Intelligent Estimating the Tree Height in Urban Forests Based on Deep Learning Combined with a Smartphone and a Comparison with UAV-LiDAR

Abstract

In this paper, a method for extracting the height of urban forest trees based on a smartphone was proposed to efficiently and accurately determine tree heights. First, a smartphone was used to obtain person–tree images, LabelImg was used to label the images, and a dataset was constructed. Secondly, based on a deep learning method called You Only Look Once v5 (YOLOv5) and the small-hole imaging and scale principles, a person–tree scale height measurement model was constructed. This approach supports recognition and mark functions based on the characteristics of a person and a tree in a single image. Finally, tree height measurements were obtained. By using this method, the heights of three species in the validation set were extracted; the range of the absolute error was 0.02 m–0.98 m, and the range of the relative error was 0.20–10.33%, with the RMSE below 0.43 m, the rRMSE below 4.96%, and the ${\mathrm{R}}^2$ above 0.93. The person–tree scale height measurement model proposed in this paper greatly improves the efficiency of tree height measurement while ensuring sufficient accuracy and provides a new method for the dynamic monitoring and investigation of urban forest resources.

1. Introduction

The concept of an urban forest originated in the United States and Canada [1,2]. Such a forest is a biological community consisting of trees, flowers, wildlife, microorganisms, and architectural structures within a city, including parks, street and unit green spaces, vertical greenery, street trees, sparse lawns, patches of forest, forest strips, flower gardens, nurseries, orchards, vegetable patches, farmland, meadows, and water areas [3]. Urban forests mitigate the decline in environmental quality associated with urban development by influencing changes in the physical space and biological environment of cities and provide important functions in urban CO₂ emission reduction, air purification, PM_2.5 absorption, urban rainfall storage, flood control, water quality improvement, noise reduction, and microclimate improvement [4,5,6]. In particular, the carbon sequestration function of urban forests plays an important role in the establishment of green and low-carbon cities [7,8,9]. In recent years, the performance of smartphones has gradually improved, and they have been widely used to extract forest parameters. For example, we use smartphones to estimate the leaf area index (LAI) [10,11] and plant area index (PAI) [12] and classify tree species [13]. In the forest parameters, the height of trees is a key factor when accurately estimating the forest biomass carbon [14,15,16]. Therefore, rapid and intelligent monitoring of urban forest tree height is a key process in urban forest resource monitoring and carbon sink estimation [17].

For many years, the instrument most commonly used to measure the height of trees in field surveys was the Blume–Leiss height gauge [18]. When measuring tree height, at least two people were required to work together, and a tape measure was needed in addition to the gauge. Other traditional tree height measurement methods, such as total station and angle meter-based methods [19,20], involve height-measurement principles similar to that of the Blume–Leiss method. Therefore, traditional tree height measurement methods based on the the Blume–Leiss approach are characterized by low measurement efficiency and high labor intensity.

With the rapid development of remote sensing (RS) technology and intelligent forestry, some researchers have achieved success in research involving tree height measurement methods. The monitoring of forest resources in China has been transformed from relying on manual field survey methods to the use of various information technologies to achieve accurate and quantitative monitoring [21]. An increasing number of specialized and intelligent devices have been used for forest resource surveys [22,23,24]. Currently, technologies such as unmanned aerial vehicles (UAVs), photogrammetric systems, and light detection and ranging (LiDAR) systems have gradually become important in forest resource monitoring [25,26,27,28]. UAV-LiDAR systems can quickly and accurately obtain forest structure parameters [29,30]. However, there are many deficiencies associated with these systems, such as poor portability and large numbers of calculations [31,32,33]. For urban forests, many difficulties will be encountered when applying UAV-LiDAR systems in surveying tasks, such as flight permissions in urban centers and a sparse LiDAR point cloud density at high flight altitudes.

