Study area
The site (50° 04’ N, 13° 59’ E) is located 30 km W from the city of Prague, Czech Republic. The elevation in the study area of 15 ha ranges from 420 to 435 m above mean sea level with mostly S or SW aspect. It is a part of a protected landscape Křivoklátsko (681 sq. km) included among the UNESCO Biosphere reserves in 1977. The study site represents a typical Central European spruce-pine forest managed for timber production (standard silvicultural treatments are carried out). Two different forest compositions can be found on the site: (i) full-grown 80–100 years old coniferous forests consisting of pine (40 %), spruce (40 %), and larch (20 %), mean crown size 7.1 ± 1.4 m; and (ii) 20–40 years old coniferous forests consisting of spruce (60 %), pine (20 %) and larch (20 %), mean crown size 4.5 ± 0.9 m, see Fig. 1.
Imagery acquisition
In total, nine flights with four sensors mounted on three different UAVs were performed in the study area (see Table 1 for details). The imagery was collected during late winter, on February 26, 2019, between 10 am and 3 pm. The flight conditions were convenient; mostly cloudy sky (6/8) with a temperature around 12°C and NW wind of 2–5 m.s− 1. A total number of nine Ground Control Points (GCP) surveyed with Leica 1200 GNSS in RTK mode were placed throughout the study site.
The three UAVs included: (a) a lightweight fixed-wing UAV Disco Pro Ag (Parrot SA, France), which is a ready-to-deploy solution for agricultural and forestry applications with a maximum take-off weight (MTOW) of 0.94 kg mounted with Sequoia camera; (b) Phantom 4 Pro (DJI, China), which is probably the most popular lightweight (MTOW of 1.39 kg) universal commercial UAV, mounted with integrated camera and (c) Matrice 210 (DJI, China) representing a professional adjustable enterprise solution with MTOW of 6.6 kg with Zenmuse X5S FC6520 camera, respectively. In addition, (d) a fourth UAS (Unmanned Aerial System, i.e. ready-to-fly solution including all necessary components) was created by mounting RedEdge-M camera on the Phantom 4 Pro platform. Except for the fixed-wing UAV where the producer regulates the elevation level, the flights were performed at 100, 150, and 200 m above ground level (AGL). Flight missions were performed using (i) the DJI Ground Station Pro application for both Phantom 4 and Matrice 210, and (ii) Pix4Dcapture application for the Parrot Disco Pro Ag. The flights were conducted using perpendicular flight lines with 80% forward (longitudinal) overlap and 70% side (lateral) overlaps. The UAVs followed predefined flight plans across the study sites. The sensors triggering option was set to Overlaps (one image acquisition per waypoint) or, where not possible, Time-lapse (one shutter shot per fixed time). To simulate the behaviour of a user without in-depth knowledge of RS techniques, we evaluated the performance of the systems in the default mode; i.e. no adjustments to the vendor-preset parameters of image acquisition (shutter speed/aperture preference, ISO, etc.) were made.
Table 1
Specifications of used UAV-mounted cameras.
|
Zenmuse X5S
FC6520
|
Phantom 4
FC6310
|
RedEdge-M
|
Sequoia
|
Manufacturer
|
SZ DJI Technology Co., Ltd.
|
MicaSense Inc.
