Page 95 - NGTU_paper_withoutVideo

Modern Geomatics Technologies and Applications

Fig. 3. Evaluation methodology

2.1 Obtaining UAV images and forming control points network
In order to carry out the evaluations, a 400-by-400m area was considered in the district area of Ardebil in the northwest of
Iran. It was imaged using a Phantom 4 Pro drone. The flight height was 80 meters and the overlap between the images was 90%.
The focal length of the camera is 8.8mm and the sensor is 1 inch, and the resolution of the images is 20 megapixels. In this area,
81 ground control points with a 50m interval were measured using a Global Positioning System (GPS). The nominal accuracy
of the measured points was one centimetre. Of these, 72 points were considered as the control point network and 9 points as
check points. Figure 4 shows the position of ground control and check points.

2.2 Formation of initial points and the measurement of their coordinates
As already mentioned, the initial points are the points that are clustered and their centre is considered as a ground control
point. In fact, they constitute the ground points of the area of interest. In this project, these points are all those whose ground
coordinates can be measured. In this research, the location of these points is measured manually on Google Earth. Figure 5 shows
the network of initial points measured manually on Google Earth. As the accuracy of this data affects the clustering results, their
accuracy should be evaluated. In a previous research, the precision of the extracted points on Google Earth was reported as 3.63
and 1.73 m, in plan and height, respectively [13]. However, as we know, this accuracy depends on the zone and the images used
to create the corresponding Google Earth map. Therefore, it is necessary to examine the precision of the initial points in the area
where clustering algorithms are applied to them. For this, the coordinates of 8 points in the area where evaluations will be
performed were specified and their coordinates were measured on Google Earth. Then, the coordinates were compared with the
corresponding values measured by GPS. Table 2 shows the results of this comparison. As can be seen, the planimetric error of
the points is about 2 cm but the elevation error on average is about twelve meters. As the height range between the points is not
much, a twelve-meter error for Z is significant. However, as can be seen, this error is systematic and is almost the same for all points. Therefore,
it will have little effect on the cluster centre selection process, which is done by comparing the distance of a centre with its surrounding control
points.

2.3 Determine the number of clusters and clustering the primary grid points
Once the initial network of the points is formed, it should be clustered. In this regard, it is first necessary to determine the
number of clusters. As mentioned, in this paper, the centres of the clusters are considered as control points. So, it is necessary to
determine the number of control points required to carry out the UAV photogrammetry project. As we know, to perform the absolute
orientation, at least five points are required.
In practice, the distance between the control points is taken one or two times the flying height. Given that the images are
taken at an altitude of 80 meters, the network of initial points was segmented into 5 (minimum number of required points), 9
(twice the flight height) and 25 (equal to the flight height) clusters. To perform the clustering, XYZ coordinates need to be used.
To this end, we used the 3D distance between the initial points as the clustering parameter. For each cluster, a point from the
GCP network whose distance to the cluster centre was minimum was considered as the selected control point. The clustering
techniques evaluated in this paper are shown in Table 1.

3