The high-definition cameras contained in smartphones can provide rapid images of single trees in urban areas, and tree height can be determined through imaging and photogrammetry principles. Compared with methods based on UAV-LiDAR and the Blume–Leiss height gauge, recently proposed methods [34,35] are much more efficient and broadly applicable. However, the method proposed by Chen et al. can only measure the height of trees on the same horizon as the measurer who use the smartphone, which limits its application in tree height measurement. Although the method proposed by Wu et al. overcomes this deficiency, their method still is insufficient in other ways. For example, the parameters of the rear camera of the smartphone must be calibrated. If the brand or model of the smartphone being used changes, recalibration is required. Additionally, in the image segmentation algorithm they used, considerable time is needed to extract the edges of trees, making this approach less convenient than methods based on target detection algorithms [21,36]. The methods proposed by Tian et al. [37] and Fan et al. [38] provide high accuracy but require special external tools, thus limiting their use by forestry investigators in daily practice.

Based upon the aforementioned limitations, an intelligent urban tree height extraction method based on monocular vision, using a smartphone is proposed and combined with You Only Look Once v5 (YOLOv5). YOLOv5 provides high target detection accuracy and a fast run speed [39,40,41]. With this approach, the smallest possible representative box can be established for trees, making real-time measurement of tree height a reality. The key contribution of this study is the use of a smartphone to measure the height of urban trees in real time, an approach with strong practicality and convenience. Data are accurately and efficiently obtained, thus reducing the field work intensity of forestry investigators.

2. Study Area and Methodology

2.1. Study Area and Data

2.1.1. Study Area

In this study, Cinnamomum camphora (CC), Ginkgo biloba (GB), and Yulania denudate (YD) in the eastern section of Wusu Street, Lin’an District, Hangzhou, Zhejiang Province, China, and the whole section of Ginkgo Road and Yulan Road, Zhejiang Agriculture and Forestry University, were used to carry out relevant studies. The locations are shown in Figure 1. GB is a deciduous tree species with a wide distribution, and it exists in a wild state in only the Tianmu Mountains, Zhejiang [42]. CC is a subtropical evergreen broad-leaved tree species [43]. YD is a deciduous tree species [44]. The three tree species are common street trees along all urban streets in Hangzhou.

2.1.2. Person–Tree Image Acquisition and Labeling

A total of 470 person–tree images were collected at a resolution of 3024 × 4032 by using smartphones to photograph street trees of the three species in the study area. A total of 396 images with high clarity and no motor vehicle obstruction were selected as the study dataset.

The constructed dataset was divided into a training set and a validation set at a ratio of 3:1 by using a random nonrepeat sampling method to train and validate the person–tree scale height measurement model. In the training set, there were 36 CCs, 33 GBs, and 30 YDs. There were 33 trees in the validation set. Each tree had three different photos. The 132 trees in the dataset were surveyed by using a traditional field survey method to obtain the measured height of trees. In addition, UAV-LiDAR was used to scan the trees from above. DSM and DEM data were used to calculate the tree height, which was used for the verification of the extraction results of the person–tree scale height measurement model. The UAV was a DJI Matrice 600Pro, the LiDAR device was a Velodyne Puck LITE^TM, the flight date was April 2022, and the flight altitude was 60 m. The standard deviation of the point cloud was 2.39 cm, which was less than 3.00 cm specified in the laser radar product manual; therefore, the relevant accuracy requirements were met. The image capture and tree height investigation processes are shown in Figure 2.

LabelImg software was used to label the images in the training set in YOLO format with “person” and “tree” categories. LabelImg generated a text file with the same name corresponding to each image when the labeling was completed. Finally, the label category, the ratio of the horizontal coordinate of the center of the label box to the width of the image, the ratio of the vertical coordinate of the center of the label box to the height of the image, the ratio of the width of the label box to the width of the image, and the ratio of the height of the label box to the height of the image were saved. The images obtained before and after labeling are shown in Figure 3.

2.2. Person–Tree Scale Height Measurement Model

2.2.1. The Construction of Person–Tree Scale Height Measurement Model

The object detection algorithms can be divided into one-stage and two-stage detection algorithms [45]. YOLOv5 is the representative of one-stage detection algorithms, which balances speed and precision [40,46,47]. The network structure of YOLOv5 consists of four parts, namely, the input, backbone, neck, and head, which are represented in yellow, blue, purple and red, respectively, in Figure 4 [48,49,50].

The person–tree images input to the network will go through the adaptive image scaling of input [51], feature extraction of backbone [45,52,53,54], feature fusion of neck [50,51,55,56], and NMS (nonmaximum suppression) of head [57] in turn, and finally the coordinates of the upper left and lower right corners of the person and the tree are output to determine the corresponding prediction box for each target.