|
Parrot SA
|
Mounted on
|
Matrice 210
|
Phantom 4
|
Phantom 4
|
Disco Pro Ag
|
Sensor
|
4/3-inch CMOS
|
1-inch CMOS
|
|
|
Resolution (MPx)
|
20.8
|
19.8
|
1.2
|
1.2
|
FOV (°)
|
72
|
84
|
46
|
49
|
F-stop*
|
3.5/5/3.5
|
3.5/4.5/4
|
2.8 (fixed)
|
2.2 (fixed)
|
Shutter*
|
120/240/160
|
80/200/100
|
270/1000/500
|
310/730/320
|
ISO*
|
100 (fixed)
|
100 (fixed)
|
100/800/800
|
100 (fixed)
|
35 mm equivalent focal length
|
30
|
24
|
39
|
29
|
Spectral bands
|
RGB
broadband
|
RGB
broadband
|
B, G, R, RE, NIR
narrowband
|
G, R, RE, NIR
narrowband
|
Spectral range (nm)
|
n/a
|
n/a
|
455–727
|
530–810
|
Image size
|
5280×3956
|
5472×3628
|
1280×960
|
1280×960
|
Image format
|
JPG
|
JPG
|
TIF
|
TIF
|
Dynamic range
per band (bit)
|
8
|
8
|
16
|
16
|
Triggering
|
Overlaps
|
Overlaps
|
Time-lapse
|
Overlaps
|
Radiometric calibration**
|
n/a
|
n/a
|
Panel + DLS
|
Panel + DLS
|
Weight (g)
|
461
|
1388***
|
170
|
72
|
Price (EUR)
|
2199
|
1699***
|
4200
|
3850
|
* Capture settings min/max/median values. ** DLS stands for Down-welling Light Sensor. *** The sensor cannot be removed from the body; thus, price and weight include the aircraft. |
Image alignment & surface reconstruction
In total, 2323 images were captured across the study site with a total disk space occupancy of approx. 31 GB. The Agisoft Metashape Professional (version 1.5.5, Agisoft LLC, Russia) image-matching software was used to generate point clouds and reconstruct 3D surface (Verhoeven, 2011). Metashape uses metadata related to the band information from EXIF to load image description including the coordinates taken from the onboard GNSS units. As the first step, we loaded the images and estimated image quality. Images taken during the take-off, landing and taxiing as well as those the quality of which (automatically evaluated by Agisoft) was less than 0.5, were excluded from further processing (Puliti et al., 2015).
After determining image orientations, i.e., geometrical processing (Snavely et al., 2008), the sparse point clouds were checked for outliers and, subsequently, densified using surveyed GCPs. Following the Metashape manual, a digital surface model (DSM) was constructed using dense point clouds and an orthomosaic was built. Ground point classification of the dense point clouds was performed to enable construction of a digital terrain model (DTM). This process was done using a trial and error approach in order to receive the best possible terrain for all of the inputs (Klápště et al., 2020).
Deriving tree variables and statistical analysis
The evaluated tree parameters (the number of detected trees and crown diameter) were derived in R (version 3.4.3, R Core Team, Austria). First, we subtracted the terrain models from the surface models to gain normalised heights, i.e., Canopy Height Models (CHM). Subsequently, CHMs from all 24 datasets (3 altitude levels, 2 sites and 4 cameras) were processed using the same workflow in the lidR package (Roussel and Auty, 2018; Roussel et al., 2020). The workflow included (i) identification of individual trees; and (ii) delimitation of the tree crown. Detection of individual trees was based on local maxima filtering using focal statistics (Ke and Quackenbush, 2011; Vauhkonen et al., 2012) while crown delineation was performed via watershed-based object detection (Panagiotidis et al., 2017; Surový et al., 2018). Subsequently, the crown diameter was calculated using automatic methodological workflow consisting of (i) approximation of individual detected crowns using circles with areas corresponding to those of the tree crown polygons; (ii) circle diameter calculation. Thus, the final output contained information about the total number of trees, where information about the location (coordinates), and crown diameter (m) of each individual tree/shrub was recorded.
Reference values for the number and diameters of the tree crowns were obtained through manual detection from the most detailed orthomosaic (Phantom 4 Pro with default camera, flight altitude 150 m). Tree crown diameters were manually measured in north-south and west-east directions also under same orthomosaic in ArcGIS software, version 10.7.1 (ESRI, Redlands, CA, USA). Due to the slight differences in the treetop position in the CHMs from individual UASs and the reference, the treetops on individual orthomosaics were automatically overlaid (Near function in ArcGIS) before further processing. Multiple treetops within a radius of 2 m were considered to represent a single tree (Nevalainen et al., 2017). In addition, the results were visually inspected.
The accuracy of individual tree detection, and thus the proportion of correctly detected trees in both study sites, was evaluated and expressed as the total accuracy with 95% confidence intervals (Table 2). In total, 100 trees were randomly selected within each study site. The normality of the distribution of individual treetop areas was tested by the Shapiro–Wilk test with outliers both included and excluded. Depending on the results, we applied Student’s t-test or Wilcoxon signed-rank test, respectively (Table 3), for detecting significant differences between the automatically detected treetop areas and the reference values, and, thus, for comparing the performance of the models. The accuracy of detected crowns was evaluated and expressed via the Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE); the model performance is expressed as 1-MAPE (Table 3). Statistical analyses were performed in STATISTICA software, version 13.4 (TIBCO Software Inc., Palo Alto, CA, USA). The workflow of the study is in Fig. 2.