3.1. RT-Based Acquisition of Channel Characteristics
Ray tracing (RT) is an accurate channel parameter calculation method based on field superposition principle and uniform theory of diffraction theory. The shooting and bouncing ray/image (SBR/IM) method is a symmetry-based RT method and it is applied to obtain channel characteristics in this paper due to its high accuracy and low computational complexity. The SBR/IM method includes four steps: scatterer reconstruction, ray decomposition, ray tracking and channel characteristics calculation.
Due to the high complexity of the realistic scenario, a scatter reconstruction method is used in this paper to reduce the complexity and computational time of the SBR/IM method. The reconstructed scenario is composed of reconstructed the surface of the terrain and buildings. Firstly, the flat surface of terrain can be reconstructed by regular triangle faces and the non-flat surface can be approximated reconstructed by irregular triangle facets. Then, the surface of buildings can also be reconstructed by regular triangle facets due to they are assumed to be rectangles. Finally, material of surfaces will be configured after all scatterer shapes are reconstructed.
The ray decomposition method is as shown in
Figure 3a, a regular icosahedron is placed inside a wave-front ball. The wave-front ball is divided to generate the ray tubes, where each tube is represented by a ray from the center of the vertex. Considering that the section of the ray tube increases with the increasing distance, which would affect the accuracy of ray tracking. Therefore, the wavefront division of the regular dodecahedron is necessary as shown in
Figure 3b. Each equilateral triangle is uniformly divided into
N wave fronts. The expression of
N is defined as
, where
n is the number of ray tube splits. In
Figure 3b, the number of ray tube splits is set as
.
For the ray tracking, the reflection, scattering, and diffraction propagations are tracked by using geometric optics theory, where the intersection operation is a key step. The intersection operation is used to determine whether the ray intersects with the triangle facets of scatterers, and then to determine the reflection or scattering points. It is also necessary to determine whether diffraction occurs. Taking the receiver as a receiving sphere, the signal is considered being able to arrive at the receiver if the ray intersects with the sphere.
The path to the receiver can be divided into the direct ray or reflected ray. For the direct ray, the electric intensity and power gain can be calculated by
where
is the electric intensity of the initial ray tube,
is the wave number, and
is the distance between the transmitter (TX) and receiver (RX),
is the power of the direct ray, and
are
antenna of the receiver and transmitter, respectively.
The electric intensity of the reflected ray can be calculated as
where
is the electric intensity of the initial ray tube,
and
are union vector of reflection and refraction coefficients for the entire ray path, respectively,
and
are the phase shift and power attenuation of the ray from the reference position, respectively,
is diffusion factor,
is the power of the reflected ray,
and
are antenna gain of the receiver and transmitter.
In this paper, the inter-path delays and angles are calculated in a deterministic method. Denote the location of signal source as
and the location vector of RX as
. The adjacent reflection points are set as the centroid and the position vector is denoted as
. The delay of m-th ray is defined as adding the intra-path delay offset
to the mean ray delay
as
where
can be expressed as
where
is the speed of light.
The azimuth AOA
and elevation AOA
of mth ray can be obtained as the same way. Furthermore, the mean ray angle
and
can be expressed as (
10) and (
11), respectively.
where
,
and
represent the
x,
y,
z component, respectively. After the channel parameter are obtained, we can further calculate the channel characteristics.
The RMS-DS is used to describe the channel dispersion phenomenon in the delay domain which is caused by the small-scale fading of multipath effect. The definition of RMS-DS can be expressed as
where
and
are the delay and power of each propagation path, respectively,
represents the received power,
L represents the number of paths, and
represents the mean delay which can be further expressed as
The
K factor, also known as the Rice factor, is the power ratio of the dominating multipath component to the summation of other multipath components. The dominating multipath component usually is a direct path in A2G communication. The
K factor can be expressed as:
where
is the power of the dominating multipath component.
In the angle domain, the multipath effect will cause angle spread. The AS is used to characterize the magnitude of the angular dispersion, which can be expressed as
where
can be expressed as:
It should be mentioned that the AS of azimuth angle of arrival (AAOA), azimuth angle of departure (AAOD), elevation angle of arrival (EAOA), and elevation angle of departure (EAOD), respectively.
3.2. Training Dataset of Height-Dependent Channel Characteristics
Different from the scenario identification of terrestrial communications, the influence of UAV height should be considered in the scenario identification of A2G communications. In this section, the channel characteristics at different height are obtained and analyzed by using the calculation method in
Section 3.1 under five typical A2G communication scenarios including over-sea, suburban and urban scenarios. The urban scenarios are further divided into urban, dense urban and high-rise urban.