Among them, the head uses complete intersection over union (CIoU) [58] to calculate the CIoU loss and incorporates it into the bounding box (Bbox) [45] loss function, which integrates the overlap area, centroid distance and aspect ratio to perform box regression prediction. A schematic diagram of ${\mathrm{CIoU}}_{\mathrm{Loss}}$ calculation is shown in Figure 5. Red box A represents the prediction box, and “a” is the center point of the prediction box. Purple box B represents the real box, and “b” is the center point of the real box. The smallest external box that can enclose the prediction box A and the real box B is denoted by C. The calculation of CIoU_LOSS is as follows,

$$\mathrm{IoU}=\frac{\left|\mathrm{A}{\displaystyle \cap }\mathrm{B}\right|}{\left|\mathrm{A}{\displaystyle \cup }\mathrm{B}\right|}$$

(1)

$$\mathrm{CIoU}=\mathrm{IoU}-\frac{{\mathsf{\rho }}^2\left(\mathrm{a},\mathrm{b}\right)}{{\mathrm{c}}^2}-\mathsf{\alpha }\mathrm{v}$$

(2)

$$\mathrm{v}=\frac{4}{{\mathsf{\pi }}^2}\left(\arctan \frac{{\mathrm{w}}^{\mathrm{gt}}}{{\mathrm{h}}^{\mathrm{gt}}}-\arctan \frac{{\mathrm{w}}^{\mathrm{p}}}{{\mathrm{h}}^{\mathrm{p}}}\right)$$

(3)

$$\mathsf{\alpha }=\frac{\mathrm{v}}{\left(1-\mathrm{IoU}\right)+\mathrm{v}}$$

(4)

$${\mathrm{CIoU}}_{\mathrm{Loss}}=1-\mathrm{IoU}+\frac{{\mathsf{\rho }}^2\left({\mathrm{a}}^{\mathrm{p}},{\mathrm{b}}^{\mathrm{gt}}\right)}{{\mathrm{c}}^2}+\mathsf{\alpha }\mathrm{v},$$

(5)

where $\left|\mathrm{A}{\displaystyle \cap }\mathrm{B}\right|$ denotes the area of intersection between the prediction box and the real box. $\left|\mathrm{A}{\displaystyle \cup }\mathrm{B}\right|$ denotes the area of the merging region of the prediction box and the real box. $\mathsf{\rho }\left(\mathrm{a},\mathrm{b}\right)$ is the Euclidean distance between the ”a” and “b” center points, “c” is the diagonal distance across C, $\mathsf{\alpha }$ is a balance scale parameter and is not involved in the gradient calculation, $v$ is a parameter associated with the consistency of the aspect ratio, $w^{gt}$ and $h^{gt}$ are the width and height of the real box, and $w^p$ and $h^p$ are the width and height of the prediction box, respectively.

Based on the coordinates of the prediction box, principle of small-aperture imaging and scale, a person–tree scale height measurement model is constructed, as shown in Figure 6. By using a person as the scale, the real height is H1. The tree will be measured, and its real height is H2. The person and the root of the tree are located on the same axis and form plane α. The smartphone camera uses a rear camera lens group, and the light reflected from the person and the top and bottom of the tree passes through the optical center O of the lens group. ABCD is the imaging plane of the smartphone, and the pixel heights of the person and tree in the imaging plane are Pix_H1 and Pix_H2, respectively. They can be calculated from the pixel coordinates of the upper left point and the lower right point of the rectangular box containing the target object. The imaging plane ABCD, smartphone in the plane, and α plane are parallel to each other. From the principle of small-aperture imaging and the principle of scale, it is easy to obtain the height measurement formula for a target tree, as follows:

$$\frac{H1}{H2}=\frac{Pix\_H1}{Pix\_H2}$$

(6)

In summary, the flow chart for extracting the height of urban forest trees based on a smartphone with the person–tree scale height measurement model as shown in Figure 7. The measured value of traditional field survey method is denoted as $\mathrm{Field}\_{\mathrm{Height}}_{\mathrm{ALL}}$. The UAV-LiDAR value is denoted as $\mathrm{LiDAR}\_{\mathrm{Height}}_{\mathrm{ALL}}$. The predicted value of smartphone is denoted as $\mathrm{Phone}\_{\mathrm{Height}}_{\mathrm{ALL}}$.

