Near Ground Pathloss Propagation Model Using Adaptive Neuro Fuzzy Inference System for Wireless Sensor Network Communication in Forest, Jungle and Open Dirt Road Environments

In Wireless Sensor Networks which are deployed in remote and isolated tropical areas; such as forest; jungle; and open dirt road environments; wireless communications usually suffer heavily because of the environmental effects on vegetation; terrain; low antenna height; and distance. Therefore; to solve this problem; the Wireless Sensor Network communication links must be designed for their best performance using the suitable electromagnetic wave behavior model in a given environment. This study introduces and analyzes the behavior of the LoRa pathloss propagation model for signals that propagate at near ground or that have low transmitter and receiver antenna heights from the ground (less than 30 cm antenna height). Using RMSE and MAE statistical analysis tools; we validate the developed model results. The developed Fuzzy ANFIS model achieves the lowest RMSE score of 0.88 at 433 MHz and the lowest MAE score of 1.61 at 433 MHz for both open dirt road environments. The Optimized FITU-R Near Ground model achieved the lowest RMSE score of 4.08 at 868 MHz for the forest environment and lowest MAE score of 14.84 at 868 MHz for the open dirt road environment. The Okumura-Hata model achieved the lowest RMSE score of 6.32 at 868 MHz and the lowest MAE score of 26.12 at 868 MHz for both forest environments. Finally; the ITU-R Maximum Attenuation Free Space model achieved the lowest RMSE score of 9.58 at 868 MHz for the forest environment and the lowest MAE score of 38.48 at 868 MHz for the jungle environment. These values indicate that the proposed Fuzzy ANFIS pathloss model has the best performance in near ground propagation for all environments compared to other benchmark models.


Introduction
In 2010, Wireless Sensor Network (WSN) technology was classified by ITU-T as a next generation network [1]. The purpose of this technological concept was to connect every sensor node using wireless technologies. This network can exchange and communicate between its nodes using wireless technologies such as WiFi, LoRa, zigbee, and many others [2]. WSN technologies have many advantages. They can monitor their environment using sensors, and their data can then be sent to the sink node and finally to the server

Models, Materials, and Methods
In this section, we would like to explain the experimental setup that we used in this research work. The experimental setup included wireless equipment, observed environment, and measurement. After the experiment was complete, the measurement data were analyzed to provide a good pathloss propagation model in near ground wireless communication environment.

Related Pathloss Propagation Models
This study attempts to provide different perspectives; therefore, we will benchmark by comparing it with other popular pathloss propagation models. In this section, we consider the Okumura-Hata model that is popular for mobile communication, the FITU-R model with foliage model, and the ITU-RMA Model that has ground reflected propagation model. These three models will be compared to the Fuzzy ANFIS model to quantify the accuracy improvement achieved by the propagation model for near ground wireless communication in forest, jungle, and open dirt road environments. The benchmark models' theoretical background is introduced below.

Okumura-Hata Model
Okumura-Hata Model is one of the popular pathloss propagation models for mobile communication applications. This model not only considers the effects of diffraction, reflection, and scattering, but it also takes urban, suburban, and open areas into account in conjunction with carrier frequency, distance, transmitter (base station) antenna height, and receiver (mobile) antenna height. This model is based on Okumura's measurements in Tokyo and the mathematical model by Hata [35].
PL hata (db) = 69.55 + (26.16 log f ) − (13.82 log ht) − A(hr) + ((44.9 − (6.55 * log ht)) log d) (1) However, because antenna correction A(hr) was for cities, we are not using any antenna correction. In this research, we measure tropical areas near cities and open road environments. Therefore, we use Okumura-Hata for open road and suburban models:  [36]. This model takes into account the plane earth model that explains direct ray in addition to ground reflected ray, which are received by the receiver, that is: In addition, Meng also added the ITU-R foliage attenuation (A ITU−R foliage (db) into plane earth model, that is: However, using measurement data taken from oil palm tree plantation using the ITU-R Model for Near Ground Forest Model optimize, the model becomes: where: PL FITU−R = Pathloss propagation using FITU-R model. f = Frequency carrier in GHz. ht = Antenna transmitter height in meters. hr = Antenna receiver height in meters. d = Distance between transmitter and receiver in meters. A, B, C = Least squared error fit from measured data, which is 0.48, 0.43, 0.13.

