plane projection at an interval of 1 cm (the black dots in
Figure 6a represent the center positions of the multiscale rings). Then, the radius range of the multiscale rings was determined, where the minimum and maximum radiuses of the large circle in each ring were 3 cm and the half of the larger side of the outer rectangle, respectively. The radius difference between the large and small circles was 1 cm. Finally, the multiscale rings were constructed with the determined circle center positions and radius sizes.
(3) DBH measurement. The ring with the largest number of points was considered as the best fitting ring, as shown in
Figure 6b. Lastly, the average of large and small circle diameters in the best fitting ring was computed as the DBH.
(4) Accuracy assessment. The accuracy of DBH is evaluated using the coefficient of determination (R
2) [
41,
42,
43,
44] and the square root of mean squared error (RMSE, in m) [
45].
where
and
represent the field-measured DBH and UAV LiDAR-derived DBH of the tree
i, respectively.
represents the average of all field-measured DBH values, and
n is the number of trees.
3.3. Influence of Scanning Angles and Modes on Trunk Point Extraction and DBH Measurement
UAV LiDAR data acquired by vertical scanning from top to bottom can cause the sparsity of obtained point clouds in the trunk in forest areas, which is not conducive to the extraction of individual tree parameters. Thus, in this study, a tilt scanning pattern was used for data acquisition.
Figure 7a–c compares the schematic diagram of vertical and tilt scanning and their point cloud profiles, showing a significantly different density of the obtained trunk point cloud with different scanning angles. When the scanning angle is too large, most of the obtained point clouds are canopy points; in contrast, when the scanning angle is too small, the number of trunk points is small due to the sheltering of tree canopies. Additionally, the point could density is affected by the scanning mode [
28].
Figure 8 depicts different scanning modes, including single-route, double-route, triple-route, and quadruple-route. In this paper, the influence of the scanning angles and scanning modes on trunk point extraction and DBH measurement were analyzed on the UAV LiDAR data. We employed the LESS [
31,
32] model to simulate UAV LiDAR data for multiple routes with a scanning angle range of 45–90 degrees (interval of 5 degrees).