*4.2. Verification of TWPE*

Propagation losses from Shanghai–Wuxi using the same environment settings and radar parameters as intercity links of Hangzhou–Wuxi are shown in Figure 17. Different LC types were considered to evaluate the PL distribution on the S–W link using the TWPE model, as shown in Figure 17a,b. The AREPS result with the IGBP LC type is shown in Figure 17c. The range-dependent PLs at a ducting layer of 60 m on the S–W link of different LCs using TWPE and AREPS are compared in Figure 17d.

The propagation distance from Shanghai to Wuxi is about 115 km. Moreover, the terrain relief of the intercity link between Shanghai to Wuxi was kept below 10 m, which fits the application condition of the paraxial approximation of PWE. Based on the plain cover characteristics of East China, the land types of terrain on this intercity link are classified into grassland and water bodies at first. We conducted WAPE considering terrain relief, tropospheric turbulence, and dielectric permittivity on the first assumption and the land types of IGBP from MODIS data.

From Figure 17a,b, we used the IGBP cover type form MODIS with a resolution of 250 m, the PL distribution basically remains at the same level as the conducted form TWPE with LC type using grassland and water bodies. Figure 17d shows the PL distribution at the transmitting height of 60 m in the trapping layer of the surface-based duct. The PL at a height of 60 m using the LC type of IGBP from MODIS between 60 and 105 km is larger than that of using the grassland and water bodies as terrain features. On the remaining links, the PL using the two LC classifications are reversed or on the same level. AREPS uses the RPO model to divide the long-distance ducting propagation into four models to improve the calculation speed and facilitate engineering applications. There was not much to consider on the tropospheric scatter and terrain scatter in AREPS. Therefore, the PL trend was similar but AREPS result was 20 dB lower than TWPE result.

Given that the industry can achieve GPS receivers with tracking sensitivity below −160 dBm, at the height of the surface-based duct within the receiver, the realistic LC classification could be an essential environmental condition to evaluate whether CCI would happen on this intercity link. The possibility of CCI could be an essential decision support of 5G base station deployment.

Propagation loss from Jiaxing to Wuxi using the same environment settings and radar parameters, the same as in the previous section, are shown in Figure 18. Different LC types were considered to evaluate the PL distribution on the J–W link using TWPE model, as shown in Figure 18a,b. The AREPS result with the IGBP LC type is shown in Figure 18c. The range-dependent PLs at a ducting layer of 60 m on the J–W link of different LCs using TWPE and AREPS are compared in Figure 18d.

The distance of the intercity link from Jiaxing to Wuxi was about 93 km. Moreover, the terrain relief was around 20 m, besides a peak, which had a maximum height of 80 m. The peak was not high enough to cut off ducting and diffraction, and the radio wave elevation angle fit the maximum limit of TWPE.

From Figure 18b, we can see that when we considered the LC types more specifically, the PL on the intercity link Jiaxing–Wuxi increased slightly. Due to the limitation of the

RPO model in AREPS using a range-independent RO model in the range-dependent environment, in extreme cases this can cause discontinuities along the RO/PE boundary [28]. This discontinuity occurred in the peak of the intercity link of Jiaxing–Wuxi.

**Figure 17.** The propagation loss with tropospheric scattering, IGBP LC, and DEM considered on Shanghai–Wuxi. (**a**) PL from Shanghai–Wuxi using TPWE and IGBP LC. (**b**) PL from Shanghai–Wuxi using TPWE, WAT, and GRA LC. (**c**) PL from Shanghai–Wuxi using AREPS and IGBP LC. (**d**) PL at the duct-trapping layer of 60 m of different LCs.

The two classifications of LC types do not give obvious differences probably due to the classifications of grasslands and waterbodies being close to those of the real LC type distributions on the intercity link. It can be seen from Figure 9d that the real LC data on the Jiaxing–Wuxi link extracted by the IGBP LC classification also consisted of most WAT—17 in IGBP LC1—and several other kinds of vegetation categories. Moreover, the classifications from MODIS data just gave more specific resolutions on the LC types. Additionally, when we used the LC type classification in the resolution of 250 m from MODIS, the fast fading was much more obvious.

### *4.3. Effects of Different Land Covers*

The previous section shows (assuming that the LC is grassland and water bodies) the PL trend and numerical range of PL at the antenna height layer when mixed ducts and tropospheric turbulence occur on the plain area. The accuracy of the results of the TWPE model with tropospheric turbulence was verified. The PL situation covered by different LCs will be discussed in this section.

