Page 94 - NGTU_paper_withoutVideo

Modern Geomatics Technologies and Applications

sensing images. Therefore, to automate the GCP section process, we decided to use a similar approach and examine the ability
of clustering methods in selecting the optimal locations of GCPs, of course on UAV images. In this study, we used 3D distance
as the clustering parameter. Taking determined cluster centres as optimal locations can provide even distribution for the location
ground control points.
Clustering is a method for grouping unlabelled data [9]. This grouping is based on the data attributes or relationships, so that
the samples within each cluster have maximum similarity with each other and maximum difference Figure.1. Number and
distribution of ground control points.

Fig. 1. Number and distribution of ground control points

Fig. 2. Clustering process

With the samples of other groups. An example of clustering is shown in Figure 2 input to a clustering technique is a set of
initial points that are grouped according to different mechanisms. The number of clusters can be defined either automatically or
manually.
So far, several clustering algorithms have been presented. In short, they can be divided into six main groups: Petitional,
Hierarchical, Model, Fuzzy, Graphic and Modern ([10], [11], [12]). The ability of some samples selected by each of the groups
listed will be examined in this paper. Selected methods are among the most commonly used methods in the relevant group whose
functions was available for implementation and testing. In the remainder of this article, the second part describes how the
evaluations should be carried out. Results are analysed in Section 3 and, finally, discussions are concluded and suggestions for
future research are made in Section 4.

2. EVALUATION METHODLOGY

The main idea of the evaluations is to see how accurate a photogrammetric model is, if the control point locations are
considered in the locations suggested by each clustering technique. Several clustering techniques were considered (Table 1).
Figures 3 shows how the evaluations are carried out in this study. The tests are carried out on images taken in Ardebil area, Iran.
The area was imaged using a Phantom4 Pro. At first, points of the study area whose location can be measured are defined as the
initial points. These points are, in fact, a representation of the GCPs in the test area. These points are clustered together. Cluster
centres are considered as proposed positions for the measurement of ground control points. It should be noted that, since cluster
centres are different in each method, measuring them would be very time-consuming. A dense network of ground control points
was therefore established in this research and the closest ground control point to the centre of the cluster was considered to be
the measured ground control point. To investigate the capability of any clustering method, UAV images were oriented using the
ground control points determined by the clustering method. This way, the accuracy of the resulting model is considered as the
precision of the clustering method. In the following, the steps carried out are described in more detail.

2