The simulation carrier frequency is set as 2.6 GHz, and the transmitting power is set as 20 dBm. Both the transmitters and receivers are equipped with a half-wave dipole antenna with vertical polarization. We set ten groups of transmitters and receivers in each scenario. In each group, the transmitter is placed at the height of 1.7 m, and the receivers are placed at the height between 10 m to 210 m with 2 m intervals. Moreover, we place three receivers at each height, so each scenario has 3000 receivers.
To obtain the needed digital map, the standardized scenario models recommended by International Telecommunication Union-Radiocommunication Sector (ITU-R) are adopted in this paper [
32]. The scenario models are related with three parameters
, which are defined as follows
indicates the proportion of the building area to the total area;
indicates the average number of buildings per unit area (buildings/km);
indicates the height of the building according to the Rayleigh distribution, where h can be calculated by
The detailed simulation parameters and the parameters of scenario models for the suburban and three urban scenarios are summarized in
Table 1. Note that there are no buildings in the over-sea scenario, we just reconstruct several ships for simplicity. The digital maps of five reconstructed scenarios are shown in
Figure 4, and the average heights of the scatterers in each scenario are 4 m, 9.44 m, 18.83 m, 24.31 m and 64.43 m. The size of all digital maps is 1 km × 1 km. The material of ship is defined as metal, and the material of buildings is concrete, and the land is defined as dry ground.
Based on the calculation method in
Section 3.1, the channel characteristics are calculated at different height. For simplicity, only the RMS-DS and AS of AAOA of one group are shown in
Figure 5 and
Figure 6, respectively. The
y-axis of
Figure 5 and
Figure 6 are the height of the UAV, and there are 30 data points per height. The
x-axis of
Figure 5 and
Figure 6 are the index of each data point, ranging in size from 1 to 30. Furthermore, there are significant changes around a certain height, which are marked with red lines in the figure, respectively. This is due to the difference in the average height of scatters of scenarios, which leads to different channel characteristics distribution with height. Furthermore, the exact mean values of different channel characteristics under different scenarios are presented in
Table 2. It can be found that the channel characteristics, i.e., RMS-DS,
K factor, ASs vary a lot under different scenarios and at different height, which makes it possible to identify the scenarios based on the channel characteristics.
Based on the above discussion, in this paper we use the channel characteristics, i.e., RMS-DS, K factor, ASs and the height of the UAV as the identification features. The dataset of i-th scenario is denoted as , where denotes the scenario label and L is the number of scenarios.
3.3. Height-Integrated Scenario Identification Method
The proposed height-integrated scenario identification method is shown in
Figure 7. It includes three steps, i.e., dataset acquisition and preprocessing, identification model training, and height-integrated model feedback. The details are shown as follows.
Step 1: Data acquisition and preprocessing.
Based on the calculation method in
Section 3.1 and
Section 3.2, the datasets of height-dependent channel characteristics are obtained. Note that the weight of different channel characteristic is different. In order to achieve better identification performance, it is required to preprocess the dataset before training. In this paper, we normalize the data by using z-scores and it can be expressed as
where
and
are the mean and variance of the input
j-th dimension data, respectively.
Furthermore, dimensionality reduction for high-dimensional input data can prevent the method from slipping into the local optimum and improve the training performance. Therefore, the principal component analysis (PCA) is adopted for dimensionality reduction. The core idea of PCA is to use orthogonal transformation to replace a set of potentially related variables with a set of linearly unrelated principal components [
33]. Low-dimensional principal components can be generated by reasonable selecting eigenvalues. The Gaussian Kernel function in Formula (
3) is employed as the core of PCA in this paper.
Step 2: Identification model training.
Then the datasets are divided into two parts, i.e., training data set and testing data set in the proportion of 3:2 as shown in
Table 3. Although SVM is a binary classifier, we can use a decomposition methods of multi-class SVM by reconstructing a multi-class classifier from binary SVM-based classifier. For
j-th binary SVM classification, it takes the scenario with
j-th label as positive class and the rest of others as negative class, where
. The results of new samples are determined by combing the labels which is predicted from all the SVM classifiers. Assuming that the multiple binary SVM classifiers are
, the final identification result of a sample
x is determined by
where
s is scenario label and “arg max” is to find the scenario label with the highest probability of the multiple binary SVM classifiers.
Step 3: Height-integrated model feedback.
After obtaining the trained scenario identification model, the new channel characteristics can be input into the model for scenario identification. If the posterior probability f is greater than the threshold , the scenario label will be output directly, otherwise the height of the UAV is changed slightly to get a new dataset and repeat the identification procedure. It should be mentioned that whether the height of the UAV is raised or lowered depends on the first scenario label. When the posterior probability is greater than the threshold or the height of the UAV exceeds the height limitation, the model outputs the predicted label of the scenario. In the scenario identification method, the threshold should be properly determined.