ITU-R Maximum Attenuation and Free Space Pathloss (ITU-R MA FSPL) Model
The ITU-R maximum attenuation model is recommended by the International Telecommunications Union for the frequency range 30 MHz-30 GHz. This model tells us that the ground reflected ray in this environment is negligible, since the forest ground is covered with shrubs that absorb the wave.
According to measurement result from Salameh, there is an indication that the direct ray is the major contributor to the received signal by the receiver which is located near the forest ground. Therefore, according to Salameh, this model is also considered as a free space model [37]. With respect to the free space model, the final Salameh model is: where: A M = Maximum excess attenuation in dB; in Salameh result was 38. R = R is the initial slope of the attenuation curve; in Salameh result was 0.9 db/m. d = Distance between transmitter and receiver in Km. f = Frequency carrier in MHz.

Measurement Equipment
For measurement, there are several radio interfaces that are usually used for WSN communication, such as LoRa [38], Wi-Fi [39], Zigbee [25], Bluetooth [40], and many more. However, in this conducted research, we used LoRa because it can achieve long range communication [41]. In a lot of reported literature, LoRa can achieve 900 m more [42] and even more in the kilometer range [43]. For measurement, we used a LoRa transmitter and receiver pair. The transmitter was paired with microcontroller ESP32. This way, we could program the LoRa transmitter using the microcontroller to send data (packet data containing "hello" word) wirelessly to the LoRa receiver every 100 microseconds. At the receiver site, the data transmission was received by the LoRa Radio receiver. Because the data that LoRa radio received was in hexadecimal format, the microcontroller ESP32 decoded it, extracted the RSSI value, and sent those values directly to our laptop every data transmission round. Measurements were performed using walk test for every 5 m from starting point up to 100 m. The measurement data were then fed into Fuzzy ANFIS as data training, and thus provided Near Ground LoRa Electromagnetic Wave Propagation model for all observed sites. The LoRa transmitter and receiver were placed directly on top of the soil with antenna height less than 30 cm from the soil (see Figure 1). The equipment parameters' detailed configurations are noted in Table 1.
from starting point up to 100 meters. The measurement data were then fed into Fuzzy ANFIS as data training, and thus provided Near Ground LoRa Electromagnetic Wave Propagation model for all observed sites. The LoRa transmitter and receiver were placed directly on top of the soil with antenna height less than 30 cm from the soil (see Figure 1). The equipment parameters' detailed configurations are noted in Table 1.

Measurement Environment
In this research's conducted measurement, 3 different sites were chosen because of their distinct or unique properties (please see Figure 2). The first site is a jungle where overgrown mass of vegetation spreads over a large area of land. This jungle has many plants on the ground between trees and larger plants. The very dense vegetation on the ground sometimes makes it difficult or impossible for humans to go around or travel through a jungle. The second site is a forest with a dense growth of trees covering a large area of land. It has many tall trees. However, paths between trees can be traveled through by humans. Finally, the last site is an open dirt road. However, on this site, the vegetation grows on the left side and on the right side of the road. These 3 measurement sites were chosen to study the effect of their distinct or unique properties and their effect on electromagnetic wave propagation. It will provide us with more information about electromagnetic wave behavior that propagates near the ground. The measurements were carried out in 3 different locations with 3 distinct characteristics. In all environments, the measurements were carried out with 54 different LoRa radio configurations. The forest environment, located in the eastern suburbs of Jakarta, has a GPS coordinate −6.366020, 106.901294. The jungle environment is also located in the eastern suburbs of Jakarta, and it has GPS coordinates −6.358641, 106.903730. The open dirt road environment is located in the western suburbs of Jakarta at GPS coordinates −6.214478, 106.747576.