In this section, three types of LCs, barren or sparsely vegetated (BSV), urban and built-up (URB), and savannas (SAV), were selected, respectively. The over-the-horizon propagation was studied on the Shanghai–Wuxi and Jiaxing–Wuxi links considering the joint effect of tropospheric turbulence and the duct. Figure 19 shows the PL distribution on Shanghai–Wuxi link. The parameters are the same as in the previous section, except that the LC dielectric parameters use the values corresponding to Table 3, respectively.

**Figure 18.** The Propagation loss with tropospheric scattering, IGBP LC, and DEM on Jiaxing–Wuxi. (**a**) PL from Jiaxing–Wuxi using TPWE and IGBP LC. (**b**) PL from Jiaxing–Wuxi using TPWE, WAT, and GRA LC. (**c**) PL from Jiaxing–Wuxi using AREPS and IGBP LC. (**d**) PL at the duct-trapping layer of 60 m of different LCs.

We used MSE to compare the PLs on Shanghai–Wuxi under three hypothetical LC types with the PLs on the S–W link under real IGBP LCs; see Figure 17a. The results are shown in Figure 20.

**Figure 19.** The PL using the improved TWPE with (**a**) BSV, (**b**) URB, (**c**) SAV LC, on Shanghai–Wuxi. (**a**) PL on S–W with BSV LC. (**b**) PL on S–W with URB LC. (**c**) PL on S–W with SAV LC.

**Figure 20.** The MSE of PL using the improved TWPE between IGBP LC and different LCs, on Shanghai–Wuxi. (**a**) BSV-IGBP MSE of PL on S–W. (**b**) URB-IGBP MSE of PL on S–W. (**c**) SAV-IGBP MSE of PL on S–W.

To clearly compare the influences of the PLs of the transmission signal in the ducting layer of the Shanghai–Wuxi link under the backgrounds of several different LC types, we compared the PLs of the link at a height of 60 m in the ducting layer, as shown in Figure 21.

According to the impedance boundary conditions applicable to irregular terrain, it is assumed that the LCs are BSV/URB/SAV. Figure 19 shows the spatial distribution of the PL on Shanghai–Wuxi link when the three types of hypothetical single LCs are distributed. Figure 20 shows the spatial distribution of the mean square error (MSE) obtained by assuming different LC types and the real LC classified by IGBP on the Shanghai–Wuxi link. It can be seen that different LCs have a grea<sup>t</sup> influence on the PL accuracy, particularly at the antenna propagation height near the surface within the duct trapping layer. From the PL distribution of the antenna height layer shown in Figure 21, it can be seen that on the Shanghai–Wuxi link, the difference between the PL coverage of different LCs is mainly reflected in the front section and the rear section. The 20–40 km and 70–100 km sections both show that the IGBP LC results are significantly larger than the PL covered by a single LC type. The research in [29] shows that terrain fluctuation has a grea<sup>t</sup> influence on the PL distribution of the signal's over-the-horizon propagation. The results in this section show that the changes of different LCs may increase the PL, especially in the section with sudden and large terrain fluctuations, the actual situation may be a sudden change of LC, so it is more necessary to consider the coverage of real terrain.

Figure 22 shows the over-the-horizon PL distribution of different LCs on the Shanghai– Wuxi link when the receiving antenna height is 40 m. Different from the results obtained with the receiving antenna height of 60 m in Figure 21, there is a significant difference on the range-dependent PL in various LCs at 40 km; that is, in the middle of the link. However, under the backgrounds of the distribution of real LCs, the calculated PL still exceeds the sensitivity threshold of the receiving antenna at a height of 40 m at range of 45 km, blocking the long-distance propagation of the signal. Therefore, the possibility of users receiving remote interference on the road section beyond the 45 km Shanghai–Wuxi link is small. If the road section undergoes urban renovation (biased towards the dark cyan line in URB,

Figure 22), or vegetation green development (biased towards the dark yellow line in SAV, Figure 22), the propagation distances of the communication signal could vary according to the specific meteorological conditions and changes in the distribution of LC features. Therefore, according to the changes of the annual live updates of the LC types, the more finely the classification of LCs that may develop in the future, and the dielectric parameter values corresponding to the more accurate LC types, the reception range of communication signals could be estimated through a more accurate model.

**Figure 21.** The propagation loss using the improved TWPE at a height of 60 m using different land covers, on Shanghai–Wuxi. (**a**) PL on S–W at different LCs at a antenna height of 60 m. (**b**) PL on S–W at a range of 20–40 km at different LCs. (**c**) PL on S–W at a range of 40–70 km at different LCs. (**d**) PL on S–W at a range of 70–115 km at different LCs.

