Comprehensive Measurement‐Based Assessment of Downlink RF‐EMF Exposure in Urban Environments: Multi‐Method Analysis and Intercomparison

ABSTRACT

This paper presents a comprehensive measurement‐based assessment of radio‐frequency (RF) electromagnetic field (EMF) exposure level in a French city. Three types of assessment methods are used to collect measurement data: drive test (DT), spot measurements, and sensor networks. The DT measurements were conducted by a portable spectrum analyzer, i.e., Tektronix RSA 306B, connected to a 3‐axis antenna mounted on the roof of the vehicle. DT system continuously recorded frequency‐dependent electric field (E‐field) values on a pre‐defined outdoor route. The spot measurements were done in the same region, covered by DT, with both broadband and frequency‐selective systems. Additionally, 19 sensors were installed on streetlamps in the same part of the city to measure the broadband E‐field level. The overall statistical analysis on raw data shows good agreement on RF‐EMF exposure level from three types of measurements. Then a distance‐based moving average method was carried out to remove the random noise in the DT data, where the optimized window size is explored using Kolmogorov‐Smirnov test. The smoothed DT data show a good correlation with nearby spot measurement values, as well as with base station antenna (BSA) density. Specific fifth‐generation (5G) spot measurements, performed with and without traffic‐attracting downloads, demonstrate the impact of beamforming on exposure levels in 5G new radio (NR) bands. Then spot measurements were used to build the exposure map using the kriging method, where the kriging prediction from the trained model is further compared with DT. Furthermore, the temporal variations observed in the sensor network were analyzed in relation to distance from the nearest BSA, revealing an inverse proportional relationship between E‐field level and proximity to the nearest BSA. This study shows good reliability in assessing the RF‐EMF exposure level using different systems. The advantages and limitations of different systems are also demonstrated by performing the intercomparison between data sets.

Summary

• Comprehensive Multi‐Method Assessment: This study evaluates RF‐EMF exposure in a urban city using three measurement methods: drive tests, spot measurements, and sensor networks, showing consistency across datasets.

• Data Smoothing and Correlation Analysis: A distance‐based moving‐average method is applied to drive test data to reduce noise, with results correlating well with spot measurements and base station antenna (BSA) density.

• 5G Beamforming Impact & Exposure Mapping: A specific 5G spot measurement assesses beamforming effects on exposure levels, and an exposure map for the city is generated using kriging interpolation from 100 spot measurement points.

1. Introduction

With the increasing use of wireless networks and radio frequency (RF) electromagnetic fields (EMF) and despite extensive daily use of cell phones, health risk perception linked to RF‐EMF persists. These fears have led to many debates on the deployment of the fifth generation (5G) infrastructures, showing that risk perception is well established (Koh et al. 2020). Even if limits exist (ICNIRP 2020; Wiart 2016) and are used by regulations (Council of the European Union 1999), to guarantee that devices put in the market and infrastructures are safe, recent studies (Eggeling‐Böcker et al. 2025) still report notable levels of public worry in some countries. In France, ANSES reports have noted significant public controversy surrounding the rollout of 5G in 2021, particularly regarding health risks (ANSES 2022). To respond to the public and for compliance purposes, long‐term massive RF‐EMF monitoring campaigns have been carried out in France (ANFR 2018), Slovenia (Gajšek et al. 2015), Japan (Onishi et al. 2021), Korea (Lee et al. 2024) and some EU countries (Tesanovic et al. 2014; Iakovidis et al. 2022) for years. In this context, a number of major research programs have been undertaken in Europe to advance scientific understanding, among them the EMF Health Cluster CLUE‐H (EMF Health Cluster n.d., https://www.emf-health-cluster.eu/), which brings together European research organizations in four major consortia (ETAIN, GOLiAT, NextGEM, and SEAWave).

In RF‐EMF exposure assessment, simulation (Leeman et al. 2025) and mathematical modeling (Gontier et al. 2024; Al Hajj et al. 2020) have provided cost‐effective solutions for decades. For example, kriging‐based reconstruction, which has been widely used in exposure analysis (Lemaire et al. 2016; Aerts et al. 2013), is typically performed using spot measurement data. They allow exposure to be assessed under diverse conditions with reduced time and resources, but they lack reliability in fully capturing the complexities of real‐world environments.