Measurement Environment
In this research's conducted measurement, 3 different sites were chosen because of their distinct or unique properties (please see Figure 2). The first site is a jungle where overgrown mass of vegetation spreads over a large area of land. This jungle has many plants on the ground between trees and larger plants. The very dense vegetation on the ground sometimes makes it difficult or impossible for humans to go around or travel through a jungle. The second site is a forest with a dense growth of trees covering a large area of land. It has many tall trees. However, paths between trees can be traveled through by humans. Finally, the last site is an open dirt road. However, on this site, the vegetation grows on the left side and on the right side of the road. These 3 measurement sites were chosen to study the effect of their distinct or unique properties and their effect on electromagnetic wave propagation. It will provide us with more information about electromagnetic wave behavior that propagates near the ground. The measurements were carried out in 3 different locations with 3 distinct characteristics. In all environments, the measurements were carried out with 54 different LoRa radio configurations. The forest environment, located in the eastern suburbs of Jakarta, has a GPS coordinate −6.366020, 106.901294. The jungle environment is also located in the eastern suburbs of Jakarta, and it

Adaptive Neuro Fuzzy Inference System Method
Fuzzy system is mainly used for control, such as controlling robot movement [44], [45], speed [46], and other factors [47][48][49]. However, fuzzy system can basically be used for anything. Adaptive Neuro Fuzzy Inference System (ANFIS) architecture based on Jang's research consists of five stages [41,42], as shows in Figure 3. The square-shaped nodes show adaptive nodes, while the circular shapes are the fixed nodes.

Adaptive Neuro Fuzzy Inference System Method
Fuzzy system is mainly used for control, such as controlling robot movement [44,45], speed [46], and other factors [47][48][49]. However, fuzzy system can basically be used for anything. Adaptive Neuro Fuzzy Inference System (ANFIS) architecture based on Jang's research consists of five stages [41,42], as shows in Figure 3. The square-shaped nodes show adaptive nodes, while the circular shapes are the fixed nodes.
The Fuzzy ANFIS architecture based on Jang's research can be seen in this equation: For stage one, each output is symbolized by O 1 i , which serves to raise the degree of membership. where: i = every node in Fuzzy ANFIS architecture.
x = is the input to node i. A, B = is the linguistic label (such as small, large, etc.). In this stage, every membership function type can be used, but in Jang's method [50], generalized bell membership functions were used to provide two outputs: maximum equal to 1 and minimum equal to 0. Therefore, we obtain: where: The second stage is constructed by multiplying the two input signals. Every node represents the firing strength of fuzzy inference.
For the next stage, normalization was applied for each firing of fuzzy inference.
where: W = is the firing strength of node. W = is the normalized firing strength of node. The next stage contains the calculation of the output based on the parameters of the rule consequent. where: P, Q, R = is the parameter set.
Finally, the last stage computes the overall output as the summation of all input signals.
Because ANFIS learns from gradient descent and chain rule, error rate needs to be known for data training for each node output. Assuming i-th position node outputs as O i , the training data set has P number of entries and the error function can be measured as: where: E p = is error measure which is the sum of squared errors. T mp = is m component from P output target vector.
O L mp = is m component from actual output vector that has been produced by P input vector.
Hence, the error rate can be calculated as: where 1 ≤ k ≤ L−1 is error rate of an internal node; it is expressed as linear combination error rate of nodes in the next stages. Therefore, for all 1 ≤ k ≤ L and 1 ≤ i ≤ #(k), we can find ∂Ep ∂O k i p , using Equations (15) and (16). Now, we have α as a parameter of the adaptive network.
∂E ∂α where: S = shows the set of nodes whose output depends on α.
Derivative for overall error measurement E with respect to α is: Therefore, we can write the updated formula for generic parameter α as follows: where: η = is a learning rate. The learning rate can be written as where: k = is the step size of length of each gradient transition in the parametric space.
According to Zadeh, in 1975, fuzzy foundation was developed from Linguistic Variable and its Application to Approximate Reasoning [51]. Using those foundations, we can say that fuzzy rule was developed in order to model the qualitative aspects of human expertise (reasoning based on experience) [52] and solve or adapt the problem [53]. Therefore, using Equations (10) and (14), we can write the propagation model for near ground wireless communication in forest, jungle, and open dirt road environment as: where: x = Input variable such as frequency, bandwidth, spreading factor, range, and others. a = Defines the width of the membership function input. b = Defines the shape of the curve on either side of the midland. c = Defines the center point of the membership function. Fi = Constant Output Level generated automatically by ANFIS.