Figure 23 shows the PLs of over-the-horizon propagation on Jiaxing–Wuxi with three hypothetical LCs considering the joint effect of the tropospheric turbulence and duct. The parameters are the same as in the previous section.

We used MSE to compare the PLs under three hypothetical LC types with the PLs under real IGBP LCs on the Jiaxing–Wuxi link, as shown in Figure 18a. The results are shown in Figure 24.

On the path link of Jiaxing–Wuxi, the propagation distance was 93 km, but there was a high terrain fluctuation at 70 km. Figure 24 shows the MSE between the parameter conditions of the assumed single LC feature distribution and the real IGBP feature distribution. As can be seen from Figure 24, the MSE between the PL calculated on the assumed single LC type parameters and the PL calculated on the real LC parameters classified by IGBP started to grow from the top of the duct, which was also the energy leaking area at a distance of 70 km from Jiaxing. It was obvious that the MSE of PL at the antenna height was the smallest when the assumed LC was SAV, which was closely related to the distribution of LCs on the intercity link. According to IGBP terrain classification in Figure 9d, Jiaxing–Wuxi was mostly covered by water bodies, and also contained many types of landform switching.

In order to clearly compare the influence of the PLs of the transmission signal in the ducting layer on Jiaxing–Wuxi link under the backgrounds of several different LC types, we compared the PLs of the link at a height of 60 m in the ducting layer, as shown in Figure 25.

Figure 25 shows the range-dependent PL at the antenna height of 60 m on different LC types. As shown in Figure 21c, the PL increased sharply, which was around 200 dB. Therefore, the difference of the PL calculated using TWPE between a single LC and on IGBP LC was not obvious at 0–20 km. From the 20–40 km PL on the J–W of different LCs shown in Figure 25a, the over-the-horizon PL on IGBP LC was close to that on BSV LC, but higher than the results of the URB or SAV LC. The PL of IGBP LC was slightly smaller in the 0–20 km section of Jiaxing–Wuxi. The PL distribution of different LCs at 0–40 km on the antenna height layer was similar to that of the Shanghai–Wuxi link, which is helpful for the deployment of short-distance 5G base stations. The middle section at a range of 40–70 km was mainly affected by terrain fluctuations and ducts, and the PL was kept within a certain range. The LC of this section can be roughly assumed using the simplified dielectric permittivity to simulate the PL distribution and could be performed to save time and costs. From the PL distribution of 70–93 km, it can be inferred that when the terrain fluctuation exceeds 80 m, the PL might exceed the minimum resolution value of the GPS receiver, so that the possibility of remote interference on the link can be ignored. Moreover, when long-distance communication is required, the corresponding 5G relay base station deployment can be reasonably arranged with reference to this result.

**Figure 22.** The propagation loss using the improved TWPE at a height of 40 m using different land covers, on Shanghai–Wuxi. (**a**) PL on S–W at different LCs at an antenna height of 40 m. (**b**) PL on S–W at a range of 20–40 km at different LCs. (**c**) PL on S–W at a range of 40–70 km at different LCs. (**d**) PL on S–W at a range of 70–115 km at different LCs.

**Figure 23.** The MSE of PL using the improved TWPE between IGBP LC and different LCs on Shanghai–Wuxi. (**a**) PL on J–W with BSV LC. (**b**) PL on J–W with URB LC. (**c**) PL on J–W with SAV LC.

**Figure 24.** The MSE of PL using the improved TWPE between IGBP LC and different LCs on Shanghai–Wuxi. (**<sup>a</sup>**–**<sup>c</sup>**) BSV-IGBP MSE of PL on J–W.

**Figure 25.** The Propagation loss using the improved TWPE at a height of 60 m using different land covers, on Jiaxing–Wuxi. (**a**) PL on J–W at different LCs at at antenna height of 60 m. (**b**) PL on J–W at range of 0–40 km at different LCs. (**c**) PL on J–W at a range of 40–70 km at different LCs. (**d**) PL on J–W at a range of 70–93 km at different LCs.

Figure 26 shows the range-dependent PL distribution at the height of the receiving antenna when the receiving antenna is 40 m on the Jiaxing–Wuxi link. At 40–70 km in the middle section of the link, the PLs simulated by different types of LC begin to show obvious differences. Compared with Figure 25 when the receiving antenna height is 60 m, the overall PL is reduced; that is, when the receiving antenna height is lower, the propagation distance of the signal increases. Within the range of signal-noise-ratio (SNO) that the antenna can distinguish, when the receiving antenna is 40 m, the signal can be transmitted to about 45 km. In addition, if the LC on the link develops in a unified manner, such as highway development, greening, desertification, etc.; that is, if the LC on the link does not change much, as shown in Figure 23b, the overall PL is reduced, and the propagation distance may increase by 10–20 km.