On the other hand, in‐situ measurements provide a more reliable and accurate approach for RF‐EMF monitoring. Depending on the study objectives, measurement tools can range from mobile phone based solutions (Liu et al. 2024) to dedicated standalone instruments (Bhatt et al. 2022). For base station‐induced RF‐EMF exposure, i.e., downlink (DL)‐related, the monitoring work is generally carried out by three types of assessment methods, namely spot measurements, sensor networks, and drive test (DT) measurements. When linked to local authorities requests, the spot measurements are often carried out by accredited laboratories, such as National Frequency Agency (ANFR) (ANFR 2023) in France or Greek Atomic Energy Commission (EEAE) in Greece (Christopoulou and Karabetsos 2015) that follow well‐defined measurement protocols (ANFR 2023) based on international standards (IEC 2022). In (Sefsouf et al. 2024), the extensive 5G measurements campaign in France was presented. This standardized methodology involves taking frequency selective measurements but also looking for the maximum exposure using a broadband probe. Due to their standardized procedures, spot measurements are usually time‐consuming and have to be carried out at a fixed location. In addition, it lacks temporal resolution and has limited spatial resolution.

To overcome the missing temporal resolution of spot measurements, EMF sensors have been installed in several European cities (EXEM 2025; EEAE 2025; Djuric et al. 2022). Those sensors are usually (but not always) placed in the outdoor environment, to continuously measure the EMF strength in that area. Since sensors are activated regularly to measure, e.g., seasonal, weekly, and daily trends can be recorded over a long duration. In (Jawad et al. 2024), the temporal trends were analyzed by using principal component analysis (PCA). This temporal trend can provide insights into network evolution, mobility, user pattern change, etc. The data collected can provide a more detailed picture of EMF exposure levels over time than the one‐time measurements, which are only accurate at the time the measurement was taken. However, the spatial resolution of sensor networks is limited due to the high cost of large‐scale installations. In (Wang and Wiart 2020), it was found that more accurate prediction on EMF exposure requires a large number of sensors.

To address the resource‐intensive constraints of spot and sensor measurements, DT measurements have been proposed (Wang et al. 2022; Onishi et al. 2023) to cover both spatially and partially temporally the exposure level within a given area. DT measurements can be performed by driving (Onishi et al. 2023), cycling, or walking along a predefined route (Chikha et al. 2024) with a portable EMF measurement system that records exposure levels in real‐time. For larger areas, such as urban cities, this approach allows for efficient assessment within a few hours when using a car. It provides us with a large‐scale spatial RF‐EMF exposure map at a much lower cost. Despite the ability of DT measurements to collect large amounts of data in a short period, they struggle to provide stable exposure levels as each drive test measurement point corresponds to a short‐duration recording $T$ (over small distance) of the instantaneous electric field on three orthogonal axes.

Although the aforementioned measurement methods have been used for several years, assessments of RF‐EMF exposure in one measurement campaign typically rely on a single type of measurement system. Inter‐comparisons between different systems, as well as evaluations of post‐processing methods, have not been thoroughly studied. In this study, we present a comprehensive analysis of multiple measurement systems for assessing downlink RF‐EMF exposure levels in an urban environment.

Since 2019, sensor networks and DT measurement campaigns have been actively conducted (Wang et al. 2022; Wang and Wiart 2020), while spot measurements have been performed since 2014 (ANFR 2018). In this study, we analyze measurement data collected over the past 3 years from all three approaches, i,e, spot measurements, sensors, and DT, carried out in the same urban environment. The sensors were installed by Télécom Paris team in collaboration with the local authorities. The spot measurements were performed by ANFR on selected 100 locations. The DTs were carried out by Télécom Paris team using the Tektronix system.

After presenting the measurement equipment setup, we evaluate the consistency between the different systems by comparing the statistical distributions of E‐field values across all downlink cellular bands, as well as the per frequency band performance. A dedicated 5G measurement campaign was carried out in the 3500 MHz band under proactive phone usage conditions. For the DT data, a distance‐based moving average method with an optimized window size was applied, whose results were compared against spot measurements. The relationship between RF‐EMF exposure and base station antennas (BSA) was further examined using both DT and sensor network data. Finally, spot measurements were used to construct a kriging interpolation model, whose predictions were compared with DT results.

In this paper, we present a comprehensive multi‐method assessment of RF‐EMF exposure by combining spot measurements, DT, and sensor networks in an urban environment. Inter‐measurement comparisons are performed using both raw and processed data. The reliability of post‐processing methods, such as distance‐based moving averaging and kriging interpolation, is evaluated by comparing processed results with datasets from other systems. The influence of BSA on exposure levels is further demonstrated through analyses of DT and sensor network data. Finally, our study highlights the complementary strengths and limitations of the different measurement systems for RF‐EMF exposure assessment.