Results and Discussion
Our initial measurement result shows varying signal strength for each measurement point and each radio configuration. This problem we encountered was caused by different propagation phenomena such as diffraction, refraction, and reflection of the transmitted signal because of the surrounding vegetation environment [54]. As stated in Section 1, in the jungle, forest, or open road environment, vegetation and other things can induce small-scale multipath fading for the electromagnetic wave. However, as we observe the multipath fading in the dense jungle, the jungle is not as static as it may seem. The spatial distribution component of the multipath loss was clearly evident in the measured values collected at different points in the jungle. The temporal distribution component of the multipath loss in the jungle was also present in the fluctuation of the received signal strength at any point. This fluctuation is due to the dynamic movement of hidden obstacles on the jungle floor, tree trunks, canopies, and treetops. The jungle is rich with livestock of birds, squirrels, monkeys, rats, and many other living animals, in addition to the humans conducting the experiments. If we also add this factor, the effect of occasional wind can also result in detectable fluctuation in the RSSI. As expected, the multipath temporal distribution component is not as severe as it is in a typical urban environment. Evidence of this environmental dynamicity can be captured using a highly sensitive microphone or high frame rate camera. Therefore, to make the measurements reliable, we transmit 10 packet data and perform averaging for its signal strength. Even though we performed averaging for each measurement point, we still found that every measurement point also had a varying measurement result from other nearby measurement points. This was reflected in the zig-zag measurement chart for every environment.
In this study, we present a comparison for every respected near ground pathloss propagation model using three different frequency bands. Figure 4 illustrates the comparison of the different pathloss models with measurements in the forest environment. Figure 5 illustrates the comparison of the pathloss models with measurements in the jungle environment, while Figure  was reflected in the zig-zag measurement chart for every environment.
In this study, we present a comparison for every respected near ground pathloss propagation model using three different frequency bands. Figure 4 illustrates the comparison of the different pathloss models with measurements in the forest environment. Figure  5 illustrates the comparison of the pathloss models with measurements in the jungle environment, while Figure 6         If we carefully observe Figures 4, 5, and 6, there are a few interesting facts. At first, even though the jungle has a dense vegetation environment in comparison to the open dirt road, its pathloss measurement shows better results. Consider the rule of thumb that decreasing the transmit frequency leads to a reduced pathloss; this is evident in the open dirt road and forest environments. As shown in Figure 6, there is less reflection and scattering, due to foliage taking place. However, as the environment changes to severe scattering and multiple reflection conditions in the jungle, this rule of thumb is somewhat violated, especially for the 433 MHz frequency band. Note that the shrub height from the jungle floor is approximately 1 meter, as shown in Figure 2a. Since the transmitter and If we carefully observe Figures 4-6, there are a few interesting facts. At first, even though the jungle has a dense vegetation environment in comparison to the open dirt road, its pathloss measurement shows better results. Consider the rule of thumb that decreasing the transmit frequency leads to a reduced pathloss; this is evident in the open dirt road and forest environments. As shown in Figure 6, there is less reflection and scattering, due to foliage taking place. However, as the environment changes to severe scattering and multiple reflection conditions in the jungle, this rule of thumb is somewhat violated, especially for the 433 MHz frequency band. Note that the shrub height from the jungle floor is approximately 1 m, as shown in Figure 2a. Since the transmitter and receiver antenna heights are only 30 cm from the jungle floor, they suffer from what is known as the Fresnel zone non clearance effect [55]. As the antenna is trapped underneath and in between the taller shrubs, this effect results in a further pathloss component caused probably by diffraction at both transmitter and receiver ends and severe scattering from the taller shrubs (resembling diffraction from a roof-top or lamb-post antenna in a city environment). Due to the larger wavelength size, and hence larger Fresnel zone diameter, this diffraction pathloss component is magnified at the lower frequency of 433 MHz, as confirmed by [55] and compared to 868 and 920 MHz bands. Furthermore, due to this dense vegetation on the jungle floor, the signal travels several shrub-edge diffraction points to reach the receiver antenna trapped under the shrubs on the other side of the jungle floor. These multiple diffraction components have weakened the signal strength and added to the overall link pathloss. Because both the transmitter and receiver were placed above the ground with only 30 cm height, we assume that in the jungle environment, most of the LoRa signal ground reflections were absorbed by wet grass on the floor, resulting in small ground reflection. On the contrary, there are stronger reflections for LoRa signals in the open dirt road environment because the road was not covered by wet grass.
Another interesting fact is the large difference between the measurement and the predicted empirical pathloss model. The example of this was the measurement of 5 m using the 868 MHz band. Based on the empirical model ITU-R MA FSPL at 5 m, the RSSI value should be at −29 dBm while in the real measurement, it is −66 dBm (a −37 dB difference). This is because the ITU-R MA FSPL empirical model was derived from free space pathloss. Thus, we can explain this phenomenon using Chyriskos formulation which states that "signal loss is the sum of two independent attenuation processes: free space loss and losses due to obstacles" [56]. The RSSI value for free space pathloss using the 868 MHz frequency band at 5 m was −25 dBm, thus −41 dB was contributed by heavy obstacle loss. There are two main components for this obstacle loss, as follows:

