1. Introduction
Casuarina equisetifolia L. is vital as a windbreak in coastal areas, with a wide natural range in Australia, Southeast Asia, and the Pacific Islands [
1]. This tree species also has characteristics that include symbiotic nitrogen fixation [
2,
3], the ability to stabilize sandy soils [
4,
5], drought and salt resistance [
6,
7], soil quality improvement, and soil rehabilitation [
2,
8]. Thus,
C. equisetifolia has become one of the most popular trees for blocking wind in sandy, coastal regions all over the world due to these various benefits [
9,
10]. In coastal shelter forests particularly,
C. equisetifolia has the ability to protect farmland and houses from natural disasters such as wind erosion and tsunamis [
11,
12], and can prevent damage from drifting sand [
10]. However, in some areas with a poor growing environment,
C. equisetifolia trees are particularly susceptible to various disturbances and prone to growing difficulties. In this context, it is necessary to obtain accurate information on
C. equisetifolia forest stands for forest management [
13]. Additionally, there is still a gap in the knowledge of how to optimize UAV flight and analysis parameters to measure tree characteristics for
C. equisetifolia. Individual trees and tree height are significant metrics for
C. equisetifolia forest management, especially considering the height [
14], density, open gaps [
15,
16], and structure [
10,
16], which can impact the sand stabilization and windbreak ability. A previous study reported that coastal vegetation was able to reduce the severity of a tsunami [
10], however, open gaps reduced this feature of
C. equisetifolia forests. Tree height determines the horizontal sheltered area, which is significant for reducing wind speed [
17]. Zhang et al. [
18] reported that quicksand can easily penetrate the forest belt and can cause sand damage if open gaps are formed inside the forest belt. Heisler and Dewalle found that, because the windbreak height reduces wind speed, the horizontal, sheltered area was determined by the windbreak height [
17]. In addition, in the up-wind direction of the windward side of a forest, the shelter area could extend to a distance of five times the height and a distance of thirty heights on the leeward side [
19]. Therefore, to manage and monitor
C. equisetifolia forests, there is a critical need for an effective tool to accurately detect forest parameters.
Unmanned Aerial Vehicles (UAVs) have become an important technology for assessing forest parameters due to their high spatial resolution, flexibility, and operability [
20,
21]. Among them, low-cost UAVs with RGB cameras have been widely applied in forest applications due to their ability to detect tree-level parameters accurately (e.g., individual trees and tree height). Notably, the structure from motion (SfM) technique is able to extract three-dimensional (3D) information from UAV imagery [
22,
23,
24]. Thus, monitoring and detecting the complex 3D structure of forests is feasible using low-cost UAVs with RGB sensors [
25,
26,
27]. A number of studies have reported the application of UAVs in forest-parameter extraction [
28]. For example, Guerra-Hernández et al. [
29] reported a relative root mean square error (r
RMSE) of 4.56% for the extracted tree height of a
Pinus pinea plantation using high-resolution UAV imagery. Combining the use of a UAV with oblique photogrammetry, Lin et al. [
30] assessed the individual tree heights of sparse, subalpine coniferous forest and yielded an accuracy of
RMSE = 1.77 m. Despite the promising accuracy of forest-parameter assessments achieved in previous studies, UAV flight planning is essential, and the flight parameters should be carefully considered [
31,
32,
33]. Many researchers have discovered that the flying altitude, image overlap, and other flight-planning settings can impact the quality or useability of final products [
33,
34,
35]. For example, the resolution of acquired images varies with flying altitude [
36,
37] and ultimately affects the identification of individual trees [
38] and tree crowns [
39]. Thus, flight parameters should be carefully considered and aligned with the actual measurement requirements.
An important aspect of processing UAV imagery for forestry applications is to select the appropriate algorithm for individual tree identification and tree height extraction from images [
40]. Several algorithms, such as the local maxima algorithm [
41], valley following [
42], extended-maxima transformation [
43], template matching [
44], Markov random fields [
44], and deep learning [
45,
46], have been successfully applied in the past decade. Among them, the local maxima algorithm is the most common detection method, primarily based on the maximum value of a moving window of a specified size that is chosen to identify treetops [
47,
48]. Pouliot et al. [
47] stated that the window size is important for detecting individual trees. If the window size is too small, a tree would be divided into several parts. Conversely, the individual trees may not be identified when the window size is too large. For example, Mohan et al. [
49] reported an inverse relationship between the fixed tree window size and tree density. Thus, the local maxima algorithm should determine an appropriate window size for object detection [
50].