2.2.2. Training of the Person–Tree Scale Height Measurement Model

The images in the constructed training set were labeled, and then the training of the person–tree scale height measurement model was performed. The hardware environment for the experiments was as follows. The CPU configuration was an Intel(R) Xeon(R) Gold 6154 CPU @ 3.00 GHz, the graphics card configuration was an NVIDIA RTX A6000, the video memory size was 47.5 GB, and the memory size was 64 GB. The software environment for the experiments was as follows. The operating system was Windows 10, the CUDA version was 11.1, the Python version was 3.9.7, and the PyTorch version was 1.8.1.

In this study, the model was trained based on Ultralytics LLC YOLOv5, and the small-hole imaging and scale principles were innovatively incorporated into the model. To prevent gradient explosion, the initial learning rate of the weights was set to 0.01, and the learning rate was dynamically reduced by using the cosine function. The momentum factor was set to 0.937, and the pretrained model was YOLOv5x.pt. The number of training epochs was set to 300, and the batch size of the model training was set to 16. The image input size was 1024 × 1024, the number of working cores was 32, and other parameters were set to the default values.

2.2.3. Evaluation Indexes for the Accuracy of the Person–Tree Scale Height Measurement Model

To evaluate whether the person–tree scale height measurement model can accurately achieve the detection of persons and trees, six evaluation indicators were used: precision (P), recall rate (R), mean average precision (mAP), box loss, object loss, and classification loss. The larger the P, R, and mAP are, the higher the precision of the person-tree scale height measurement model in detecting and recognizing persons and trees. The smaller the values of the box loss, object loss, and classification loss are, the better the performance of the person–tree scale height measurement model in detecting and recognizing persons and trees.

P is also known as the accuracy rate, which is based on the model results and indicates the proportion of the samples predicted to be positive that are true positive, as calculated with Equation (7):

$$\mathrm{P}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}$$

(7)

R is also known as the check-all rate, which is based on the original samples and indicates the proportion of the samples predicted to be true positive in the original samples, as calculated with Equation (8):

$$\mathrm{R}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$$

(8)

TP, FP, and FN in Equations (7) and (8) denote the true positive samples, false-positive samples, and false-negative samples, respectively.

mAP@[0.5:0.95] represents dividing the IoU threshold range from 0.5 to 0.95 in steps of 0.05 and calculating accuracy statistics in different ranges to determine the average accuracy for different IoU threshold ranges. The corresponding formulae are shown in Equations (9) and (10), in which AP is the area under the accuracy-recall curve:

$$\mathrm{AP}={{\displaystyle \int }}_0^1\mathrm{PdR}$$

(9)

$$\mathrm{mAP}=\frac{{{\displaystyle \sum }}_{\mathrm{i}=1}^{\mathrm{N}}{\mathrm{AP}}_{\mathrm{i}}}{\mathrm{N}}$$

(10)

In addition, the box loss represents the ability of the model to locate the center of the target object and the extent to which the boundaries of the predicted box cover the target object. The object loss is essentially a measure of the probability that the target object is present in the proposed region of interest. If the objectivity is high, it means that the image window is likely to contain a target object. The classification loss provides a way for the model to predict the correct class of a given object.

2.3. Evaluation of the Tree Height Extraction Accuracy

Height measurement experiments were performed with the person–tree images in the dataset to obtain the corresponding model predictions for each tree. The $\mathrm{Field}\_{\mathrm{Height}}_{\mathrm{ALL}}$ was measured by using a band tape and a Blume–Leiss height gauge [59,60]. In addition, it has been noted that LiDAR can accurately extract single-tree height information [61,62,63]. Therefore, the $\mathrm{Phone}\_{\mathrm{Height}}_{\mathrm{ALL}}$ was compared with the $\mathrm{Field}\_{\mathrm{Height}}_{\mathrm{ALL}}$ and the $\mathrm{LiDAR}\_{\mathrm{Height}}_{\mathrm{ALL}}$ to determine the accuracy of the person–tree scale height measurement model. The absolute error (AE), relative error (RE), root mean square error (RMSE), relative root mean square error (rRMSE), and coefficient of determination R² between the three types of tree heights were calculated as indices to evaluate the accuracy of tree height extraction by using Equations (11)–(15),

$$AE=\left|h_i-h_{ri}\right|$$

(11)

$$RE=\frac{\left|h_i-h_{ri}\right|}{h_{ri}}$$

(12)

$$RMSE=\sqrt{\frac{{{\displaystyle \sum }}_{i=1}^n{\left(h_i-h_{ri}\right)}^2}{n}}$$