The paper is organized as follows: Section 2.1 and 2.2 introduce the measurement systems for DT, sensors, and spot measurements. Section 3 gives the main methods used in the data analysis. Section 4 presents the analysis of collected data. Section 5 concludes the paper.

2. Measurement Systems

This section provides a detailed description of three main measurement campaigns: measurements carried out by car, i.e., DT, measurements from sensor acquisitions, and those obtained from spot measurements. All three types of measurements were performed in Massy, a suburban city in the Greater Paris area (known as Ile‐de‐France).

2.1. Drive Test Measurement

A portable Real‐Time Spectrum Analyzer (RTSA) was used in the DT, i.e., RSA306B from Tektronix (Tektronix 2025). A wide frequency band from 700 MHz to 3800 MHz was selected to cover all active cellular frequency bands. The choice of one single wide band was due to the consideration of acquisition speed during the DT measurements. The RTSA was connected to a 3‐axis dipole antenna, i.e., TAS‐1208‐01 probe from MVG (MVG n.d) (9 kHz ‐ 6.2 GHz) via a switch controlled by Arduino. The isotropic measurements were realized by combining the root mean square (RMS) value from each axis. In the measurement setup, the 3‐axis antenna was fixed on the top of the vehicle while the rest of the measurement system stayed inside the car with the experimenter, as shown in Figure 1.

Figure 1.

Installation of 3‐axis antenna on the top of the vehicle.

An example of the DT carried out in January 2023 is shown in Figure 2. The predefined routes were designed to cover main outdoor facilities, including a train station, commercial centers, residential areas, and office buildings. The color bar indicates the measured E‐field level in V/m. The broadband spectrum data were recorded instantly during the DT. In the post‐processing stage, only the downlink cellular bands were extracted for analysis. For each selected band, the corresponding power samples were filtered and then used to calculate E‐field values including the central frequency, reference level, span and resolution bandwidth (RBW). The route shown in Figure 2 covers a total distance of 16 km. The average driving speed was approximately 20 km/h (i.e., about 5.6 m/s). For each recording, the RSA306B system recorded a complete 3‐axis measurement within 0.8 s, from which the total electric field was calculated. Given the vehicle's motion, successive measurements were spaced by an average of 4.4 m, so each point represented a spatial average over this distance.

Figure 2.

Map of the city with three types of measurements, where spot, sensor, and DT locations are indicated with diamond, square, and circle markers. The color of the circle marker represents the E‐field values from DT with a unit of V/m.

Since the RTSA did not include a built‐in GPS module, GPS data were recorded separately using the ‘Geotracker’, i.e., Android application on a pre‐synchronized mobile phone. Before the drive test, clock synchronization is done between the RTSA and the GPS recorder. Then after the drive test, it was used to align the two datasets and apply the GPS interpolation to assign accurate longitude and latitude coordinates to each measurement point. All GPS traces were manually checked to ensure that the recorded positions followed the actual street trajectories. Additionally, the data recorded over a single wide band was post‐processed to obtain E‐field values for each frequency band. The active frequency bands considered in the DT are listed in Table 1.

Table 1.

Cellular frequency bands in France.

Band Name	B700	B800	B900	B1800	B2100	B2600	B3500
${\mathrm{F}}_{\text{min}}$ (MHz)	758	791	925	1805	2100	2620	3490
${\mathrm{F}}_{\text{max}}$ (MHz)	788	821	960	1880	2170	2690	3800

2.2. Sensor Networks

Sensor networks have been widely deployed in French cities since 2019 (EXEM 2025). These sensors, designed by EXEM, are typically installed on streetlamps in urban areas. As shown in Figure 3, the monitoring sensor is installed in the measurement environment. The sensor itself, designed by Exem, internally consists of three orthogonal E‐field probes connected to diode detectors, an acquisition system, a Sigfox‐based wireless communication module, and batteries (EXEM n.d). The sensors are installed at a height of about 4 meters to ensure they remain out of reach of pedestrians. The 3‐axis broadband sensors measure the spatial components of the E‐field across a frequency range from 250 kHz to 6 GHz, which includes all active bands listed in Table 1. Every 2 h, the sensors are activated to perform a measurement lasting 6 min, ensuring stable performance in recording E‐field strength. Since April 2022, 15 such sensors have been installed in the urban area of the city under study, as shown in Figure 2.

