1. Introduction
Municipal forest management occurs at various scales throughout the urban landscape, but it often includes the direct management of individual trees. While conventional forestry is generally focused on increasing production at the stand level, urban forestry is more concerned with maximizing public forest benefits (e.g., aesthetics, shade, and property value) and minimizing tree-related risks. More recently, urban forest management practices have been placing a greater emphasis on the value of urban forests as related to their associated ecosystem services [
1,
2]. While models have been created to place a monetary value on the ecosystem services offered by urban forests [
3], municipal management efforts are still largely shaped by a desire to limit property damage and operation expenses [
4,
5]. As such, urban forest managers often focus their management efforts on roadside planting areas, sites where trees may require increased levels of care to become established and grow [
6,
7]. This management often includes the collection of street tree inventory data as a means of establishing a baseline for future monitoring and assessment [
8].
The tree and site conditions monitored in a tree inventory can vary greatly among municipalities based on management objectives. However, tree species, vitality, risk rating, geographic location (i.e., coordinates), and diameter at breast height (DBH) were shown by Östberg et al. [
9] to be the most common inventory parameters by researchers, city officials, and arborists that work within urban forestry. More detailed data collection for these parameters comes at the price of increased labor, and there is always a trade-off between allocating resources toward conducting inventories and other management efforts that could benefit from inventory data (e.g., planting, pruning, and removal). Limited municipal resources can limit inventory efforts leading to many municipalities foregoing street tree inventories altogether [
10,
11].
Recent integration and advances in the accessibility of geospatial techniques, monitoring strategies, remote sensing, and geographic information systems (GIS) have led to new and more cost-effective monitoring strategies in urban forestry [
12]. Coordinates for individual trees have become increasingly integral to urban tree inventories with the availability of global navigation satellite systems (GNSS), such as the US Global Positioning System (GPS); allowing for an easier transfer to GIS programs and integration with complementary datasets [
8]. Urban tree canopy (UTC) cover, transportation and infrastructure, and property boundaries can be interpreted and measured together [
13,
14,
15]. Techniques utilizing remotely-sensed aerial and satellite data have worked well to characterize UTC and prioritize tree planting efforts [
16,
17]. At the same time, terrestrial remote sensing techniques have even been utilized by non-professional volunteers to attain urban forest inventory data [
18]. Advanced data processing techniques for analyzing forest crowns have taken advantage of high-resolution imagery [
19], hyperspectral imagery [
20,
21], three-dimensional scanned data [
22,
23,
24], and combinations of these datasets [
25,
26,
27] to identify individual trees and collect measurements (e.g., species identification, crown dimensions, leaf area index, etc.) [
28].
Aerial remote sensing data has fundamentally changed the analysis of urban forests, but traditional inventory methods are often still preferred for detailed data collection on individual trees (e.g., street trees) [
1]. Nielsen et al. [
29] noted that studies employing aerial and satellite data generally require ground-level field data for interpretation. Even with the integration of high-quality imagery and aerial laser scanning data, Tanhuanpää et al. [
24] were only able to accurately identify 88.8% of all trees in an inventory along roadways. Similarly, Wu et al. [
30] had a detection rate of 83.6% from a purely aerial laser scanning dataset. These results suggest that individual tree identification methods from aerial remote sensing are progressing, but have some room for improvement. Moreover, measurements like DBH, ground coordinates, and species identification from aerial laser scanning and aerial imagery still typically lack the accuracy and precision that can be attained from traditional field inventories conducted on the individual tree level [
1,
29,
31].
Terrestrial remote sensing is less common but may be more applicable for managing individual trees. Terrestrial laser scanning utilizes light detection and ranging (LiDAR) to rapidly measure distances between a scanning unit and illuminated objects to create 3D representations of scanned areas. This technology has been utilized for many surveying applications to measure geologic features, ground surface, vegetation, archeological sites, and property planning for various types of infrastructure [
32,
33,
34,
35]. It has been adapted to accurately measure key forest inventory parameters (e.g., tree height, DBH, and leaf area index) [
36]. In contrast with managed forest stands that serve as the backdrop for most existing research [
36], urban environments are often compositionally more heterogeneous, and problems related to occlusion and poor scanning geometry tend to be more common. For example, Moskal and Zheng [
37] showed urban terrestrial laser scanning point cloud measurements for DBH to be correlated (
R2 = 0.68) with physical measurement, while Yao et al. [
38] showed a range of correlations (
R2 = 0.62−0.88) taken from a forest plot within a national park.