(13)

$$RMSE\%=\frac{RMSE}{\overline{h_r}}\times 100\%$$

(14)

$$R^2=\frac{{{\displaystyle \sum }}_{i=1}^n{\left(h_i-h_{ri}\right)}^2}{{{\displaystyle \sum }}_{i=1}^n{\left(h_i-\overline{h_r}\right)}^2}$$

(15)

where $h_i$ and $h_{ri}$ are the values of different types for the $i^{\mathrm{th}}$ tree, $\overline{h_r}$ is the average of the values, and n is the number of trees.

3. Results and Analysis

3.1. Person–Tree Scale Height Measurement Model

Figure 8a shows the curves of P, R, and mAP@[0.5:0.95] for the model after 300 training epochs. As shown in Figure 8a, the values of the three indicators fluctuate and display a decreasing trend until the 60th training epoch, but as the number of training epochs increases and the learning rate of the cosine function is dynamically adjusted, the model gradually transitions to a convergence state. At the end of the last iteration, the accuracy rate, the recall rate, and the mAP indicate high accuracy.

The curves of the box loss, object loss, and classification loss for the training set are shown in Figure 8b. Notably, the curves of the three metrics for the training set decrease sharply before the 50th training epoch. The curves of the box loss and object loss stabilize and decrease to less than 0.01 after approximately the 250th training epoch. The curve of the classification loss stabilizes and decreases to less than 0.0025 after approximately the 100th training epoch.

3.2. Tree-Height Extraction Results

The trained model was used to extract a total of 132 tree heights for the three tree species in the dataset. The model-labeled person–tree images are shown in Figure 9, and the corresponding stand labels for (a), (b), and (c) are GB36, CC18, and YD8, respectively. Figure 9 illustrates that the confidence level of the tree height measurement model for person and tree recognition is above 0.94, which is high.

The Field_Height and Phone_Height, AE, and RE for the three tree species of validation set are shown in

Table 1. The Field_Height_GB, Phone_Height_GB and error analysis.

Tree No.	Field_HeightGB /m	Phone_HeightGB /m	Absolute Error/m	Relative Error/%
GB 6	8.30	8.43	0.13	1.54%
GB 8	11.20	11.31	0.11	0.98%
GB 13	10.60	11.52	0.92	8.68%
GB 16	11.10	11.28	0.18	1.62%
GB 24	9.60	9.83	0.23	2.34%
GB 27	11.60	12.58	0.98	8.45%
GB 30	12.20	12.39	0.19	1.56%
GB 36	10.50	10.61	0.11	1.05%
GB 38	11.20	11.13	0.07	0.63%
GB 42	10.20	10.22	0.02	0.20%
GB 44	8.75	9.05	0.30	3.42%

,

Table 2. The Field_Height_CC, Phone_Height_CC and error analysis.

Tree No.	Field_HeightCC /m	Phone_HeightCC /m	Absolute Error/m	Relative Error/%
CC 3	7.20	7.50	0.30	4.14%
CC 5	6.20	6.70	0.50	8.03%
CC 6	6.70	6.66	0.04	0.58%
CC 13	5.90	5.99	0.09	1.58%
CC 18	6.80	6.77	0.03	0.44%
CC 26	90	9.46	0.46	5.09%
CC 28	6.10	5.80	0.30	4.98%
CC 30	6.00	6.04	0.04	0.63%
CC 32	6.90	6.80	0.10	1.49%
CC 45	8.30	8.41	0.11	1.33%
CC 46	8.80	8.57	0.23	2.63%
CC 47	10.80	11.74	0.94	8.70%

and

Table 3. The Field_Height_YD, Phone_Height_YD and error analysis.