Figure 3.

EMF Sensor installed on streetlamps and the sensor networks map in Massy.

2.3. Spot Measurements

The spot measurements were conducted at 100 selected locations in November and December of 2022, using two types of EMF field meters. The Narda NBM 550 equipped with probe EF‐0691 from Narda was used for broadband measurements, referred to as “Case A”, covering a frequency band from 100 kHz to 6 GHz. The Narda SRM 3006 (Narda n.d) is used for frequency‐selective measurements, referred to as “Case B”, and can be equipped with several 3‐axis antennas, i.e., probe 3501/03 from Narda covering 27 MHz to 3 GHz and probe 3502/01 from Narda covering 420 MHz to 6 GHz. At each spot, measurements were performed at three heights (1.1 m, 1.5 m, and 1.7 m), with each height measured for approximately 6 min, in accordance with the French in‐situ measurement protocol defined by ANFR (ANFR 2017). The E‐field values obtained at the three heights were then averaged and used in the subsequent analysis throughout the paper.

The locations of the spot measurements were selected with a separation of 150 m to 200 m, as shown in Figure 2. The 100 spots were chosen to be distributed as regularly as possible and to overlap with the driving route and sensor locations, enabling later comparison with the other two measurement results.

The detailed measurement protocol and description can be found on the website of Cartoradio (Cartoradio 2025). We notice that both “Case A” and “Case B” cover the active DL cellular frequency bands listed in Table 1, which facilitate further comparison with DT and sensor measurements. In addition to the “Case A” and “Case B” measurements, an extra 5G measurement at B3500 was conducted in the band B3500, with a mobile terminal nearby downloading files.

2.4. Base Station Antenna Analysis

The strength of DL EMF exposure emitted by base station antennas BSA depends on multiple factors. From a propagation perspective, exposure levels are influenced by the surrounding environment, the antenna characteristics, and the distance between the antenna and the exposed person or object. From a temporal perspective, exposure levels vary with the traffic load, which typically changes on a daily basis.

Among these factors, we have partial access to BSA information through Cartoradio (Cartoradio 2025). The accessible data include the coordinates, azimuth angle, height, and operating frequency of each antenna. However, important parameters such as transmit power, tilt angle, traffic load, and instantaneous antenna patterns are not disclosed. Therefore, in this study, BSA information is used only as a proxy indicator of potential exposure.

The analysis in this paper considers both BSA density and the distance to the nearest BS site. Since each BS site may host multiple antennas from different frequency bands and operators, we define BSA density as the number of active antennas within a given 2‐dimensional area. The distance to the nearest BS site is calculated as the 3‐dimensional distance from the measurement point to the site, where the BS site height is taken as the average of the antenna heights installed at that location.

3. Statistical Methods

In this section, the main statistical methods used in the data analysis are explained, including the moving average method, the Kolmogorov‐Smirnov (KS) test, and the kriging method.

3.1. Distance‐Based Moving RMS

Moving average (MA) is a widely used statistical technique employed in signal processing by smoothing out fluctuations. It involves calculating the averaged value of $N$ data points over a window with a defined window size, where averaging methods may vary, e.g., simple MA, exponential MA, and weighted MA.

In this paper, we use the distance‐based MA to process the DT measurement data. For each target data point $x_i$ in DT measurements, we calculate the RMS average of the E‐field at point $x_i$ (see Equation 1 for more details) over all data points within the disk area $\mathrm{ℬ}(x_i,r)$, where $r$ represents the radius of the selected circle. We adopt the distance‐based moving RMS since the data points collected from the measurements are usually not evenly spaced.

$${\hat{E}}_{x_i}=\sqrt{\frac{1}{\mid \mathrm{ℬ}\left(x_i,r\right)\mid }\sum _{x_j\in \mathrm{ℬ}\left(x_i,r\right)}{E_{x_j}}^2}$$ (1)

where $\mid \mathrm{ℬ}\left(x_i,r\right)\mid$ is the number of points inside $\mathrm{ℬ}\left(x_i,r\right)$, centered at $x_i$ with radius $r$.

The distance‐based moving RMS method is only applied during the post‐processing of data collected from DT measurements. Due to the short acquisition duration for each instance in the DT, the measured data exhibits significant fluctuations, with values varying rapidly over short time intervals. These fluctuations are attributed to the dynamic changes in the surrounding environment.