**Figure 26.** The propagation loss using the improved TWPE at a height of 40 m using different land covers, on Jiaxing–Wuxi. (**a**) PL on J–W at different LCs at an antenna height of 40 m. (**b**) PL on J–W at a range of 0–40 km at different LCs. (**c**) PL on J–W at a range of 40–70 km at different LCs. (**d**) PL on J–W at a range of 70–93 km at different LCs.

### *4.4. Analysis of Remote Interference between Different Intercity Links*

Comparing Figures 22–25, the Shanghai–Wuxi link is relatively close to the Jiaxing– Wuxi link, and base stations on one of the links from 0 to 45 km might receive an uplink signal from another city, resulting in remote signal interference. Therefore, this section compares the PL of the four links with the receiving antenna heights of 60 m and 40 m, respectively (the heights are both within the trapping layer of the surface-based duct assumed in this paper), as shown in Figure 27. The probability of possible remote interference of inter-city signals was analyzed.

As shown in Figure 27, due to large terrain fluctuations, the propagation distance on Hangzhou–Wuxi and Nanjing–Wuxi links were about 15 and 40 km, respectively. From the city orientation shown in Figure 8, it can be seen that Hangzhou, Jiaxing, and Shanghai are all located within a radius of about 100 km southeast of Wuxi, and signal serialization may occur between the links from each city to Wuxi. As shown in Figure 9d, the Jiaxing–Wuxi link had a sudden increase in PL at different receiving antenna heights due to relatively

large terrain fluctuations and switching of LCs. In the 0–40 km stage, the signals transmitted from Hangzhou, Jiaxing, and Shanghai were relatively stable under the effect of surfacebased duct and tropospheric turbulence. Therefore, within a certain radius close to Wuxi, remote interference may occur among the signals of the three urban links, where a useful signal from one city may become an interfering signal received by another user. Based on the analysis of this phenomenon, it can be considered that when the base stations are deployed among relatively close cities, combined with a variety of possible ducting conditions, the tropospheric turbulence effect, and the annual update of the distribution of LCs, the simulation prediction of the propagation of over-the-horizon signals between cities can be carried out. According to the results, within the relevant radius, methods such as high obstacle blocking, alternating high and low antennas, and flexible site selections of base stations could be adopted to carry out preventive 5G base station deployment between cities with similar terrains to what we simulated in this paper, where remote interference may occur.

**Figure 27.** The range-dependent PL on four intercity links at antenna heights of (**a**) 60 m and (**b**) 40 m. (**a**) Range-dependent PL from four cities to Wuxi at ANH = 60 m. (**b**) Range-dependent PL from four cities to Wuxi at ANH = 40 m.

### *4.5. Deep Learning Model Predicts Land-Based Ducting Propagation Using Geomorphology Data*

It was found that the intercity links whose terrains are similar to that of the intercity link analyzed in the previous sections, such as Shanghai–Wuxi or Jiaxing–Wuxi, are prone to the ducting effects caused by meteorological changes, causing remote interference at the antenna height. Since most of the relevant cities and landform data are less available, in this section, we use d1300 randomly generated datasets with different propagation distances, terrains, and land covers to perform physical simulations of over-the-horizon PL distributions under the same radar parameters, duct parameters, and random turbulence as the previous sections. The intercity link maximum distance was randomly generated within 90–100 km. The terrain relief height was randomly generated within 0–20 m. The LC type was randomly generated between 9 and 17 according to the IGBP classification scheme, and the corresponding dielectric parameters are listed in Table 3.

The simulated 1300 groups of terrain, land cover, and dielectric parameters at different distances and the PL at the corresponding antenna height were used for deep learning modeling. We intended to build a deep learning model to predict the PL on the intercity link with input parameters of terrain and land cover characteristics, and the ducting and radar environment settings remained the same as in previous sections.

Before training, because the randomly generated data may have been inconsistent with the actual situation, the obvious errors of the forward-propagation results were eliminated. A total of 56 sets of data were eliminated, and the remaining 1244 sets of data were used for deep learning modeling. After shuffling the order of datasets, the first 900 datasets were used as training sets, and the last 344 sets of data were used to test the training model. The

deep multilayer perceptron (DMLP) network was used here for deep learning modeling. We continuously optimized the parameters to obtain the optimal model, and a total of 4000, 2000, 3000, and 999 neural networks were built for training. The schematic diagram of the deep learning forward prediction model is shown in Figure 28.