Tree No.	Field_HeightYD /m	Phone_HeightYD /m	Absolute Error/m	Relative Error/%
YD 3	10.20	10.23	0.03	0.29%
YD 4	9.20	10.15	0.95	10.33%
YD 8	8.00	8.22	0.22	2.69%
YD 11	7.30	7.62	0.32	4.32%
YD 21	6.20	6.58	0.38	6.13%
YD 28	8.60	8.98	0.38	4.44%
YD 31	6.70	7.11	0.41	6.07%
YD 34	9.60	9.69	0.09	0.90%
YD 36	10.50	10.97	0.47	4.48%
YD 37	10.20	10.53	0.33	3.24%

. The distribution and fitting results for the Field_Height and Phone_Height of the three tree heights are shown in Figure 10a–d.

The $\mathrm{Phone}\_{\mathrm{Height}}_{\mathrm{GB}}$ based on the person–tree scale measurement height model are shown in Figure 10a. An analysis of Figure 10a indicates that the ${\mathrm{R}}^2$ values of the model for the training set and validation set are 0.9710 and 0.9311, respectively. Additionally, the RMSEs are 0.29 m and 0.43 m, respectively, and the rRMSEs are 2.84% and 4.13%. The tree height predictions based on the validation set are further analyzed in Table 1. The minimum and maximum values of the $\mathrm{Phone}\_{\mathrm{Height}}_{\mathrm{GB}}$ are 8.43 m and 12.58 m, respectively, with a range of absolute error from 0.02 m to 0.98 m and a range of relative error from 0.20% to 8.68%.

The $\mathrm{Phone}\_{\mathrm{Height}}_{\mathrm{CC}}$ estimated by the height measurement model are shown in Figure 10b. An analysis of Figure 10b indicates that the accuracy (${\mathrm{R}}^2$) of the model for the training set and validation set is 0.9728 and 0.9726, respectively. The RMSEs are 0.26 m and 0.37 m, and the rRMSEs are 3.53% and 4.96%, respectively. The tree height predictions based on the validation set are further analyzed in Table 2. The minimum and maximum values of $\mathrm{Phone}\_{\mathrm{Height}}_{\mathrm{CC}}$ are 5.80 m and 11.74 m, respectively, with absolute errors ranging from 0.03 m to 0.94 m and relative errors ranging from 0.44% to 8.70%.

The $\mathrm{Phone}\_{\mathrm{Height}}_{\mathrm{YD}}$ estimated by the height measurement model are shown in Figure 10c. An analysis of Figure 10c indicates that the accuracy (${\mathrm{R}}^2$) of the model for the training set and validation set is 0.9362 and 0.9736, respectively. The RMSEs are 0.40 m and 0.43 m, respectively, and the rRMSEs are 4.71% and 4.96%, respectively. The $\mathrm{Phone}\_{\mathrm{Height}}_{\mathrm{YD}}$ based on the validation set are further analyzed in Table 3, and the minimum and maximum values are 6.58 m and 10.97 m, respectively, with absolute errors ranging from 0.03 m to 0.95 m and relative errors ranging from 0.29% to 10.33%.

The $\mathrm{Phone}\_{\mathrm{Height}}_{\mathrm{ALL}}$ of the three tree species estimated by the tree height measurement model are shown in Figure 10d. As shown in Figure 10d, the accuracy (${\mathrm{R}}^2$) of the model for the training set and validation set is 0.9746 and 0.9766, respectively, the RMSEs are 0.32 m and 0.41 m, and the rRMSEs are 3.67% and 4.64%, respectively. In the validation set, the minimum and maximum values of $\mathrm{Phone}\_{\mathrm{Height}}_{\mathrm{ALL}}$ predicted by the model are 5.98 m and 12.58 m, respectively, with absolute errors ranging from 0.02 m to 0.98 m and relative errors ranging from 0.20% to 10.33%.

The distribution and fitting results for the Phone_Height and the LiDAR_Height are shown in Figure 10e–h. Some tree point cloud data are shown in Figure 11.

Based on Figure 10e–g, the ${\mathrm{R}}^2$ range of the training set is 0.9473~0.9737, the RMSE range is 0.28~0.41 m, and the rRMSE range is 3.26~4.60%. The ${\mathrm{R}}^2$ range of the validation set is 0.9861~0.9973, the RMSE range is 0.16~0.27 m, and the rRMSE range is 1.72~3.68%. Figure 10h is obtained from a comprehensive analysis of the trees in the dataset. The ${\mathrm{R}}^2$ value of training set is 0.9775, the RMSE is 0.34 m, and the rRMSE is 4.02%. The ${\mathrm{R}}^2$ value of validation set is 0.9929, the RMSE is 0.21 m, and the rRMSE is 2.39%.