3.2. Kolmogorov–Smirnov Test

The KS test is a non‐parametric statistical tool used to assess similarity by comparing the cumulative distribution functions (CDFs) of two distributions. It quantifies the maximum vertical deviation between two CDFs. Therefore, it is useful for determining whether the empirical data set originates from a specific distribution.

In this paper, the KS test was used to assess whether the measured E‐field values within a spatial window could be approximated by a Gaussian distribution. A spatial window was selected such that large‐scale fading was assumed to remain constant inside the window, with only the random variations contributing to the measured data. The KS test compared the empirical distribution of data inside selected window to the Gaussian model. Validating this assumption helped determine whether the chosen window size was appropriate for subsequent moving‐average processing.

The metric used to determine whether accepting or rejecting hypothesis is the p‐value, with a significance level of 0.05 set as the threshold. If $p<0.05$, the null hypothesis is rejected, suggesting that the data set does not follow a Gaussian distribution. On the contrary, if $p>0.05$, the null hypothesis is accepted, denoting the data set follows a Gaussian distribution.

3.3. Universal Kriging

Kriging is a geostatistical method used for interpolation and estimation in spatial data, where universal kriging is the most commonly used type. A key component in building a kriging model is the variogram, which models how spatial correlation varies with distance. By computing the distance and value differences between each selected pair of points, the empirical variogram can be built. Then, the empirical variogram is fitted into an existing theoretical model, such as spherical, exponential, or Gaussian. Afterward, the weights ${\lambda }_i$ are determined based on the spatial correlations among the points and their distances to the target point. Then, the interpolated value at the target location $x_0$ is calculated as follows:

$$Z^{*}\left(x_0\right)=\sum _{i=1}^N{\lambda }_{i}Z\left(x_i\right)$$ (2)

where $Z^{*}\left(x_0\right)$ is the estimated value at the target location, and $Z\left(x_i\right)$ are the observations at the known data points.

In the present paper, the 100 points from spot measurements are used to build the kriging model, whose interpolation results are displayed in Section 4.

4. Results

In this section, we first present the statistical analysis and comparisons of data collected from all three measurement approaches. Next, the effect of the distance‐based MA is evaluated using the KS test, with results further validated against spot measurements. Finally, a kriging‐based interpolation model is constructed from spot measurement data and its predictions are compared with DT measurements.

4.1. Overview of Measurement Results

The CDF results are presented in Figure 4, showing the distributions of E‐field values for all three types of measurements. The DT data include results from three measurement campaigns conducted in 2022, 2023, and 2024, while the sensor data have been collected since their first deployment in the city, covering the period from 2022 to 2024. From the CDF plot, it can be observed that the median E‐field values for the different types of measurements range between 0.3 V/m and 0.5 V/m.

Figure 4.

CDF and boxplot comparison between spot, sensor, and DT measurements from 2022 to 2024.

A similarity in distributions across all measurement types can be observed in the boxplot shown in Figure 4. The box in the plot is defined by the lower bound, middle line, and upper bound, representing the 25th, 50th, and 75th percentiles of the data, respectively. The boxes for different measurements show good consistency. Generally, spot and sensor measurements exhibit lower variability, as indicated by the reduced dispersion of outliers in Figure 4. This reduced variability can be attributed to the longer measurement duration of each sample. On the contrary, DT data show a higher number of outliers because the real‐time spectrum analyzer acquires instantaneous broadband samples within a very short duration (0.8 s on average), in contrast to the long integration times (e.g., 6 min) used in spot and sensor measurements. This motivates the necessity of applying the MA method to process DT data. Additionally, as shown in Table 2, spot measurements have the lowest number of samples, which explains the fewer outliers in Figure 4.

Table 2.

Number of measurement samples in each type.

	Spot “Case A”	Spot “Case B”	DT 2022	DT 2023	DT 2024	Sensor
Numbers	100	100	411	3672	3654	186478

As mentioned in Section 1, the limited number of spot measurements may fail to capture the exposure distribution across the entire selected area. This is evident from the roughness in the CDF of “Spot A” and “Spot B” in Figure 4. On the other hand, fewer sensors are installed at fixed locations, which reduces their spatial representativeness. Despite differences in devices, installation heights, time of measurement, and locations, we provide an overall comparison for reference purposes.