Additionally, terrestrial laser scanning data were collected from a stationary, tripod-mounted unit for many of the past studies [
39] making data collection a time-intensive process [
36]. Mobile laser scanning from a terrestrial perspective has allowed for datasets to be collected over large areas with survey-grade, centimeter-level accuracy under relatively short durations of time [
40]. Holopainen et al. [
41] showed that 79.22% of trees could be manually located in mobile laser scanning datasets with a root mean square error (RMSE) of 0.49 m when compared to terrestrial laser scanned datasets.
Applications of photogrammetry predate the use of LiDAR, but until recently, programs did not exist that could quickly produce photogrammetric models. However, this has changed with the development of inexpensive and commercially-available Structure-from-Motion (SfM) programs that use computer vision image-matching algorithms and user-interpreted ‘targets’ to generate three-dimensional photogrammetric models. These programs make it possible to create three-dimensional models with little more than a consumer-grade camera and a home desktop computer. Computer vision-assisted photogrammetry programs have been used from an aerial perspective to generate digital surface models (DSM) for applications in forestry and environmental monitoring [
42,
43]. Mapping with these models can be used to create estimates for tree height and other metrics of canopy cover [
44].
More recently, ground-based, stationary photogrammetry has also been used to measure volume, height, diameter, and surface area for individual trees [
45,
46,
47], trees positioned within forest plots [
48,
49,
50,
51,
52], and root systems [
53] (
Table 1). Even when the photogrammetric process is conducted to create models of stems along roadways [
54], the process for attaining model measurements is more intensive than traditional field measurements. Beyond this limitation, photogrammetric methods may have difficulty in assessing certain aboveground vegetation parameters (e.g., canopy width, tree height, etc.) due to shade, geometric limitations, and canopy occlusion; as Surový et al. [
49] showed, stems modeled within a close-range photogrammetry program from a forest study plot showed a stem diameter RMSE of 0.59 cm, and noted error was associated with low visibility areas.
Close-range photogrammetry has the potential to become an efficient means of inventorying and measuring street trees while avoiding the high equipment costs associated with terrestrial laser scanning that range from tens of thousands to ~$/€100,000. While others have demonstrated the potential of using fixed cameras to capture the imagery needed to model trees, there are few if any studies that have looked at mobile SfM photogrammetry techniques to measure woody vegetation in urban environments. For this study, we compared the location and stem diameter measurements derived from both georeferenced mobile SfM photogrammetry and stationary SfM data. The results from both methods were then compared to field-derived DBH and location measurements. In conducting this study, we hoped to demonstrate the potential of mobile SfM photogrammetry to inventory street trees in urban conditions.
4. Discussion
The mobile photogrammetric point cloud models in this study were suitably accurate for collecting DBH measurements, as seen with the strong correlation between the modeled and field-derived measurements (i.e.,
R2 < ~0.95) and relatively low error (RMSEs < 12.0%). These results compare favorably to field measurements, especially in light of evidence that high precision between individuals inventorying can be difficult to attain. For example, Kitahara et al. [
58] showed the difference between median field DBH measurements between groups of inexperienced surveyors could significantly improve from 2.3% to 0.9% for conifers and 1.7% to 0.9% for broad-leaved trees with more training on basic sampling procedures. Luoma et al. [
59] showed a standard deviation of 0.3 cm (1.5%) for DBH measurements taken independently by four experienced foresters in a boreal forest. In theory, physical tree measurements may necessitate high degrees of accuracy and precision, but past studies on SfM and terrestrial laser scanning data have generally concluded RMSEs of <10% are acceptable for most forest inventory purposes [
39,
45,
46,
48,
53,
60]. Given this precedent of acceptable error held by past studies on point cloud accuracy, measurements from this study compare well to previous applications of SfM photogrammetry for tree measurements (
Table 1), as the RMSE value for DBH was <11% for both urban and windbreak sites (
Table 2).