Few studies have considered both the influence of flight parameters and algorithm parameters for detecting individual trees and tree heights. This study explores the influence of flying altitude and the window size of the local maxima algorithm for detecting individual trees and tree heights for C. equisetifolia in the Pingtan Comprehensive Pilot Zone. The main purpose of this study is as follows:
(1) To determine the optimal flight altitude for C. equisetifolia identification and tree-height estimation;
(2) To assess the optimal extraction parameters of the local maxima algorithm for C. equisetifolia identification and tree height estimation.
4. Discussion
This study focused on determining the optimal flight altitude and parameters of the local maxima algorithm to improve the accuracy of C. equisetifolia identification and tree-height estimation results. Although previous studies have mentioned the ability of C. equisetifolia to provide a windbreak, few studies reported monitoring this tree species effectively in the coastal regions. Our results showed the optimal accuracy of ITD and tree-height estimation when the flight altitude was 60 m and the combination of the CSWS and FCWS was 0.1 m × 0.8 m for the local maxima algorithm. This study can help acquire accurate forest-characterization information from UAVs at the tree-level, which contributes to forest management and forest tree identification.
Flight altitude is an important variable affecting UAV image resolution, ultimately affecting the accuracy of ITD and tree-height estimation [
36,
37,
63]. Our study found that the image quality could meet the accuracy requirement of ITD and tree-height estimation when the flight altitude was between 60 m and 80 m. Similar results have been reported in previous studies. For example, Dandois et al. [
36,
37,
63] stated that a flight altitude of 80 m is optimal for mapping forest structure, considering image collection and processing efficiency. Comparing three altitudes above ground level (65 m, 90 m, and 115 m) for UAV image collection, Swayze et al. [
64] reported that 65 m was the optimal flight altitude for detecting forest parameters (e.g., tree height, DBH, and density).
Image overlap influences the forest-parameter estimation less than flight altitude. There is a consensus that >75% forward overlap and 60%–80% side-lap are recommended for UAV image acquisition [
35]. Similarly, Tu et al. [
33] demonstrated that when the forward overlap was less than 80% the accuracy of the tree-height estimation decreased significantly, and that the optimal side-lap was between 70% and 80%. In this context, the forward overlap and side-lap in this study were set at 85% and 75%, respectively.
Previous studies have reported that a higher accuracy of forest-parameter estimation results when a lower flight altitude was used [
37]. However, our study indicated that a flight altitude of 40 m is not optimal for detecting
C. equisetifolia at the individual-tree level. This may be explained by declining image overlap in higher tree-height areas, which could decrease the accuracy of a 3D-generated reconstruction by SfM (
Figure 6). Contrarily, it has few impacts on the areas of lower tree height. Therefore, the appropriate flight height of a UAV should be determined for monitoring different tree species. For instance, Johansen et al. [
37] suggested that a flight height of 30 m was the optimal UAV parameter for tree crown perimeter, area, and height estimation compared to flight heights of 50 m and 70 m. This can be explained by the surveyed tree heights in the study by Johansen et al., which were mainly between 2 m and 3 m and would be subject to few impacts on image overlap.
The UAV flight altitude of 80 m in this study had clear advantages over a 60 m flight altitude. One advantage is that one single flight was able to cover the study area of approximately 6.5 ha using the charge of a single battery at the height of 80 m, while two batteries are needed to complete the flight at the height of 60 m. Moreover, a longer flight time with battery replacement would increase the risk of shadows from cloud cover. In addition, the number of obtained UAV images at a height of 80 m (3252) is less than the number of images acquired at 60 m (2178). Having fewer images can also reduce image-processing time. The SfM times were 11.02 and 9.11 minutes under the heights of 60 m and 80 m, respectively. Torres-Sanchez et al. [
65] compared imagery collected at 50 m and 100 m above ground level for mapping the structural parameters of olive trees and reported that both the time of flight as well as the multi-spectral images and RGB image processing were reduced from 47 min to 13 min and from 5 h 15 min to 1 h 8 min, respectively. Therefore, considering the efficiency of UAV image acquisition and processing, as well as the accuracy of ITD and tree-height detection, a flight altitude of 80 m was selected for ITD and tree-height estimation.