The uncertainty of the measurements arises from both environmental and instrumental factors. Environmental uncertainty reflects the propagation conditions at the time of measurement. Due to the large difference in measurement time for drive test, sensor, and spot measurements, a direct comparison of their environmental uncertainty is not possible. Instrumental uncertainty, on the other hand, differs across devices: broadband sensors and “case A” spot measurements generally exhibit higher uncertainty than frequency‐selective devices. Despite these differences, the results in Figure 4 show good agreement between “case A” and “case B” spot measurements, where environmental uncertainty can be considered comparable since the measurements were performed at nearly the same time at each location.

Since “Case B” measurements represent frequency‐selective results, the comparison of per frequency band behaviors is performed between spot “Case B” and DT measurements collected in 2023 and 2024 is shown in Figure 5. It is noted that the DT data from 2022 is not included due to its significantly smaller sample size compared to 2023 and 2024, as the driving route in 2022 is not long enough to cover the entire city.

Figure 5.

RMS averaged frequency selective comparison from spot measurements and DT.

Therefore, the E‐field strength is extracted for each cellular band from DT measurements. The same bands are selected from “Case B” of the spot measurements. Bear in mind that the spot measurements were performed at the end of 2022, which is temporally close to the DT campaign conducted in early 2023. The inter‐measurement equipment comparison of E‐field values shows a good agreement between the spot and DT measurements in both 2023 and 2024. Furthermore, we observe a slight increase in E‐field strength in almost every band in DT from 2024, compared to 2023.

The reported E‐field levels for all measured cellular bands (700‐3500 MHz) are significantly below the International Commission on Nonionizing Radiation Protection (ICNIRP) reference limits, indicating minimal exposure relative to the regulatory standards. Among those bands, we observe slightly higher values in 700 MHz and 800 MHz bands than other bands but still significantly below the ICNIRP limits. This could be attributed to the higher usage of lower frequency bands by operators. In addition, the E‐field levels in the B3500 band were observed to be significantly lower than those in the other bands possibly due to the low usage in the 5G NR band. The impact of 5G beamforming will be analyzed later in Section 4.3.

4.2. Drive Test Measurements

Since the DT measurement done in 2023 is temporally closest to the spot measurement campaign, we use DT in 2023 to validate the MA algorithm. As explained in Section 2.1, the total DL RF‐EMF exposure is calculated based on RMS values from all DL bands. The RMS average of the total E‐field level recorded from the DT is 0.83 V/m with a standard deviation (std) of 0.59 V/m. The maximum value recorded is 5.8 V/m. Measurement points near the BSA generally have a higher possibility of experiencing higher exposure levels. However, this also depends on the local environment surrounding the receiving probe and nearby BSA characteristics, e.g., azimuth of the main beam.

Since the DT measurements record instantaneous 3‐axis E‐field values at each sampling point along the route, the data include fluctuations caused by the dynamic urban environment (e.g., moving vehicles, buses, pedestrians) and thermal noise from the device. These variations can obscure the underlying large‐scale trends. Therefore, we applied the moving‐average (MA) method in Equation (1) to smooth out short‐term fluctuations and obtain a more stable representation of the large‐scale field strength. In Figure 6, the sliding window size is set to be $\pm$20 m to $\pm$300 m, where 20 m represents the radius of circle mentioned before. With a bigger window size, the curves are smoother while more variations are removed. Note that we should not consider very large sliding window sizes to avoid removing the variation caused by large landscapes. Since the unwanted noisy variation follows Gaussian distribution due to the noisy nature of the data, we performed the KS test to find the optimal size of the sliding window and therefore smooth only the noise level.

Figure 6.

Moving average using different sliding window sizes.

In Figure 7, we present the results of KS tests for five considered sliding window sizes. For each sliding window size, we apply the KS test for all data inside the window and check if the data falls into Gaussian distribution. If it is close to Gaussian distribution, the p‐value, given by the KS test, will be higher than 0.05, which means that the null hypothesis should be accepted. The CDF in Figure 7 plots the distribution of p‐values in all windows, compared with different window sizes. With a smaller window size, most of the p‐values from all windowed data are higher than 0.05, which indicates more data are Gaussian distributed. It also verifies the necessity of adopting MA to remove the randomness. Comparing sliding window size, measurement data inside a window size of $\pm 100\mathrm{m}$ has around 60% probability to be Gaussian‐distributed. For size of $\left\{\pm 200\mathrm{m},\pm 300\mathrm{m}\right\}$, less than 20% can be considered as Gaussian‐distributed. To strike a balance between smoothing local noise and preserving variations from large‐scale environmental changes, a window size as large as possible is chosen, to minimize noise while retaining key variations. Therefore, we adopt the sliding window of 100 m in this study.