This study utilized many of the same techniques from past work on measuring trees with SfM point clouds like using high image overlap, high quantity of photos for the subjects being measured, and multiple angles for camera field of vision—but past terrestrial SfM studies have primarily operated within a local coordinate system with field-measured distances between specified reference points [
45,
46,
48,
60] or artificial targets [
48] to scale the models. Alternatively, Surový et al. [
49] used ground control points and a magnetic motion tracker points to georeference models via coordinates.
Unlike previous studies employing close-range photogrammetry for stem measurements, our mobile terrestrial photogrammetric set-up shared many features of mobile laser scanning since both do not rely on ground control points or local coordinate systems and scaling [
61]. However, the mobile set-up in this study did not utilize an inertial measurement unit. These units are almost universally available in professional mobile laser scanning systems, and work in concert with the GPS to increase positional data accuracy. An inertial measurement unit is particularly important where GPS technology may become susceptible to signal related error (i.e., multipathing, occlusion, etc.), and urban areas can be challenging environments to record accurate data, due to the prevalence of tree canopy and large structures. Although the results in this study were not attained with an IMU, adding a unit to a similar set-up would likely enhance accuracy by compensating for poor satellite reception along highly-interrupted roadways.
Mobile laser scanning data is generally taken with the expressed intent of accuracy suitable for conventional property and infrastructure surveys (i.e., at the centimeter level of error). Street tree inventories do not typically require this level of positional accuracy to identify specific trees [
10,
31]. Tree positions along streets are often measured along the direction of the street based on the beginning of the block or the property boundary [
31], hence, a similar technique was used in this study; however, this approach was modified to incorporate the use of ground objects and orthoimages. Unfortunately, the technique outlined in this paper using measurements of azimuth and distance from visible control objects to stems also has a major shortcoming, since orthoimages are often subject to georeferencing errors [
62]. Such positional errors cause general localized directional shifts of the imagery relative to the physical coordinates of the depicted landscape and objects within the imagery. This may have been evidenced in the windbreak site, where stem positions taken from the handheld-GPS and the photogrammetric model showed consistent negative bias along the X-axis (
Figure 4). The urban sites were distributed throughout the landscape of an entire city, and correspondingly, urban residuals were observed much less uniformly than the windbreak site (
Figure 5), which may have been due to the use of multiple orthoimages having translational errors in a variety of different directions. Photogrammetric stem positions data showed higher-precision via associated standard deviations (
Table 3 and
Table 4).
Past close-range photogrammetry studies on individual tree measurements have typically incorporated 360° image coverage of trees to model all sides of the stems [
45,
46,
48,
49,
53,
60]. The vehicle-mounted photogrammetry system from this study exchanges overall coverage and favorable camera geometry for expediency. Similar to the limited perspective of a vehicle-mounted laser scanner, portions of the stems not facing the roadway were not modeled by mobile photogrammetry. However, mobile terrestrial photogrammetry differs from mobile laser scanning with regard to other advantages and limitations. The equipment used in this dataset is much less expensive than a vehicle outfitted with mobile laser scanning, as the creation of a georeferenced streetscape point cloud requires little more than a survey grade GPS, a timing GPS, two DSLR cameras, a laptop computer, a vehicle with a suitable mount, and a program license for SfM software [
54]. Although the overall expense is less for mobile photogrammetry, there are still certain limitations associated with the SfM workflow. Relative to standard field measurements, photographic and geospatial data collection and model creation are still relatively work-intensive processes. The mobile photogrammetric system in this manuscript, and therefore the modeled area of the point cloud, are limited by camera field-of-vision.
Similar to mobile laser scanning data coverage, 360° reconstruction of stems was not possible; however, there was typically enough image coverage to measure the diameter from opposite sides of modeled stems (
Figure 6). This roadside perspective is also typical of mobile laser scanning data, and measurements from these point clouds must be performed manually or from programmed circle-fitting functions [
36]. Lack of perspective on irregularly shaped stems can lead to a potential source of error when measuring the diameter of a point cloud stem (
Figure 6), and this can be particularly problematic for large stems, since their associated error is proportionally greater within a dataset.