The local maxima algorithm is the most commonly used classical method, and window size is a critical factor affecting the accuracy of the ITD and tree-height estimation [
50]. Our study determined the optimal combination of a 0.1 m CSWS and a 0.8 m FCWS for ITD and tree-height estimation. From
Table 4 and
Table 7, it can be seen that when the FCWS was 0.8 m, the accuracy of the ITD increased with an increase in the CSWS. In contrast, the accuracy of tree-height estimation decreased with an increase in the CSWS. This is because the smoothed image becomes smoother with an increase in CSWS and outliers have less influence on the ITD, while the UAV generally underestimated the tree height (
Figure 4) for
C. equisetifolia and the accuracy of the tree-height estimation would be lower if a larger CSWS was used. Thus, a CSWS of 0.1 m was adopted in this study to balance the accuracy of the ITD and tree-height estimation.
In this study, the UAV-derived tree heights were underestimated at all flight altitudes. Previous studies have reported two reasons to explain the tree-height underestimation. One possible explanation is that the DTM is overestimated, which serves to cover the base of trunk information. This mainly occurs in areas with high forest coverage or abundant understory vegetation where it is challenging to obtain the ground position accurately [
57,
66,
67]. Another reason is that the resolution of the CHM is relative coarse compared to the size of treetops, which can result in the loss of treetops in the CHM [
57,
62,
67]. In our study, it was possible to obtain the ground position accurately because of the evenly distributed trees. Therefore, the underestimated tree heights in our study are most likely explained by the coarse resolution of CHM, which resulted in the loss of treetops. In this study, the resolution of CHM was 9.7 cm when the flight altitude was 80 m. The coarse resolution of the CHM relative to the thin treetop results in the underestimation of tree heights. The tree heights of
Figure 7a,b were 4.7 m and 5.1 m in the field measurement, respectively. The estimated tree heights from the UAV image were 2.98 m and 2.92 m at the altitude of 80 m, respectively, and the tree heights were underestimated by 1.72 m and 2.18 m. As can be seen from
Figure 7, the tree in
Figure 7b was more seriously underestimated than the tree in
Figure 7a due to the slender and obvious treetop.
The study site is located in a sensitive island environment, which results in the treetops of
C. equisetifolia being exposed to heavy wind erosion. The top of a tall
C. equisetifolia is more likely become full than the shorter trees under heavy wind erosion. Therefore, the identification of shorter trees is more often underestimated in the study area. Similar studies have reported that UAV-derived tree height was underestimated, but this deviation can be modified if it is consistent [
67,
68].
Casuarina equisetifolia was widely introduced into the coastal areas of Guangdong and Fujian, China, because of its ability to stabilize sandy soils [
4,
5], drought and salt resistance [
6,
7], soil-quality improvement potential, and soil-rehabilitation ability [
2,
8]. However, some
C. equisetifolia is susceptible to plant diseases and insect pests due to being planted in a pure stand. Therefore, detecting individual trees is helpful for monitoring the status of
C. equisetifolia forest stands. In addition, monitoring can identify open forest gaps and provide information on dead trees, such as their number and position, to support forest management (
Figure 5).
In addition to individual tree identification and tree-height estimation, other forest parameters can be estimated using UAV-derived DSMs and multi-spectral imagery, including tree crowns [
29,
56], DBH [
64,
69,
70], and biomass [
71,
72]. More detailed information can be detected rapidly and accurately using UAVs in the future as the capabilities of UAVs increase over time. Additionally, UAVs can track long-term dynamic changes in forest parameters using flights over time [
73,
74,
75,
76]. Low-cost consumer UAVs have been widely used in forestry applications because of their affordable characteristics. In this study, the proposed workflow was designed for the Phantom4-Multispectral UAV, so the flight parameters may not be directly suitable for other sensors. Nevertheless, the proposed procedure for evaluating the optimal flight altitude and extraction parameters will inform other workflows.