Figure 7.

KS test for data inside one sliding window.

4.3. Spot Measurements

In Figure 5, both DT and spot measurements display only the passive received exposure from 5G NR bands. Although there is an increase in E‐field level in the 3.5 GHz band from DT in 2023 to 2024, the contribution of 5G to total E‐field remains relatively small (with E‐field value around 0.1 V/m) compared to other frequency bands (with E‐field value no less than 0.3 V/m). As we know, 5G networks at B3500 use dynamic beamforming, which directs radio signals toward active users rather than broadcasting uniformly in all directions, as was the case with previous generations.

To better assess E‐field levels at 3.5 GHz, Table 3 compares exposure levels from 5G B3500 bands, i.e., B3500, in scenarios with and without mobile phone attracting the traffic. SIM cards from the four main operators in France were tested, and the results with successful data transmissions were recorded. The traffic attraction was done by downloading the same file at different locations, where the downloading duration is also noted in Table 3. Measurement points were selected based on 5G network availability across all telecommunication operators. Each row represents the time average E‐field level at the same location under both conditions. In most locations, exposure levels increase significantly during file downloads. However, locations 3, 6, and 11 show only a slight increase, likely due to weak connections with the 5G BSA. This is supported by the observation that lower E‐field levels generally correspond to longer file download durations.

Table 3.

E‐Field in 5G band (B3500) with and without mobile phone downloading files.

Location	E‐field (V/m) without downloading	E‐Field (V/m) with downloading	Duration (s) of downloading
1	0.13	1.03	21
2	0.36	2.64	17
3	0.02	0.1	50
4	0.18	3.97	14
5	0.18	2.09	18
6	0.08	0.41	36
7	0.03	0.4	18
8	0.17	1.18	32
9	0.59	9.75	16
10	0.17	0.51	26
11	0.32	0.38	37

4.4. Sensor Networks

The sensor networks deployed in the urban area of Massy capture the hourly variation of E‐field exposure. We analyze the mean value and std from all sensors. The averaging duration for each sensor is different, as 15 out of 19 sensors were installed in April 2022, while the rest were deployed in January 2023. From Figure 8, we observe that most sensors measure a quite low exposure level, around 0.5 V/m. The 3‐D distance to the nearest physical BS site is also plotted in the figure. It should be noted that two of them stand out, i.e., 04 & 16, since they are in the line‐of‐sight (LoS) and main beam direction of nearby active BSA. While the distance to the nearest BS partially explains the E‐field levels, other factors, such as the main beam direction of the antenna, the number of active antennas at the site, and the surrounding environment near the sensor, also play crucial roles.

Figure 8.

Mean values from all sensors installed in Massy, considering a total duration from 6 months to 32 months.

For the sensors that record a mean value higher than 0.5 V/m, we performed a temporal variation analysis. In Figure 9, the scaled temporal variation is provided. Each sensor measurement is scaled with its mean and then averaged over 24 h. It can be seen that all sensors exhibit a clear decrease in E‐field levels at night and an increase during the day.

Figure 9.

Scaled temporal variation from sensor networks with significant E‐field level.

4.5. Inter‐Measurement Comparison

4.5.1. DT versus Spot

Here, we compare the total E‐field from the spot measurement (“Case A”) with the total E‐field from DT after MA. Although the spot measurements were conducted at different times of the day, the temporal variation during working hours (9:00–18:00), as shown in Figure 9, is limited. The increase from 9:00 to 18:00 is less than 20%. Therefore, for comparison, we selected the nearest DT point to each spot measurement point within a radius of 50 m. The correlation between the paired data was analyzed, yielding a moderate‐to‐strong correlation coefficient of 0.70. This indicates that the good agreement between spot and DT measurements is not only reflected in the overall distribution but is also evident in a per‐location analysis. Additionally, the MA algorithm with a window size of $\pm$ 100 m performs effectively in processing the DT data.

4.5.2. DT versus BSA Density

We analyze the correlation between the total E‐field from DT (after MA) and BSA density. Here, we choose the circle with radius of 200 m to calculate the BSA density. The choice of the radius was deduced from BSA density in the urban area of Massy, which is derived by dividing the total number of BS sites by the area of the central part of Massy. This indicates the average coverage area for each BS site is around 200 m. The correlation coefficient between the smoothed E‐field and BSA density is 0.59, indicating a moderate correlation between the received exposure levels and local BSA density.