One of the results calculated by the TWPE scattering model in the test set was used to compare the prediction result by the deep learning model, as shown in Figure 29a, with the antenna height at 60 m, and Figure 30a with the antenna height at 40 m. Using the mean square error as the loss function, the formula of the specific loss function is as follows:

$$loss = \frac{\sum\_{i=1}^{n} (pred(i) - y(i))^2}{n} \tag{20}$$

where *pred*(*i*) represents the range-dependent PL predicted by the network, *y*(*i*) is the true range-dependent PL calculated by the TWPE model, and *n* is the number of groups of training datasets.

**Figure 28.** The schematic diagram of the deep learning forward-prediction model.

Using the following definition to give the model prediction accuracy functions to evaluate the quality of the established model

$$acc = 1 - \frac{y\_{pred} - y\_{true}}{y\_{true}} \tag{21}$$

The accuracy of each range step was solved at antenna heights of 60 or 40 m, and the mean value was defined as the predicted accuracy of the sample. Then the predicted accuracy distribution was obtained on each tested sample. The accuracy results of test datasets are shown in Figures 29b and 30b, respectively.

When the antenna height was 60 m, the MSE of the DMLP prediction was 118.202749. When the antenna height was 40 m, the MSE was 108.696593. After the optimized deep learning parameters were well trained, they could quickly and efficiently predict (at a certain fixed antenna height, terrain, and landform) whether there would be communication interference problems between cities around 100 km apart with ducting and turbulence.

It can be seen from Figures 29b and 30b that the prediction accuracy function of the test set of the DMLP model simulation results is mostly kept between 0.8 and 1.0, and a good prediction effect was achieved. We selected two groups of DMLP predictions with relatively good test results and compared them with the TWPE simulation results. Figures 29a and 30a show that the DMLP prediction make large errors after 80 km, and the DMLP prediction results fluctuate greatly.

To improve the prediction effect of the model, we introduced the long- and shortterm memory (LSTM) network model to perform deep learning modeling on the original 1244 sets of data, and used two layers of LSTM networks with 80 and 60 neural units, respectively. The LSTM has requirements for the network input dimension. It is a matrix

form of [number of samples, number of expansions in time steps, number of sample features at each time step]. To meet the requirement of this parameter dimension, a total of 1244 groups of training and testing data of the original fully-connected layer were rearranged according to the four input features, which was distance, terrain fluctuation, the real part, and imaginary part of the dielectric parameters of the corresponding land covers. Then the input parameter dimensions were converted into the three-dimensional matrix. After the deep training using LSTM network, the prediction range-dependent PLs at antenna heights of 60 and 40 m were obtained, as shown in Figures 31 and 32.

**Figure 29.** DMLP prediction at a height of 60 m and accuracy distribution. (**a**) PL Prediction using DMLP at an antenna height of 60 m. (**b**) DMLP of PL at a height of 60 m of accuracy on the test samples.

**Figure 30.** DMLP prediction at a height of 40 m and accuracy distribution. (**a**) PL prediction using DMLP at an antenna height of 40 m. (**b**) DMLP of PL at a height of 40 m accuracy on test samples.

The LSTM model still maintains a good prediction effect. The MSE of LSTM prediction of PLs at an antenna height of 60 m was 93.749120. Moreover, the MSE of LSTM prediction of PLs at an antenna height of 40 m was 90.245694. According to the loss value, the MSE of LSTM was smaller than that of DMLP. The prediction accuracy of LSTM remained mostly between 0.8 and 1.0. However, the shortcoming of the rapid fluctuation and the large error after 80 km of DMLP were greatly improved in LSTM.

**Figure 31.** LSTM prediction at a height of 60 m and accuracy distribution. (**a**) PL prediction using LSTM at an antenna height of 60 m. (**b**) LSTM of PL at a height of 60 m accuracy on test samples.

**Figure 32.** LSTM prediction at a height of 40 m and accuracy distribution. (**a**) PL prediction using LSTM at an antenna height of 40 m. (**b**) LSTM of PL at a height of 40 m accuracy on test samples.

The terrain and land cover datasets used in the deep learning model were randomly generated. There might have been some random terrain and LCs that did not conform to the actual continuous distribution, so the model training results have certain errors. However, it can be seen that the topography and LC types have grea<sup>t</sup> influence on the PL distribution over the horizon. Therefore, it is effective and efficient to use the deep learning model to predict the over-the-horizon PL of the intercity link in the ducting environment with different landforms. The results also show that topography and LC types have grea<sup>t</sup> influence on the over-the-horizon propagation when a terrestrial duct happens. Subsequent continuous improvements of parameters to optimize the deep learning model may improve the prediction accuracy.