•
Obstacle loss due to the vegetation environment that obstructed the signal. As stated by Salameh, "direct ray is the major contributor to the received signal by the receiver which is located near the ground of the forest. The implication here is that the ground reflected ray in this environment is negligible, since the forest ground is covered with shrubs that can absorb the wave" [37]. • Obstacle loss due to the Fresnel zone that was caused by low antenna height. Because our measurement was placed with an antenna height of less than 30 cm, this Fresnel zone acted as an obstacle according to Adi and Kitagawa. They stated that the Fresnel zone area is influenced by antenna height: the higher the value of the antenna height, the greater the percentage of the Fresnel zone clearance [55]. They further state that the lower frequency causes a bigger Fresnel zone; thus, it is no wonder that the measurement on 433 MHz generated a lower RSSI signal compared to 868 MHz and 920 MHz.
If we carefully observe the Okumura-Hata pathloss propagation model, we find that its behavior does not match the measurement results. The presented results in Figure 4 through Figure 6 show that the Okumura-Hata model has a curved line with between 5 and 50 m between the transmitter and receiver. Although the measurement results' trendline is a zig-zag line, overall measurement behavior does not show any curved shape anywhere. Moreover, there is a gap in the predicted values between  [57]. Table 2 shows the statistical evaluation for each pathloss propagation model using RMSE while Table 3 shows us statistical evaluation for each pathloss propagation model using MAE.  The statistical error analysis concept tells us that the smallest value indicates that the model has the best matching performance between the predicted values and observed values. We can see that the best model based on statistical evaluation is fuzzy ANFIS with the lowest RMSE score of 0.

Conclusions
In this study, we investigated and analyzed the behavior of LoRa pathloss propagation models in the near ground with low transmitter and receiver antenna heights from the ground. Furthermore, we introduced a fuzzy ANFIS model to predict the near ground pathloss in different tropical environments. We can see from the presented results that the most accurate prediction model, that agrees with the measurement results, in forest, jungle, and open dirt road environments, is the proposed fuzzy ANFIS model. We validated the performance using RMSE and MAE statistical analysis tools. The fuzzy ANFIS model achieves the lowest RMSE score of 0.