4.5.3. Mapping Using Kriging

Figure 10 shows the results of kriging interpolation using data from 100 spot measurements. First, the Leave‐one‐out cross‐validation (LOOCV) was used to determine the best variogram model to be used. Then, we selected the “spherical” model in the universal kriging. The interpolation result was plotted in a grid of 100$\times$100 points. As shown in Figure 10, the kriging interpolation reproduces the original spot measurements, including the high‐value hotspots, which demonstrates the interpolation nature of kriging.

Figure 10.

Exposure map generated by kriging interpolation, where the color in the map indicates E‐field level in V/m.

We used the built kriging model to predict the E‐field level at the locations collected from DT, whose results are shown in Figure 11. The CDF plots exhibit an overall good alignment between the spot and DT data, although a small gap is present. This discrepancy may arise because kriging is good at interpolation but not at extrapolation. Figure 12 shows the kriging prediction with 95% confidence intervals, together with the ground truth from DT measurements after applying the moving average. The results confirm that kriging performs well for interpolation in general, but also highlight that the model tends to underestimate locations with higher DT values. The inherent assumption for kriging determines that spatial data follows a smooth, continuous trend. Therefore, kriging may sometimes miss high values in predictions, particularly in datasets where extreme values are rare and not well captured by the model. Consequently, DT locations outside the spot region may be inaccurately predicted. Another possible reason is the absence of environmental data, as kriging relies only on spatial correlation.

Figure 11.

Comparison between kriging prediction and DT.

Figure 12.

Kriging prediction of E‐field values with 95% confidence interval (shaded band) and DT measurement after MA (dots).

5. Conclusion

In this paper, we present a comprehensive RF‐EMF exposure assessment in the city of Massy, France, using three types of measurement systems: DT, fixed sensors, and spot measurements. The study covers all active downlink cellular frequency bands, with DT measurements performed repeatedly over 3 years, 19 sensors monitoring broadband exposure, and spot measurements conducted in both broadband and frequency‐selective manners.

A general comparison across systems shows good consistency in the distribution of total E‐field levels, confirming that average exposure levels over the city can be characterized by any of the three approaches. Beyond this, our spatial analysis of DT data, enhanced by MA post‐processing, demonstrated strong correlations with both local BSA density and nearby spot measurements. Similarly, in sensor networks, higher mean exposure values were typically associated with shorter distances to the nearest BS site. Specific 5G spot measurements in B3500 band were conducted with and without continuous mobile phone file downloads. A significant difference in the measured E‐field levels was observed between the two cases. Additionally, kriging interpolation, based on 100 spot measurements, was used to generate an exposure map.

The novelty of this study lies in its multi‐method comparison under real‐world conditions, showing how different measurement systems can provide consistent exposure characterizations while also offering distinct advantages. Each approach, however, has its own limitations:

• DT measurements benefit from rapid data collection over large areas. However, it is affected by the short acquisition times and instantaneous random noise, which require post‐processing.

• Spot measurements are reliable but limited by cost, effort, and sparse spatial coverage. Nevertheless, strategically selected spot locations can support interpolation models (e.g., kriging) for spatial prediction.

• Sensor networks provide high temporal resolution and automation, but limited spatial representativeness.

In practice, no single system is sufficient on its own. A judicious combination of complementary methods, leveraging the quick spatial coverage of DT, the reliability of spot measurements, and the temporal continuity of sensor networks, can offer the most robust strategy for future RF‐EMF exposure assessments.

Nevertheless, our results highlight that exposure levels are influenced not only by BSA distance and density, but also by additional environmental factors. Further investigation into the correlation between downlink exposure levels and local environmental characteristics will be considered in future work. Furthermore, we will extend this study by integrating kriging from multiple data sources to enable spatial‐temporal prediction, by assessing specific exposure in 5G B3500 band with both downlink and uplink considered, and by conducting distribution analyses of exposure collected from additional cities to evaluate the impact of diverse environments.

Conflicts of Interest

The authors declare no conflicts of interest.

Wang, S. , Zhang Y., Liu Y., et al. 2025. “Comprehensive Measurement‐Based Assessment of Downlink RF‐EMF Exposure in Urban Environments: Multi‐Method Analysis and Intercomparison.” Bioelectromagnetics 46: 1–12. 10.1002/bem.70033.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.