Mobile laser scanning datasets can often produce models that account for the full height of trees, while stems modeled via mobile SfM photogrammetry in this experiment were limited by camera field-of-view. Some compensation may be available by adding more cameras to the mobile set-up. However, multiple georeferenced datasets can also be merged to potentially attain multiple biometrics that would otherwise be impossible to attain from stand-alone datasets (i.e., height from terrestrial and stem-related measurements from aerial). Omasa et al. [
63] demonstrated this in an urban park with data taken from aerial and terrestrial LiDAR, but SfM point cloud dataset merges have also been performed from imagery captured independently from the ground and an unmanned aerial vehicle (UAV) [
64]. Another possibility is to utilize imagery derived from unmanned aerial vehicles since the imagery can be taken at a gradient of heights from above and facing the sides of trees. Different combinations of imagery perspective were explored by Gatziolis et al. [
65] to create SfM models for individual trees, and relatively uniform camera to tree distance was considered to be an important factor for creating SfM point clouds. Another advantage of mobile laser scanning datasets is their ability to account somewhat for woody portions of tree structure since some of the points can pass superficial leafy vegetation [
66]—however, it is much more difficult for photogrammetric models to account for underlying woody structure during leaf-on conditions [
66], or when they are occluded by nearby vegetation (
Figure 7). Liang et al. [
60], also noted that measurements from a SfM point cloud were generally inaccurate for stem measurements beyond 20 m from the camera, while unit specifications vary, mobile laser scanning datasets are generally considered accurate to a distance of 100 m before point densities significantly drop-off [
67].
Processing imagery into point clouds suitable for measurements is relatively time-intensive. Processed SfM point clouds often have certain characteristics that can limit their overall usefulness. Residual error can occur from poor interpretation of depth, which typically presents itself as ‘artifact’ points from distant sources attached to other locations of the SfM model. When creating models along roadsides, the high overlap between photos (i.e., high volume of photos throughout the entire site) was important for maintaining continuity throughout the model. Trials leading to the utilized SfM methodology in this study showed poor reconstruction and continuity along the modeled right-of-way when not enough overlap between photos was present [
54]. Post-processed dataset size is also a consideration, as SfM models tend to create much denser point clouds than mobile laser scanning. Terrestrial laser scanning data creates point clouds with much higher densities than aerial laser scanning data [
68], but the amount of data associated with terrestrial-based SfM point clouds can be extremely cumbersome for computers to process and manipulate. Currently, mobile SfM photogrammetric datasets beyond a few hundred meters of roadway will typically overburden processing for consumer grade computers unless point densities are reduced [
54]. The current solution to this problem is to process sections of roadway independently, and then to merge the point clouds after models have been created and evaluated to a user-defined level of acceptable accuracy.
Environmental conditions and camera settings can also highly impact the potential quality of the processed point cloud. Poor illumination and shadow in imagery can be difficult for computer programs to interpret [
48,
53]. Backgrounds with uniform composition, color, and texture can also be difficult for SfM software to interpret and model—particularly regarding depth (
Figure 8). Vehicle speed is also a consideration, as the vehicle used in this manuscript had to travel at a relatively slow traffic speed of ~10 km/h. Occlusion and multipathing of satellite signals to GPS receivers can also result in relatively erroneous spatial data associated with image location. Future work in areas with canopy and tall buildings would greatly benefit from the inclusion of data from an inertial measurement unit.
Close-range remote sensing of roadways may see increasing utility with the proliferation of open-source and publicly available datasets. Currently, the most encompassing of existing street-level data is Google Street View, and existing survey and monitoring efforts assessed with Street View datasets likely illustrate a precedent for future work involving three-dimensional datasets from mobile laser scanning and mobile photogrammetry. There have been examples of using Google Street View to ascertain overall greenness [
69] and to collect certain street tree parameters like species identification [
70]. However, the use of Street View has been somewhat limited by its two-dimensional point of view, as Berland and Lange [
70] also determined that it was not suitable for acquiring measurements. Google Street View imagery has already been demonstrated for cataloging and checking existing urban tree inventories [
71] and as an additional means to identify trees by species when working with aerial and satellite data [
72]. Inventories will likely become increasingly automated. Already, a machine learning approach utilizing publicly available aerial and street view imagery from Google Maps™ has been shown to detect street trees correctly and classify with >70% accuracy [
73].