The distribution and fitting results for the Field_Height and the LiDAR_Height are shown in Figure 10i–l. Based on Figure 10i–k, the ${\mathrm{R}}^2$ range of the training set is 0.9245~0.9474, the RMSE range is 0.30~0.43 m, and the rRMSE range is 3.58~5.83%. The ${\mathrm{R}}^2$ range of the validation set is 0.9640~0.9857, the RMSE range is 0.28~0.31 m, and the rRMSE range is 2.66~4.21%. Figure 10l is obtained from a comprehensive analysis of the trees in the dataset. The ${\mathrm{R}}^2$ value of training set is 0.9598, the RMSE is 0.37 m, and the rRMSE is 4.32%. The ${\mathrm{R}}^2$ value of validation set is 0.9822, the RMSE is 0.30 m, and the rRMSE is 3.38%.

Therefore, the results involving Field_Height, Phone_Height, and LiDAR_Height indicate that the height of the trees extracted by the person-tree scale height measurement model are relatively accurate. The person–tree scale height measurement model constructed in this study based on a smartphone combined with YOLOv5 predicts the tree height of GB, CC, and YD tree species with high accuracy and low errors, reflecting good performance.

4. Conclusions and Discussion

In this study, a person–tree scale height measurement model that combines a smartphone with YOLOv5 and uses the principles of small-aperture imaging and scale is proposed. These principles correspond to the ratio of the actual height of a person to the pixel difference associated with the person in the imaging plane to achieve accurate measurements of tree height in urban forests, with ${\mathrm{R}}^2$ values exceeding 0.97 and low relative error at less than 10.33%. Further analysis of three tree height-extraction methods for three tree species shows that there is no significant difference (Figure 12). Thus, the person–tree scale height measurement model performs well in extracting tree height, which is no different from the tree height obtained by traditional field survey and UAV-LiDAR.

By analyzing the tree height results for the validation set, it was found that the relative error was 8.00% or more for only 5 of the 33 trees in the validation set. Figure 13 shows the measurement results for YD4. The red boxes denote the main sources of error. The box used to encompass the tree in the model is a little large, resulting in Pix_H2 in Formula 6 being large. Weeds around the root of the tree block the image of the person, resulting in Pix_H2 in Formula 6 being small. Therefore, the measurement result for YD 4 is a little large.

The main reasons for error are as follows. (1) When the person–tree images were acquired, the shooting distance was improper or the plane formed by the person and the tree to be measured was not parallel to the camera imaging plane. (2) The model did not accurately mark the person and the tree.

In other cases, the proposed altimetric model cannot successfully acquire a complete person–tree image in a limited space, and the measurement accuracy may be highly susceptible to the influence of the canopy when performing person–tree image acquisition [64]. In the actual measurement process, the location of a tree is most likely a street with a high traffic flow, and the tree height is likely large; therefore, the photographer must remain a certain distance from the target object to capture the entire tree. However, in the real space, this distance may not be achieved due to obstructions, and the acquisition of person-tree images will fail [65]. If the distance is achievable, the photographer may be located in the middle of the road, which increases the danger of operations. In addition, when acquiring a person–tree image, it is important to keep the plane formed by the person and the tree parallel to the camera imaging plane.

In addition, due to the hardware limitations of the monocular lens, the depth information cannot be effectively extracted in all cases, resulting in the height measurement model not being able to effectively distinguish between the tree to be measured and background trees. When creating the dataset, it was found that when capturing the 3D space in 2D images, the canopy of the tree to be measured and the canopy of background trees were similar in color, and it was difficult to distinguish their contours, which could not be marked effectively by using LabelImg, thus affecting the recognition accuracy of the trees in the altimetric model.

In summary, in future research on tree height extraction, a measuring distance function can be implemented for effective extraction of image depth information by using a binocular camera [66]. Additionally, YOLOv5 can be improved to use the extracted depth information to effectively exclude the interference from background trees when identifying trees and then process multiple tree images or videos to realize the fast and intelligent extraction of forest structure parameters such as tree height and crown canopy width without reference.